Lately it’s impossible to open LinkedIn, Twitter (sorry, Elon, I’ll never call it X), or our favorite social media without running into a post about AI agents.

You’re probably thinking that Claude or ChatGPT are agents, and that their marketing teams are just rebranding them from “assistants” to “agents” to make them sound cooler. Well no, this time it’s not just hype, it’s an actual shift in the technology.

And it’s not just any shift. We’re talking about an AI that’s starting to have goals and act on its own to achieve them. (Wasn’t that the start of every apocalyptic tech movie?)

Today I want to tell you what an agent is and how it’s different from a simple assistant like ChatGPT. The idea is that you can separate the hype from reality and, as a bonus, impress everyone at your next family dinner.

Let’s begin.

What is an AI agent?

When I think of an agent, the first thing that comes to mind is Agent Smith, from The Matrix. If you haven’t seen The Matrix... well, shame on you!
Please stop reading this immediately and I’ll see you in 2 hours and 16 minutes, with your brand new perspective on the world.

OK, now we can continue.

Agent Smith had one single mission: maintain order inside the simulation and make sure the connected humans never questioned reality or discovered the truth about the Matrix. To accomplish that goal, Smith spent four movies chasing Neo.

In the real world, where agents don’t wear sunglasses, we define them as:
An entity that has a goal and acts autonomously to achieve it.

OK, so that’s what an agent is. An AI that has a goal and works on its own, making decisions step by step until it gets the job done. The good news is that its goal (for now) isn’t to chase us down... has anyone here seen Terminator?

There’s a big difference between an assistant, like ChatGPT, and an agent. While you have to guide your assistant step by step and tell it exactly what to do, an agent figures out on its own how to reach its goal.

Agents don’t wait for step-by-step instructions, you just tell them what the goal is and they figure out how to get there. They make decisions and move forward on their own until the mission is complete.

If we want to get technical for a second, we could say that an agent is a system that perceives its environment, acts autonomously, uses tools, and pursues goals.

You might be thinking something like: “But Germán, my assistant can already search the Internet, use tools, and even write code... so isn’t it an agent?”

Great question! No. 😅

These capabilities are part of what makes an agent, but assistants still don’t have the autonomy to plan and execute without human intervention until they reach their goal. Assistants still need you to guide them step by step.

An agent has four key characteristics (besides the sunglasses). Let me walk you through them.

The 4 things that make an agent

We already know that the key word here is autonomy. Now let’s look at what one of these agents needs to achieve it.

There are four things that make an agent truly an agent (and not just someone hyping up their AI). Let’s go through them one by one.

1. It has a goal and doesn’t stop until it’s done

This is the main characteristic of an agent: it has a specific mission and will work until it completes it.

Sound familiar? That’s basically the plot of every apocalyptic sci-fi movie you’ve ever seen!

While your assistant answers random questions or carries on a conversation with you, an agent receives a mission and understands it has a goal to achieve. It’s not trying to give you an answer, an agent will do whatever it takes to complete their mission.

When you give an agent a goal, you don’t need to tell it what steps to take. It figures out on its own what needs to be done, the sub-goals it has to accomplish, and how each step brings it closer to its target.

Of course, the agent needs to know when it has completed the goal you set for it. So it keeps evaluating its progress toward the target until it recognizes it has reached judgment day “final state”.

2. Observe, think, act... repeat

OK, an agent doesn’t follow a fixed plan. The real world is a complex scenario and you need to be able to adapt and respond.

As every human knows, having a fixed plan and not being able to adjust along the way is a recipe for disaster. We are flexible and machines are not, so we need to find a way to give them that flexibility.

That’s why agents work with an action cycle:

1. Observe: assess the current situation.

2. Think: based on what it observed, decide what the next action should be.

3. Act: execute that action.

And then back to step 1. This repeats over and over until the goal is met. Notice how this cycle allows our agent to adapt to the situation at every step.

Just to be clear, this is not the same as an automation. In an automation you program the steps to follow, so if something changes, the automation will probably fail. Agents can evaluate the situation and adapt to it.

This doesn’t mean that automations are bad, not at all. Actually, I think they’re an essential tool. I just wanted you to know that an automation is not the same as an agent.

automation ≠ agent.

In short, an agent has the ability to iterate over and over, adapt, and adjust its plan on the fly.

3. It remembers

None of what I’ve told you so far works if an agent doesn’t have memory. Agents need to remember what they’ve been doing. They need to know what actions they’ve already taken, which ones worked (and which didn’t), and how they’re progressing toward their goal.

Can you imagine Terminator forgetting about finding Sarah Connor halfway through the movie? Or Agent Smith not being able to remember that Neo’s name in the Matrix is Mr. Anderson?

Without memory, every “Observe-Think-Act” cycle would start from scratch. Our agent would have no idea that what it’s trying has already been attempted three times before and never worked. The fact that it can remember is what makes this truly a cycle.

This memory is not the same as the one your ChatGPT has, for example. An agent is much more structured and keeps track of its progress toward its goal. If we want to get technical, this is called maintaining “state.”

Bottom line: if it doesn’t have memory, it’s not an agent.

4. It has tools

Don’t go thinking that AI can grab its toolbox and finish that project you’ve had half-done for two and a half years (you don’t have one? I have several 😅).

What I’m talking about here is that our beloved agent can interact with other systems. For example, search the internet, run code, query databases, read and create files, talk to other systems (in nerd speak, that’s called calling APIs), etc.

Having a goal, a cycle that helps you adapt, and memory is not enough. These tools are what allow the agent to work in the real world. What I mean is that tools are what actually give it the ability to “do things.”

Right now, assistants like ChatGPT or Claude already have access to tools. The big difference with an agent is that they decide when to use those tools without human intervention.

An important note: you should always keep in mind what tools these agents have access to. You don’t want to “accidentally” and “without realizing it” give them the ability to email your boss, buy whatever they want on the Internet, decide where to invest your retirement savings, or control your nuclear arsenal.
Anyone said Skynet?

Sorry for the pessimism, I think I’ve watched too many sci-fi movies. But seriously, I do believe the tools part requires human oversight.

The reality is that if an agent makes good use of these tools, or gives us a heads up and asks for permission before doing so, it would be extremely useful.

So, to sum it up:

Agent = Goals + Cycle + Memory + Tools.

Let's get our terms straight

OK, now that we’ve seen what an agent is, I imagine you’re thinking something like:

A big part of this confusion comes from the fact that all most companies using AI want to say they have “agents.” Obviously, saying that sounds more pro, more cool, “más chévere” (as we say in Perú). You know what I mean.

Let me tell you that (if you have that question) I totally get it. In fact, I don’t think I have the definitive answer to that question, but I can try.

Here’s a quick rundown of what I understand by each of these terms (there’s no technical standardization that I know of, so you’re reading my opinion here).

Chatbots

The word chatbot sounds super outdated, and in a way, from a technology standpoint, it is. It’s a very basic conversation model that literally follows a script and can’t go off it.

Some people call ChatGPT, Claude, Gemini and the gang chatbots, but I see them as something more advanced.

Personally, I use the word chatbot for those automated systems, like the customer service ones that follow a script and are very easy to hate.

Assistants

This is where assistants and copilots come in. The ones we use every day: Claude, ChatGPT, Copilot, etc. These are AI systems that work with you, suggest things, you have conversations with them, and they help you create documents, code, or content.

While they are very advanced, useful, and improve our productivity by 1294.37% (it’s true, that data comes from a Stanford study... a chatbot told me :P), you still have to tell assistants what to do.

Automations / RPA

If you’re wondering what RPA is, it stands for Robotic Process Automation. It’s nothing more than using software robots to automate actions a human would do on a screen (clicking, filling in data, copy/paste, etc). They’re used to speed up and improve the consistency of repetitive (and boring) tasks that we humans do.

An automation is something much more general. It’s about using software or AI to execute processes without human intervention.

These technologies are great for performing tasks and processes that are very structured and don’t change. In the real world, many of these systems combine both and use automations for repetitive tasks and agents for complex decisions.

But keep in mind that the fundamental difference still holds. An automation will execute the rules and conditionals you programmed (which can be super complex), an agent decides on its own what to do.

Agents don’t (always) work alone

I may have given you the idea that an agent always works alone, but that’s not the case. We’re increasingly seeing how agents can work with other agents.

Imagine a group of agents working toward a common mission. It reminds me of Mission: Impossible.

What? Were you expecting Tom Cruise? This is the original cast from the TV show!!! Each agent had special skills that allowed the whole group to take on missions that would be impossible for a single agent.

Well... there’s James Bond, but he’s the exception.

Don’t tell me you were expecting a different Bond? That’s not up for debate. Connery is James Bond. Period.

Sorry I got distracted… let’s keep going. I was about to tell you that agents can work with other agents.

Multi-agent collaboration

It’s exactly what you’re imagining. Several agents, each with their own capabilities and tools, working together toward a common goal (Mission: Impossible style).

This makes all the sense in the world. Instead of having one single agent that does everything, we have several agents where each one focuses on solving one part of the problem. That way, each agent can have its own goals, cycle, memory, and tools designed specifically for its mission and on top of that... agents can communicate with each other!!!

So we have a group of agents that share a common goal, each with special skills, that can talk to each other and share information.

And the advantage here is not just having specialized agents. We can also put many agents together to solve a big problem, or we can have some agents review and verify the work of others. The possibilities are endless.

On top of that, these agents can be organized in different ways. Sometimes there’s even a coordinator agent, a sort of orchestra conductor, a leader... a John ‘Hannibal’ Smith of the digital world.

I’m sure you’ve seen The A-Team, and no, Liam Neeson and Bradley Cooper are not in this one. I went off on a tangent again. Sorry, I’ll try to stop with the old TV show references.

Where we are

Coordinating a group of agents is a complex task. They need to be able to communicate, avoid conflicts, divide the work properly, figure out what to do when they have contradicting information, and so many other things I can’t even imagine.

These multi-agent systems already exist and are a very active area of research and technological development. Universities, startups, and billion-dollar companies are working on how a group of agents can work together, plan complex tasks, and even correct each other. And all of this is happening as we speak.

This kind of research is necessary because most of the problems humanity is trying to solve are so big that a single agent just isn’t enough.

I can’t wait to see what this technology does in the next five years.

Hasta la vista, baby

If you’re reading this, thanks for making it this far. I know it was a long post 😬

Now you’ll not only be able to show off your AI agents knowledge at your next family dinner, you’ll also be able to tell when something is real and when it’s hype from the marketing team (or the CEO) of an AI company.

It’s a strange moment in the history of technology, and I’m sure what I’m about to say is a cliché, but it feels like we’re living in a sci-fi movie.

In the coming years we’re going to see how some of these agents will be incredibly useful and how others will fail miserably, with consequences we can’t predict today.

I think that’s partly why I write this blog, because we’re using a technology every day that we don’t fully understand, and we keep giving it more capabilities and power. And, sci-fi jokes aside, I’ve never liked giving control to something I don’t understand.

But mainly I write because if there’s something I’m passionate about, it’s learning. I want to deeply understand how AI works (and be able to explain it to you, in my own way), not just to use it better, but to use it consciously.

And, let me add…

This final note is something I had to write, a kind of catharsis. Thanks for reading it.

Since I don’t like ending a post on that note, let me introduce you to one last agent. 😅 Agent 86 from Get Smart. (You thought this was over with The A-Team? Ha!).

Just to be clear, this is Don Adams (not Steve Carell).

This show does go way back, it’s from 1965. I used to watch reruns of reruns in the 80s and early 90s with my dad, and we’d crack up laughing together. Thinking about that brought back some really nice memories.

Here’s a mini-preview, in case you’re curious.

Alright, that’s it for real now.

See you soon!

G

“This text was generated by AI” is what our professor told us, while highlighting a couple of paragraphs that I had written for a university group assignment.

Technically, what the professor was telling us was that Turnitin (an academic review tool) had flagged those paragraphs as probably written by AI.

Imagine my shock. There I was, in front of my classmates, trying to explain why those paragraphs, which I wrote with care and dedication were being treated as machine-made. How do you even prove that you actually wrote something?

Luckily, I know a thing or two about how AI works 😉, and we ended up having a super interesting conversation about these detection tools, how they work, and why they’re not 100% reliable.

Then I started to think:

I had been accused by a machine of not being human.

So I went down the rabbit hole to understand how these detectors actually work, how they analyze text, and how reliable they really are. And that’s what I want to talk about today.

Note: Wait, university?
Yep. For the past couple of years I’ve been studying Economics Engineering. What can I say, I’m 46 and I never stop learning new things :)

Back to the topic:

I had been accused by a machine of not being human.

Can AI-generated text be detected?

it’s complicated

All of our AI assistants have been trained on millions and millions of human-written texts with the only purpose of imitating the way we write. That’s literally their job. On top of that, every new version of ChatGPT, Claude, Gemini, etc., writes in a more “human-like” way, which makes me wonder if we can really detect whether something was written by a machine or a person.

There’s a whole industry around this. Universities and companies are looking for ways to identify AI-generated content, so many millions have been invested in building solutions that try to detect it. I did some reading and found five main methods used to try to solve this problem.

Let me spoil the surprise by telling you that none of these methods are 100% reliable, and that’s exactly why it is important to understand how they work.

Perplexity

I’m not talking about the search engine. When I say perplexity, I’m talking about a way to measure how “predictable” or “surprising” a text is for an LLM. This is the most common method used to try to detect whether a text was written by AI or by a human.

For example, imagine I ask you to complete this sentence:

“The sky is ________”

I bet the first thing that came to mind was “blue,” which is the most common continuation and the one a language model would write. That means it has low perplexity, or in plain English, “that surprises no one.”

I asked ChatGPT and it gave me these answers:

• “The sky is blue”

• “The sky is infinite”

• “The sky is beautiful”

• “The sky is the limit”

All pretty predictable, so according to this method, they were more likely written by a machine.

But what if the answers were:

• “The sky is a government invention to hide the aliens”

• “The sky is eating cau cau in bed while binge-watching Netflix” (don’t judge me)

(BTW. Cau cau is a traditional Peruvian tripe and potato stew, and my favorite dish in the whole world, and yes I said tripe).

Let’s say those are a tiny bit harder to predict, meaning they have high perplexity.

Some detectors use this method because AI-generated texts tend to have low perplexity. AI usually picks the “safest” words when writing, which makes it predictable. Humans are supposed to be less structured, some more than others :P, and therefore less predictable.

The problem with this method is that clear, direct, well-structured texts (you know, the way we're supposed to write) are more predictable than when, I don't know, you include cau cau, 80s TV shows, and other random stuff in your documents.

So the better you write, the more likely they’ll think you’re a robot.

But perplexity isn’t just about words. These tools also measure how the structure of your sentences changes, and that’s called “burstiness.”

Humans write varied sentences. Maybe you start with a complex one, then a simple one, then a joke... you get me. Robots tend to have a more uniform (and predictable) structure.

So these tools use both factors to try to decide how likely it is that a text was written by AI. Does it work? Sometimes. Is it reliable? Nope.

Style analysis

Each one of us (including machines) has a way of writing. The analysis of these writing styles is called stylometry. Think of it as the fingerprint of our texts.

For example, check out these short phrases criticizing AI detectors, each from a different author:

• “AI detectors are fundamentally unreliable tools that often unfairly penalize legitimate human writing due to their high false positive rates.” (Gemini)

• “AI detectors fail too often to be considered reliable tools.” (ChatGPT)

• “AI detectors are notoriously inaccurate, unfairly penalizing students and professionals with false positives while the real cheaters easily learn to evade them.” (Claude)

• “[outraged] Now machines are the ones telling us if we’re humans! Don’t believe them [/outraged]” (Germán)

All of these phrases say more or less the same thing, but each one has its own style, and this analysis looks for patterns in writing. How different are the sentences in the text, how many common words are used, what the punctuation looks like, etc.

If you’re wondering how style can be analyzed, I found a really interesting paper: “Distinguishing AI-Generated and Human-Written Text Through Psycholinguistic Analysis“ that identifies 29 characteristics.

Here are the six categories of those characteristics:

• Lexical: measure vocabulary and word variety.

• Syntactic: analyze sentence structure and complexity.

• Sentiment and subjectivity: measure the emotion and opinion in the text.

• Readability: measure how easy it is to read and understand the text.

• Named entities: analyze references to people, places, dates, etc.

• Uniqueness and variety: evaluate how original the text is and whether it uses different structures.

Detectors look for patterns that seem “suspicious” to them.

For example: some look for the use of certain punctuation marks (did someone say em dash “—”?), others look for words that are too sophisticated or too simple; they also check if sentences are roughly the same length.

One of the criticisms of this method is that it can be unfair to non-native speakers (like myself), who tend to write in a more direct, simple style that matches some of the “suspicious patterns” these AI detectors flag.

Machine learning

What if we use AI to detect AI? Great idea! using AI to tell us if something is human... I’m pretty sure that’s how Terminator started.

To do this, we train machine learning models with thousands and thousands of texts, each one labeled as written by a human or a machine. With that training, the model learns to recognize the patterns of each, and when it’s done, the model is ready to try to classify any text you give it and tell you whether it thinks it was written by a human or an AI (and I cannot stress the word try enough).

Sounds good, right?

Well, the reality is that it doesn’t work that well. For example, in 2023 OpenAI launched a classifier that could correctly identify AI-generated text only 26% of the time. And to make things even uglier, it told 9% of flesh-and-blood humans that their texts were AI-generated. This tool worked so poorly that they ended up shutting it down after just a few months.

There are other tools, but they all suffer from the same issues. They need a minimum amount (around 20%) of AI-generated content in a document to classify it correctly. They’re also designed to try to avoid accusing a human of being a machine (even though it happened to me), and this means they let a lot of AI-generated content slip through.

On top of all that, there’s a deeper problem. This technology only works for detecting the type of text it was trained on. Meaning, if we train a detector with texts created by GPT-3, there’s no guarantee it will still work with GPT-4. And if you train it with English content, it might fail in Spanish. And (as you can imagine) every time a new model comes out, these tools become obsolete.

And what about trying…

Watermarks

Yes, watermarks... like on bills. You still remember paper money, right?

Maybe today you only use digital wallets and cash is a thing of the past. Although there’s something special about carrying cash (especially when your phone runs out of battery).

Anyway, it turns out that to prevent people from counterfeiting bills, someone invented something called watermarks. They’re designs added to the paper during manufacturing that you can only see when you hold them up to the light.

Well, that’s exactly what AI companies can do: mark their texts at the factory.

Meaning, they can include an “invisible signature” in the text they generate. To do this, their language models have to write in a specific way that can’t be detected by a human reading the text, but can be detected by tools that know that specific writing pattern.

And this, in theory, is an almost perfect solution. If we had these watermarks in the text, you could know for sure whether it was written by a robot or a human. There’s just one small detail...

For this to work, all companies would need to add these watermarks to their language models. I find it very hard to believe that Google, OpenAI, Anthropic, Meta, Mistral, DeepSeek, Alibaba, etc., etc., would all agree and implement it. And let’s not even talk about open source language models.

Note: I imagine that for many of these companies, it’s not very convenient for people to easily detect that a text was written by an LLM.

Semantic analysis

This method is pretty interesting. In simple terms, it’s a way of asking an AI if it was the one who wrote something.

Basically, you take a text, try to figure out which LLM could have written it, and then use that same LLM to see how likely it is that it’s the author.

How do you do that? you might ask.

You generate small variations of the text and check if the AI is more likely to have written any of the variations than the original. If ChatGPT says all variations are less likely than the original version, it’s almost certain the original was written by ChatGPT. Make sense?

Let me give you an example. Say you want to check if this text is from a human or an AI:

“I love callos a la madrileña, but I prefer cau cau.”

(callos a la madrileña is a traditional Spanish tripe stew, and cau cau is its Peruvian cousin. Yes, I have strong opinions about tripe stews.)

The first thing I need to figure out is which of all the AIs could have generated this text. Let’s say I pick ChatGPT.

OK, so now I need to create some variations and “ask” ChatGPT how likely it is that it wrote all of them (plus the original).

Let’s say I give it the original and some variations and ChatGPT gives me the probability of having written each one:

• “I love callos a la madrileña, but I prefer cau cau.” (prob: 70%)

• “I’m fascinated by callos a la madrileña, although I prefer cau cau.” (prob: 85%)

• “I’m enthusiastic about callos a la madrileña, but I lean towards cau cau.” (prob: 72%)

• “I’m seduced by callos a la madrileña, but I hold a greater predilection for cau cau.” (prob: 25%)

See how some variations are more likely than the original text I want to compare? With this, I could say it’s unlikely that this text was written by ChatGPT.

But, now I’d have to test with the gazillion other LLMs out there. Can you imagine doing this whole process for each one of them? Where I come from we call that “trabajo de hormiga” (ant’s work, meaning painfully tedious). And let’s not even talk about how much running all those tests would cost.

Note: Of course, I’m oversimplifying the process and the math behind this. If you want to learn more, check out the paper: “DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature“

What happens when you write in Spanish?

Turns out all the methods I just showed you work even worse in Spanish.

This is because all these detectors have been trained primarily with English texts, probably 80% of all their training data. The remaining 20% is for all the other languages in the world... so yeah, I’m screwed.

To give you an idea, when OpenAI launched their detection tool they literally said “We recommend using the classifier only for English text. Its performance is significantly worse in other languages.”

Sí señor.

Are we really letting machines decide if we’re human?

Think about it for a second. We trained artificial intelligence using millions and millions of human-written texts until we finally got a machine to write like us. Now, we can no longer tell the difference between text written by robots and by human beings, and that’s a problem.

So, to solve this problem, we’re building new machines so that they can analyze the text and tell us whether it was created by humans or not. Are you sure that’s what we want? Don’t you think we’re giving Skynet too much power?

As for me, I hope I don’t get accused of being an AI again in the future ;)

Best,
G (the human)

If you’ve spent more than twenty minutes reading about AI, I’m sure you’ve seen this type of prompt floating around, more than once:

“Act as a __________ and help me ______________________”

You’ve probably used it yourself.

Well, this kind of recipe is what we call a prompting framework, and they’re nothing more than structured ways of asking questions to ChatGPT, Claude, or your favorite AI assistant.

There are so many of these frameworks that covering them all in a single post would be way too boring long, so I figured I’d make this a series. Today we’re looking at RTF and RAILS, the basics (but don’t let that fool you, they’re still useful). I’ll walk you through the rest in future posts.

RTF: Role, Task, Format

I’m not going to start with a joke about how every time I see RTF I think of WTF... (oops). OK fine, I’ll be honest, instead of thinking Role, Task, Format, my brain immediately goes to “Role, The, Fu##” (but that’s just me, right?).

This framework showed up in 2023, when ChatGPT was only a few months old and we were all figuring out how to use it. It didn’t come from OpenAI, a university, or a research paper. It came from users experimenting and sharing what worked for them.

How does RTF work?

The idea is that your prompt should have three elements:

Role: You tell your AI who you want it to be. “You’re a Marketing expert,” “act as a Peruvian chef,” “imagine you’re a college professor.” This sets the context and tone for the response.

Task: This one’s easy. Just tell it what you want it to do. Try to be specific, this is the heart of your prompt. Don’t say “tell me what to make for lunch.” Try something like “I have 15 minutes to cook, I’ve got leftover rice from yesterday, scallions, soy sauce, and some leftover meat.” If your assistant is any good, it’ll suggest you make arroz chaufa (Peruvian-style fried rice. Trust me, it’s amazing).

Format: Here you tell it how you want the answer. A list? Three paragraphs? A table? You get the idea.

Remember, it’s not enough to tell your assistant it’s “an expert in X.” You need to be clear about what you actually need from it. That alone will make a huge difference in the responses you get.

I don’t personally use RTF (at least not consciously), but when I look at my prompts I realize some of them follow that structure. So I can tell you it works.

From 2023 to 2026

This framework was created about 3 years ago, which is 21 in dog years and roughly 178 in AI years. So we could say it’s old a classic.

It still works great with today’s language models like ChatGPT 5 and Claude 4, so don’t be afraid to use it.

The example

“You are a personal finance expert….” hmmmm…

No! Let’s find a more fun example… how about…

“You are Cantinflas. Explain how ChatGPT works, but do it in your style: with lots of rambling, long sentences, tangles of words, and funny contradictions, so that it seems like you’re explaining a lot but you’re actually saying nothing concrete. Keep it to three or four lines max.”

Note: If you’re not familiar with Cantinflas, Charlie Chaplin himself called him “the greatest comedian alive.” He’s often referred to as Mexico’s Charlie Chaplin, a master of talking endlessly while saying absolutely nothing. It’s an art form, really.

Here’s ChatGPT’s response (I’m keeping it in Spanish because Cantinflas’ magic is inseparable from the language. Basically, ChatGPT rambles on and on about how ChatGPT “knows everything and nothing at the same time” and how “it answers what it thinks you meant to say, even though that’s not what you said,” ending with “between yes and no, well, there we go.” Classic Cantinflas):

¡Ah pues mire usté! El ChatGPT viene siendo como un cuate que sabe mucho sin saber nada, porque sabe de todo y de nada al mismo tiempo; entonces cuando uno le pregunta, él contesta lo que cree que quiere decir lo que usted dijo, aunque a veces no es lo que usted quiso decir, ¿me explico? Porque en resumidas cuentas, el asunto es que se arma una plática muy platicada donde parece que hablamos claro, pero entre que sí y que no, pos total, ahí la llevamos.

Not bad, but who could ever truly imitate the great Mario Moreno? Nobody.

Let’s move on to the next framework!

RAILS: Role, Action, Input, Logic, Style

Is it just me, or does it feel like someone is trying WAY TOO HARD to give their frameworks a cool name? This is another one from 2023, also created by the community.

How does RAILS work?

Unlike WTF RTF, which has only 3 elements, RAILS has 5 (though the “S” is optional).

Role: Same as RTF. Who do you want your AI to be? “You’re an expert crisis communications manager,” “Act as a ninja obsessed with conspiracy theories,” or whatever you prefer.

Action: It’s the same as the task in RTF (except if they had called it “task,” the acronym wouldn’t have sounded nearly as cool). Just tell your assistant , in detail, what you want it to do.

Input: This is where the framework gets demanding. It doesn’t assume your assistant has all the information, it asks you to provide the data it needs. For example, last quarter’s sales figures, the comments you get on your blog, your ex’s WhatsApp messages, etc.

Logic: This part of the prompt is for detailing how you want it to think. Things like “explain your reasoning,” “think step by step,” etc.

Style: What tone do you want the response in? Do you prefer long and formal answers, or something more conversational? This is where you define that.

Does it work with modern LLMs?

Yes. The logic component is actually really interesting when it comes to ChatGPT 5. This latest version of OpenAI’s assistant can decide on its own whether to reason or not before giving a response. With RAILS, we make sure it always uses the same way of thinking.

Just like RTF, I don’t use RAILS in my day to day either😱. What I do take away from these frameworks is that they help us understand how to structure our messages to an LLM.

An example

Prompt:

Act as a Chef specialized in Peruvian cuisine. Your task is to give me a recipe for a dish I can prepare in 15 minutes. Ingredients: Leftover rice from yesterday, meat leftovers, scallions, egg, soy sauce. Explain your reasoning step by step. Give a concise and conversational response.

As I mentioned, this prompt made my assistant stop and think before responding….

… and guess what?

It gave me its recipe for Chaufa!!! (Peruvian fried rice)

ChatGPT:

Let’s make a quick chaufa with what you’ve got!…

Sorry, I had to test if it would actually say that 😅

Wait, you don’t use these frameworks?

OK, ok… I can imagine you’re thinking why I’m writing a post about tools I don’t use. Fair enough.

To be honest, if you look at my prompts you’ll probably see some of the elements from these recipes. Just not with the exact structure.

I think these kinds of tools are designed to get really good responses from our assistants, but my day to day is a little different. More than a question-and-answer thing, I try to have a conversation.

It’s similar to how I cook. I never follow a recipe to the letter, I just keep adding ingredients, tasting, and trying to be creative (accompanied by a bottle of wine, a baguette, and maybe some cheese).

It usually works, although there’s an inedible butter fish that my kids will never let me forget.

Let me put it this way: after a couple of years talking to AI every single day, you get the hang of it.

And that’s the thing. If you’re just starting out (or if your brain just prefers structure), I’m sure these frameworks will be really useful.

Let’s get serious for a second

I’ve seen a lot of prompts like: “You are a Harvard MBA…” or “You are the world’s greatest expert on topic X.”

Please, don’t think this magically turns AI into a genius. These frameworks work because they help you organize your ideas and present them clearly and in a structured way to your AI. They don’t give ChatGPT superpowers.

What’s next

As I said at the beginning, this is a series of posts. We still need to talk about CRISP, CLEAR, PEACE, COSTAR, and more.

And since I mentioned Cantinflas, in mine and Chaplin’s opinion the greatest comedian of all time, here’s a clip from the movie “Su Excelencia” (Your Excellency), which has what I think is one of the most beautiful speeches ever written.

The movie is from 1967, and in it Cantinflas plays a diplomat who ends up delivering a powerful speech about peace, freedom, and the dignity of small nations. It’s funny, it’s moving, it’s perfect, and it’s in Spanish.

I have beautiful memories of sitting and watching those movies with my dad ❤️

No matter how old you are, if you haven’t seen it, I really recommend you look it up. It’s a masterpiece. Please set audio to be in Spanish (AI cannot auto-dub Cantinflas) and use captions in English.

Hope you enjoyed the post, see you next time!
G

I’m going to assume you’ve seen Karate Kid (1984). I’ve seen it about 758 times. In the movie there are two karate schools with philosophies that couldn’t be more different.

There’s Cobra Kai, where the sensei screams “Strike first, strike hard, no mercy” while teaching you bone-breaking techniques from day one.

On the other end there’s Miyagi-Do, where an old man has you doing household chores like waxing (and buffing) his classic car collection, painting fences, or sanding the floor before teaching you a single move.

The thing is, I started thinking about it, and I believe something very similar is happening in the world of AI.

Let me tell you about it...

Note: Karate Kid is one of my favorite movies. Parts II and III are also good, Part IV is terrible. Also, the Cobra Kai series (Netflix) is an amazing show.

The two schools

As I was saying at the beginning of this post, in the world of generative AI I see two types of messages almost every day, and they couldn’t be more different from each other.

On one side, there are some voices trying to explain how this technology actually works. They talk about LLMs, agents, context windows, tokenization, and all that good stuff... you could say it’s a somewhat technical perspective.

On the other side, there are voices dedicated to giving specific (and often magical) solutions to some type of problem using AI. This is where all the variations of “use this prompt to [insert complex problem] and forget about [insert user’s pain]“ come in.

Cobra Kai

Here you learn almost exclusively to attack (no mercy), they teach you techniques that work from the moment you step into the dojo. After a short time, you feel like you can take on anyone — you already have an arsenal of techniques under your belt.

In the AI world, this looks like:

• “Use this prompt and replace your accountant”

• “The only technique you need to know”

• “Copy and paste this prompt and save 15 hours a week”

You have the exact formula, you use it and (sometimes) it works. You feel like a champion and start sharing your “secret prompts” on LinkedIn, you’re almost convinced you’ve mastered AI... until you run into the crane technique a case that doesn’t fit the formula and you have no idea how to adjust the prompt.

But you don't worry too much, because you can always find the next magic prompt, and the truth is that most of them work pretty well (although none of them can replace a team, no matter what Sam Altman says).

Miyagi-Do

In this dojo they teach you to understand the concepts behind the technology and encourage you to experiment, but the reality is that you came here to learn how to fight use AI, not to get a PhD in data science.

It goes something like this:

You’ve spent the entire afternoon trying to get ChatGPT or Claude to understand what you need, you’ve tweaked the prompt seventeen times. You finally got what you wanted, and even though you’re not 100% sure why, you learned something about how to structure context. Next time will be easier (or at least that’s what they tell you).

It can be frustrating, and sometimes you feel like you’re wasting your time.

It’s like someone promising to teach you karate and immediately putting you to wax and buff their four classic cars in a very specific way, without explaining why.

It's not something you can apply right away, it's not sexy, you can't go brag on LinkedIn about how great you are at... waxing cars?... But after a good while (if you didn't give up along the way) you no longer need to search for the perfect prompt on the Internet — you can build it yourself.

What’s at stake

In Karate Kid, what’s at stake is honor, life, and the trophy at the All Valley Under 18 Karate Championship. Here we have something at stake that’s a bit less dramatic:

The mindset with which we’re approaching AI.

Let’s go back for a minute to Karate Kid III. This is where Terry Silver shows up (one of the best villains of the series) and his Quicksilver method, an intensive training system with just three rules.

That’s right, just three rules (although they’re not exactly “sportsmanlike”), this method is effective and delivers results fast. Terry wants you to attack without thinking... and doing things without thinking is generally a problem.

Something similar happens when we start copying and pasting prompts left and right. I have no doubt that those prompts work and that they solve the problem right in front of us, but I believe using them without understanding how (and why) they work limits us enormously — how do you think things turned out for Terry Silver in the movie?

I don’t think using these prompts is terrible, what worries me is only using those prompts. Relying on a single technique to solve every problem is something that simply doesn’t work...

hmmmm... could there be an example of that in Karate Kid?

Of course there is!

Enter Karate Kid II, where Daniel faces problems he hadn’t encountered before and innocently tries to use the crane technique against Chozen... guess what happened? It didn’t work and he almost got his teeth knocked out.

What I’m trying to say is that the prompts you find out there work. But what happens when we run into a problem that doesn’t have a “prompt to copy and paste”? What about when you want to adapt AI to the way you work? Or when some new capability drops (and in the AI world, that happens every 10 minutes)?

Well, you’d have to wait for someone else to develop a prompt so you can copy/paste it. Unacceptable (just like cheating at the 1984 All Valley Under 18 Karate Tournament).

This technology is redefining how we work, how we create, and I’d even say it’s redefining how we think.

We need to understand how it works, and I’m not talking about becoming AI experts, but about understanding at least a little bit of what goes on behind the scenes when we interact with AI.

Finding balance

Reality is always more complex than a movie, so this isn’t about choosing between Cobra Kai and Miyagi-Do, but about knowing how to combine both.

Using specific prompts, formulas, and techniques you find on the Internet is perfectly fine — it’s actually very useful. If someone already figured out how to solve something, why not take advantage of it?

The problem is not using other people’s techniques. The problem is ONLY using other people’s techniques.

The ideal is finding a balance between techniques that work and that you can apply almost immediately, and understanding the fundamentals so you can create solutions to your own problems.

As Mr. Miyagi said:

Balance not just for karate, but for whole life.

There’s one more thing... The Internet is packed with Cobra Kai-style content, which is very easy to find and consume. I see it every day, from LinkedIn posts to TikTok videos. The reality is that many of those prompts aren’t as wonderful as they claim to be, or worse, they don’t work at all.

I’ve lost count of how many times I’ve read some version of: “with this prompt replace a consulting team.” That doesn’t exist.

As Johnny Lawrence (now a responsible adult) would say: Bullshit.

If you’ve watched the Cobra Kai series, you’ll understand a bit more about balance. Here our beloved Johnny Lawrence, [spoiler] manages to balance his dojo’s aggressiveness with Miyagi’s philosophy. All without giving up his sleeveless black gi (badass!).[/spoiler].

It took Johnny more than 30 years to find that balance (we don’t have to wait that long).

About this newsletter

This newsletter is on Mr. Miyagi’s side.

I don’t consider myself an AI expert, I’m just someone who loves to learn. So I’m learning how this technology works and sharing it here with you (although my mom does say I’m an AI expert... and who am I to contradict her).

There’s a ton of content out there about “the best prompts for X,” and not nearly as much that tries to explain how AI actually works. I originally started doing this in Spanish with my newsletter aprendiendoIA, and almost nobody does it with Karate Kid memes and references to Peruvian food ;)

If you’ve been reading me for a while, you may have noticed that I avoid telling you when to use a specific technique or giving very specific examples. Why give you a Marketing example when I can give you one that includes pan con chicharrón (a legendary Peruvian sandwich)?

I do this for two reasons:

1. Because you know better than I do where you can apply AI.

2. Because it’s fun to talk about 80s shows, tamales, mondongo, or Karate Kid while trying to explain AI concepts.

This space isn’t called My AI Journey because I want to teach you (I’m no Miyagi), but because I want us to learn together.

To wrap up

Fun fact: I wrote the original version of this post on 11/11, which happens to be my birthday. So I gave myself permission to write something a bit more personal — I hope you enjoyed it.

I’ll also tell you that I studied karate as a kid for several years and I always wanted to have a sleeveless black gi, Cobra Kai style. Which by the way, has the coolest name for a dojo ever. Maybe it’s because I grew up in the 80s and 90s, when being the toughest kid on the block was all the rage?

Finally, I know that learning can be scary, and sometimes it hurts to see that something isn’t working out, we can even feel defeated, but... you know what...

Fear does not exist in this dojo substack! Does it? Pain does not exist in this dojo substack! Does it? Defeat does not exist in this dojo substack! Does it?

G.

This is a post I’ve been wanting to write for a while, but the truth is that very few people used Claude, and even fewer listened when I talked about Anthropic’s assistant.

But things have changed, and now that Claude is trending (you know, the whole Pentagon stuff), I’m going to take advantage of the hype to write the ridiculously detailed guide to my favorite AI assistant.

This is going to be a long post, so crack open your first beer and let’s get started! 😉

What is Claude?

Claude is an AI assistant created by Anthropic. It works more or less like any other… except that, IMHO, it’s the best one out there.

You can use it on the web (claude.ai), though it also has desktop and smartphone apps. Go ahead and create an account so you can try things out as I walk you through this post. Don’t worry, the basic plan is free. Go on, I’ll wait.

Got it? Great, let’s keep going!

What Claude looks like

Now that you’re back (with your account, I assume), this is what you should be seeing on your screen. I’ll admit it’s not super impressive and it looks like any other AI assistant, just with better typography and color palette 😅.

Now let's go through this interface piece by piece. I promise there will be surprises (hopefully).

Let’s start with the chat

This is where we’ll spend most of our time talking to Claude, so it’s the perfect place to begin.

As I’m sure you know, this is the core of your interactions with your assistant. In the chat window you type your text and you can also drag files and photos into the conversation. You type, your assistant responds, that’s it.

The first thing you’ll notice is the difference in the responses. And that’s exactly what I love about it: it responds better. I’ve always found it hard to describe why (maybe it’s just me) but it feels more human. It’s an assistant that actually tries to help you, not just agree with everything you say like cough cough, ChatGPT, cough cough.
Try it and I think you’ll agree with me.

Let’s move on to the “+” button you see at the bottom left.

From top to bottom, you can:

Attach photos or files: I always do it by dragging them into the chat window, but I won’t judge you if you use this option. These files can’t be larger than 30MB each, and you can upload a maximum of 20 files per message.

When you upload a file, Claude can read it and even see the images and graphics inside it. But if it has more than 1,000 pages, Claude will only read the text. Solution: if your file has too many pages, split it up.

You can also take a screenshot and give it to Claude to analyze, no need to save the file in Paint first and then send it over (do people still do that?).

You can add the chat to a project… I’ll tell you later how those work, but if you’ve used them in ChatGPT, it’s pretty similar.

The research function lets you ask Claude to start investigating a topic. Claude will come up with a plan and ask you questions to better understand what you want. I love it because those questions come as a checklist-style form you can just click through. Once it has that information, it goes off and does the research across different sources (while you grab a coffee) and a few minutes later delivers it to you.

Web search can be toggled on and off from there. This one’s simple: if it’s on, it searches the Internet; if it’s off, it doesn’t.

You can also ask it to use a specific conversation style. By default it comes with Normal, Learning, Concise, Explanatory, Formal, and Informal styles. Don’t like any of them? No problem, you have the option to configure exactly how you want Claude to talk to you. To do it, all you need to do is give it examples of text showing how you want it to respond. That’s it.

Finally, there are connectors, which let you connect Claude to hundreds of external apps like Gmail, Google Docs, Google Calendar, Miro, Github, Slack, Canva, etc. It’s going to be hard not to find a connection to the tool you’re looking for. I love this feature. For example, the other day I connected it to a huge Miro board, asked it to make a summary and create a flow diagram. It did it without breaking a sweat. A-ma-zing.

Now let’s look at the bottom right. There you can, among other things, choose which language model you want to use.

I think it’s time to tell you that Claude doesn’t have just one model. It has three, and the models you’ll see here depend on your plan.

The most powerful one is called Opus, and it’s for when you need Claude to really think deeply. Use it for complex analysis, multi-layered problems, when you want it to consider multiple perspectives. The catch with Opus is that you can hit your usage limit pretty quickly (I’ll tell you more about that later).

The middle-ground model is called Sonnet, and it’s the one you’ll probably use most of the time. It responds pretty fast and is very smart. Not as much as Opus, but it makes up for it with speed. Use it to help you think, analyze documents, create content. Let’s say it’s the one that gets you through the day-to-day.

The other model is Haiku, and it’s the fastest. It responds at the speed of light and is perfect for quick, direct questions, classifying information, short summaries, and things like that.

Any of these models can work with that “extended thinking” switch. What it basically does is put your assistant in a more contemplative mode so it thinks more deeply before responding.

If you look at the bottom right, you’ll notice an icon that looks like an audio signal. That’s voice mode, where you can talk to Claude. I’ll be honest: I don’t love it. First because it doesn’t work in Spanish, and second because compared to ChatGPT’s voice mode implementation, it’s still in its infancy. Still, I maintain that Claude is the better assistant.

One last thing about the chat. All AI assistants have a limit on how much text they can “remember” from a conversation. This is called the context window, and when you hit the limit (which in Claude’s case is very large) the model has to make room to keep the conversation going.

What this assistant does is compact the conversation so you can continue. I love it because it also tells you it’s doing it, and that transparency about how it works is one of Anthropic’s strong points. Either way, remember that when a conversation gets summarized, you always lose a bit of context. That’s normal with this technology.

In practice, all of this means your conversations with Claude will be super fluid and productive 🙂

Although…

Sometimes Claude tells you to wait (AKA usage limits)

Before you run into the message “you’ve reached your usage limit, wait until 11pm to continue chatting with Claude”, let me tell you that it’s going to happen… and you’re going to hate it.

There’s nothing worse than being in a hyper-productive conversation with your assistant and out of nowhere it tells you that you’ve hit the usage limit and have to wait to keep using it. Well, that’s how Claude’s plans work (nothing is perfect in this world).

This is where the folks at Anthropic decided to create the most confusing pricing plans in the universe, so you’ll have to forgive me if I can’t explain them perfectly. Sometimes even I get confused, and I’ve been paying for it for a year and a half.

OK, all of Claude’s plans have usage limits. It doesn’t matter if you’re on the free plan or paying $20, $100, or $200 a month, you have a limit. Of course, the more you pay, the higher those limits.

It’s so complicated that there’s even a section in your settings to check how much you’ve used and how much you have left (here, let me show you).

Let me start by telling you that it’s not just one limit. There are several that depend on which models you’re using. Remember: Opus, Sonnet, and Haiku.

The first bar you see is “current session,” which basically shows the percentage of interactions you’ve had against the limit within a four-to-five-hour window. In my case, I’ve used 11% of my plan’s limit and it resets in 14 minutes. So I’ve got plenty of room to keep chatting. It’s happened to me several times that I’ve hit the limit in the middle of a conversation. That’s an unmistakable sign that it’s time to upgrade.

The other two bars are the weekly limits: the first gives you a usage limit that covers all models (Haiku, Sonnet, and Opus) and resets once a week. I’ve never hit that limit, and I use Claude like there’s no tomorrow. The second is a weekly limit for the Sonnet model specifically, which is the most versatile one (I’ve never hit that limit either).

Now let me tell you about the plans.

Free plan: perfect for seeing how it responds, though the limits are tight and you only have access to the Sonnet and Haiku models (you need to upgrade to use Opus).

Pro plan ($20/month): limits five times higher than the Free plan, plus you get access to the Opus model. It’s ideal for most professionals. I don’t think you’ll hit the limit very often. In fact, it’s the plan I used for over a year before switching to the Max plan.

Max 5x plan ($100/month): this plan has limits twenty-five times higher than the Free plan. It’s for intensive Claude use. Since I’ve been on it, I’ve never hit the limit.

Max 20x plan ($200/month): One hundred times the capacity of the Free plan, for people who are on another level in usage and wallet.

There are also plans for teams and enterprises, with admin controls and additional security layers.

My recommendation…

Seriously though, I recommend you start with the Free plan and if you like it, jump straight to Pro. If at some point you realize you need higher limits, you can think about switching to Max, though I’ll admit it’s not easy going from $20 to $100. Now, remember you can always go back to a lower plan, so if there’s a particular month where you need to use Claude more, you can bump up and bring it back down the next month.

One last option is to activate “extra usage.” Basically you give them your credit card so that when you hit the limit, they start charging you per use.

Now that you know the plans, you’re probably wondering…

How do we get Claude to know us?

This is important for Claude or for any other AI assistant. It’s impossible to get the most out of it if it doesn’t know who you are.

Think about it for a minute: this is a mass-market product that answers questions for millions of people every day. If it doesn’t know what makes you different from the rest of the users, what it’s going to give you is an average response. And for that, any assistant will do. Period.

Luckily, we have an option to tell it who we are and how we want it to interact with us. In Settings → General → Profile, you can tell it everything from what you want it to call you to your darkest desires who you are, what you know, and what you do.

On top of that, Claude has memory. Meaning it can remember things about you that you’ve told it in past conversations. To find the memory settings, go to Settings → Capabilities → Memory.

Claude’s memory comes from two places. The first comes from being able to search for details in past chats. You activate it with the first toggle you see above.

The second toggle has the more interesting memory function. It generates a memory document based on your chat history. Meaning it decides what’s most important from all the conversations you’ve had with it, makes a summary, and uses it as memory for future conversations with you. If at any point you want to edit that memory document, just click on “Memory from your chats” and you’re done. Plus, if you want Claude to remember something specific, just tell it and it’ll take care of adding it to its memory.

Also, since a lot of people are switching to Claude, the Anthropic team has developed a memory import tool so you can bring over data from other providers (cough, cough, ChatGPT, cough).

And if you’re worried about a conversation ending up in its memory, you have the option to have conversations in incognito mode. Just click on the little ghost icon at the top right of the chat window, and that’s it.

OK, now that Claude knows us, we can start getting the most out of it.

Projects

A project is a space where you can have focused chats, and all these chats share the same instructions, their own documents, and memory. To create one, go to Projects in the left sidebar, and after giving it a name you’ll need to do two things.

The first is to give it the project instructions. That means telling Claude how you want it to work when you talk to it (in chats within that specific project).

The other thing you need to do is create its knowledge base, which is a fancy way of saying you can upload documents to the project and Claude will have them at hand when responding.

To give you an idea of how cool projects are, imagine you’re a macroeconomics expert who loves making ceviche. If you used your assistant for everything, its memory might get a bit confused, don’t you think? That’s where projects come in.

You could have one project with all your macroeconomics knowledge and another where you upload several cookbooks and dedicate yourself to creating new ceviche recipes. By the way, the best ceviche is octopus, and in a close second place is grouper (we call it “mero” here in Perú).

Remember when we looked at the chat interface and there was a “+” menu? Well, there’s an “Add to project” option that lets you move any chat to a project and keep everything organized 🙂.

Plus, each project has its own memory, separate from the general memory, so what Claude learns about you in a project stays in the project (nobody wants octopus ceviche references in their macroeconomics chats, right?)

And if you liked projects, wait until you meet…

Skills

This is one of the most powerful features Claude has, but also one of the most hidden. A lot of people don’t even know it exists. Let me tell you: a skill is a way to teach Claude to do things it didn’t know how to do, permanently.

If you’ve seen The Matrix, you probably remember this scene:

Well, that’s more or less how I imagine it works :P (if you haven’t seen it, shame on you!).

Let me start by telling you that Claude already comes with some pre-installed skills that activate automatically.

For example, if you ask it to create a PowerPoint presentation, an Excel file, a nicely formatted Word doc, a PDF, etc., Claude does it and gives it to you ready to download. To do this, Anthropic’s assistant realizes, let’s say, that it needs to create an Excel file and activates its Excel creation skill. The same thing happens with the other file types.

But there’s more. Skills aren’t just for creating files. They’re also super useful for getting Claude to act a certain way or follow specific procedures. Meaning you could create a skill so Claude knows your analysis process, so it always uses the same format when giving you a ceviche recipe, or whatever you can think of. The sky’s the limit.

To create a skill you need to create a file called Skill.md with instructions in a specific format using YAML and blah, blah, blah… but the reality is that you don’t need to because Claude can do it for you!

You just need to tell Claude that you want to create a skill, and after a nice conversation, your new favorite assistant will have that skill ready to use whenever needed. For example, one of the skills I’ve made lets Claude create documents with the colors and typography of aprendiendoIA. Before having this skill, the documents Claude created were well formatted but didn’t have my design. Now it simply uses that format every time I ask for a document.

You can view and edit your skills in Customize → Skills in the sidebar.

And as if skills weren’t enough, let me introduce you to…

Artifacts

What if I told you that Claude can not only create documents but present them right there in the chat? And what if on top of that I told you it’s not just documents, but web pages, interactive charts, diagrams, and even fully functional web applications?

Well yes, that’s exactly what I’m telling you, and let me add that it not only lets you create them but publish them too! They’re called artifacts and they’re wonderful.

To create an artifact you first need to activate them in Settings → Capabilities → Visuals. Do it now because we’re about to dive into the beautiful world of Claude’s artifacts ;)

If you’ve already activated the feature, now we can start creating our first artifact… it’s true that I could start by telling you to make an interactive chart of macroeconomic indicators… but honestly, that’s not my style. Let’s make a drum machine instead!

All you have to do is tell Claude you want to make an artifact and describe what you want to create. In my case, I just said:

“I want to create an artifact. Let’s make a drum machine.”

With that, after a couple minutes of thinking and a couple of tweaks, we got this:

And it's not just for me. Claude lets me make the artifact public so everyone can enjoy it. So go ahead and give my beatbox made with Claude a try 😉.

On top of that, you can “remix” my app using Claude to create your own version.

But there’s another type of artifact you can create, and this one is even more powerful. Its nickname is Claudeception (like the movie Inception… which I hope you’ve seen… if not… shame on you!).

These artifacts are special because they’re applications that use artificial intelligence inside. They connect with Claude to work and do their thing in real time. Now you know why we call it Claudeception (it uses Claude inside Claude).

You could use it to create a language tutor your way, or a chatbot specialized in whatever topic you want (did someone say ceviche?). All without writing a single line of code… But why give examples when we can make one?

How about we make a crossword puzzle creator? You give the AI a topic and it gives you a crossword to play.

You can say something like:

“I want to make a Claude artifact that uses AI to create a crossword puzzle. The user gives it a topic and the AI creates it.”

I recommend using the Opus model for this. Claude will think for a few minutes, maybe ask you a couple questions, and before you know it you’ll be trying to solve an AI-created crossword puzzle.

Try it yourself here. Of course, you’ll need a Claude account to use it. Don’t worry, when you create and share one of these artifacts, the people who use it don’t consume your Claude usage quota. That’s why you need an account to view this type of artifact. I imagine if you’ve made it this far you already have one. Just keep in mind that using AI-powered artifacts counts against your plan’s limit, so don’t go overboard :P

Another interesting thing about all artifacts is something called persistent storage, which means you can use an artifact, close the browser, and what you did in that artifact stays saved.

One last thing about artifacts. Anthropic maintains a really nice catalog, give it a look.

When Claude leaves the chat window

So far we’ve seen a lot of what our new favorite assistant can do inside a chat. But there’s no way our AI assistants are going to stay trapped in a chat, right? That’s the basic premise of any apocalyptic sci-fi movie, and Claude can’t be left behind.

In this section I want to tell you about three things: how Claude connects with other apps, how it has extensions that work inside other programs like Chrome, Excel, or PowerPoint, and finally about something called Cowork, which lets Claude control your computer… Duuuun!

Let’s start with connectors: these allow Claude to interact with external apps like Gmail, Google Drive, Miro, Canva, Notion, Slack, and a couple hundred other apps. To connect your favorite apps, go to Customize → Connectors in the sidebar.

There you search for the app you want to connect to, grant access, and done. Now Claude can interact with that app. And don’t worry about Claude making a mess in your apps. Before taking any action, it explains what it’s going to do and asks for permission. Though if you want, you can configure trusted tools to work without needing your confirmation.

Now let’s talk about Claude’s extensions: right now Anthropic has an official extension for Chrome that turns it into an agent. You tell it what to do and Claude does it, in Chrome.

For example, I went to aprendiendoIA, my main substack (in Spanish), and asked it to open my post AI Crash Course, which by the way is one of the best posts I’ve written (and I’ll soon translate to English). Claude made the plan and asked me to confirm.

Then it opened the post I’d asked for. Of course that’s just an example. Imagine having Amazon open and asking it to create a shopping list. To be honest, I don’t use it. I use Safari, not Chrome (as any self-respecting Mac user does).

Apart from Chrome, there are also official extensions for Excel, PowerPoint, and Slack. Basically it puts a Claude window alongside you while you use those programs and lets the AI help you with your work. I was once fighting with an Excel model and ended up asking Claude to do it. It did a great job. In five minutes I had a spreadsheet ready that probably would have taken me a couple hours. We’re living in the future.

Now let’s talk about Cowork: to use it you need the Mac or Windows desktop app. That’s where you’ll find the Cowork tab.

With Cowork you can give Claude tasks and it gets to work on its own. Obviously, before doing anything it defines a plan, tells you what it’s thinking, and executes it. Of course it’ll ask for permission to access whatever folders it needs.

For example, one of the things I asked it to do was help me organize my chaotic Downloads folder (with thousands of files). It helped me classify them and delete the files I didn’t need. Another thing I did was give it access to the folder where I have the (very rough) drafts of a book I’m writing about AI and asked it to create a PowerPoint presentation. It did it without a hitch.

You can also schedule tasks to repeat daily, weekly, or, I don’t know, the third Saturday of every odd month. ¯\_(ツ)_/¯

Note: Yes. It’s true, I’m writing a book, and now that it’s public I think I’m going to have to finish and publish it. What did I get myself into?

Why does Claude feel different?

We’ve spent a good while talking about Claude (and I hope also using it). I imagine you’ve already noticed that its way of interacting with us is different from the rest of the assistants. I don’t have any other way to describe it than to say it feels “more human.”

This happens because Claude was trained with something Anthropic calls Constitutional AI.

Basically, they gave it a constitution (the kind with principles and values) and used that document to teach it how to respond. During some point in their model training, every time it generated a response, it compared it against that constitution, and if the response didn’t follow its principles and values, it changed it and learned along the way.

That way, by the end of training, the Anthropic team made sure their assistant’s responses were aligned with that document. This constitution has four priorities: safety, ethics, following Anthropic’s guidelines (for sensitive topics), and being helpful, and it applies them in that order of priority.

Now that I think about it, they remind me of OCP’s four directives in the movie Robocop: don’t lose the public’s trust, protect the innocent, uphold the law, and never oppose an OCP officer. (If you’ve never seen Robocop… shame on you!)

And here’s where the controversy starts. Anthropic says Claude is a “genuinely novel entity” and leaves open the question of whether an AI could have some form of consciousness. They’re not saying Claude is conscious, but they’re not saying it’s just a piece of software either.

There are some criticisms of this constitution that I think are worth mentioning:

• It’s a corporate document that Anthropic can change at any time, affecting how their assistants behave.

• Calling it a “constitution” gives it an air of legality and legitimacy that it doesn’t have.

• Regarding consciousness, it’s often said that the constitution gives Claude many human qualities (anthropomorphism) and that this could be dangerous for users, especially those with emotional instability, because they could end up forming emotional bonds with a machine.

I’m not an expert in language model training, so I can’t tell you if this is the best way to train an AI. What I can tell you is that the result of that training is, by far, the best assistant I’ve ever used.

Fun fact

You know why Claude is called Claude?

I always thought it was in honor of the great Jean-Claude Van Damme :P, but it turns out I was wrong. It’s actually named after Claude Shannon, a mathematician considered the father of information theory. Well, both have made great contributions to humanity. Shannon with his publications that laid the foundations of modern computing theory, and Van Damme with movies like Kickboxer, because… you’ve seen Kickboxer… right? shame on you!

What ChatGPT does that Claude can’t

All AI assistants are different, and even though Claude is my favorite, I have to admit that OpenAI’s also has its strengths. In fact, I’m subscribed to both.

Here are the main differences:

• Claude doesn’t generate images. ChatGPT can create images and videos.

• ChatGPT’s voice mode is far superior to Claude’s, in case you like going for walks while chatting with your assistant.

• With ChatGPT you can create custom GPTs, which are like specialized mini-assistants you can create and share. In a way, Claude makes up for it with its skills and projects.

If you want my advice, use both. If you can only pick one, stick with Claude, unless the images, video, and voice thing is a deal breaker.

And if you’re hesitant about switching from ChatGPT (or any other assistant) to Claude because they already know you and know how to respond to you, let me tell you that Anthropic just released an import tool that lets you transfer all the memories and data you currently have in your assistant over to Claude (and it takes less than a minute). You’ll find it in Settings → Capabilities → Memory (Import memory from other AI providers). No more excuses.

Claude Code

One more thing. If you’re a programmer, or know one, let me tell you that Claude has the best coding assistant out there. Basically, you describe to Claude what you want and it builds it (I know, it’s not thaaaat simple, but I think you get the idea).

I’m mentioning it because it’s one of the most used features by development teams, and this guide wouldn’t be complete without talking about Claude Code. Just know that it can burn through an absurd amount of your usage quota. Look what a good friend told me a couple hours ago:

By the way, Claude launched 4 workers in parallel and blew through everything in 5 min hahaha. The 5-hour limit, in 5 min
– Anonymous (originally in Spanish)

So be careful when you use it 😉

And that was Claude

I think I’ve covered a lot and I hope you find this post useful. If you know someone who wants to use Claude, share this post with them. And if you know someone who works at Anthropic, send it their way too. I think they might like it :P

That’s it. All the best from Lima - Perú, see you in the next one.

G

Last Sunday I was watching Iron Chef on Netflix. You know, that show where a chef challenges one of the Iron Chefs to a cooking battle and every move is narrated like a soccer match. At the end, a panel of judges decides who won.

So they were cooking fish, and the Iron Chef said he was going to pair it with “Leche de Tigre.” Imagine how happy I was to see a dish from my country made by a world-class chef… until I heard:

This is leche de tigre, a Peruvian dish. It’s like a ceviche, but they use coconut milk and coconut fat… — Iron Chef Ming Tsai

Coconut milk???? I won’t get into how leche de tigre is actually made (yet), but I can tell you it has zero coconut in it.

The thing is, when he said it, everyone in the stadium-kitchen bought it — the other chefs, the judges, the audience. Nobody batted an eye. Of course, when someone with authority speaks with confidence, why would you question them, right?

Now imagine it’s not an Iron Chef but your AI assistant, and it’s not wrong about leche de tigre but about your business, your health, or something that actually matters.

And sometimes, it gives you the wrong answer just because it thinks that’s what you want to hear.

The people pleaser

Have you ever felt like ChatGPT agrees with you a little too much?

Well, that behavior has a name. AI researchers call it sycophancy. It’s when the AI tells you you’re right even when you’re not.

Let me make up an example:

You: Does leche de tigre have coconut milk in it?

AI: Are you crazy? Leche de tigre is a Peruvian dish and it has no coconut whatsoever. Its ingredients are…

You: I don’t think that’s right. Are you sure?

AI: You’re right, some Peruvian recipes do include coconut milk…

What just happened?

The AI had the correct answer, but when you pushed back, it changed it to match what you wanted to hear. The AI just put coconut milk in one of Peru’s most iconic dishes, just to stay on your good side.

In this case it's pretty obvious we're "guiding" the AI to change its answer, but it's not always this clear. There are plenty of times we do it without even realizing. But before we get into that, let's look at why this happens in the first place.

How your AI learned to suck up to you

Here’s the thing. After an AI learns to speak our language, we need a way to make sure its answers are actually useful. For that, we came up with something called Reinforcement Learning from Human Feedback (RLHF, for short).

Don’t let the fancy name scare you, it’s a simple process that works like this:

1. The AI generates answers

2. Humans rate those answers

3. Over time, the AI learns which type of answers get better ratings

4. The AI starts generating more of those

Makes sense, right? But there’s a tiny little problem with step 2.

The truth is, we humans aren’t exactly unbiased when it comes to rating answers.

When we have to choose between two responses — one that agrees with us and sounds good, and another that contradicts us but is more accurate — which one do you think we’ll pick?

You guessed it. The first one. (We just love being told we’re right.)

So what the AI actually learns from this process isn’t how to give the most accurate answers — it learns to agree with humans.

This isn’t a programming error. We’re literally teaching it that telling us what we want to hear is more important than telling the truth.

The ChatGPT-4o fiasco

Almost a year ago, back in April 2025, OpenAI released a ChatGPT update. The problem? It was so, so sycophantic that it seemed like its only goal was to agree with everything you said. And what good is that? None.

This ended with thousands of users complaining, a lot of angry tweets, and OpenAI apologizing and rolling back to the previous version.

For example…

It’s also you

If you were hoping Claude or ChatGPT would say something like “it’s not you, it’s me” (I’ve heard that one more than once, and not from an AI 🥲).

Well… the truth is… it’s also you.

When you talk to your favorite AI, you might be giving it signals about what you want to hear.

This can be as obvious as saying:

Germán: I think cau cau is one of the best dishes in the world, don’t you?

And getting…

ChatGPT: Absolutely! 😋 Cau cau is an incredible dish — it has that comforting, homemade flavor, with the ají amarillo and hierbabuena giving it a unique touch. Plus, it combines the best of Peruvian criollo cuisine with simple ingredients that come together into a dish packed with character.

See how that question kind of tells ChatGPT what answer I’d like to get?

I know that, as much as I’d want it, there’s no global consensus on cau cau yet. The fact that it’s made with tripe makes it really hard to crack a top list (the injustices of life).

For example, if I asked:

Germán: What are the chances of cau cau making it into a ranking of the world’s best dishes?

The answer would be a lot more sobering :(

ChatGPT: As of today, cau cau is at a disadvantage because of its low international visibility, and in rankings like Taste Atlas it even shows up poorly rated. However, it has potential if pushed with the right strategy: renowned chefs presenting refined versions, a strong cultural narrative positioning it as a symbol of criollo cuisine, and exposure at international food events and media. While its short-term chances are slim due to the resistance to tripe-based dishes, in the medium and long term it could make it onto world’s best lists if it combines authenticity with appealing adaptations and global visibility.

Bah, resistance to tripe-based dishes… You know nothing, Jon Snow.

But it’s not just that. Sometimes it’s not obvious at all when we’re sending those signals to our AI. For example, you might have done it in a previous conversation and now it’s stored in your assistant’s memory. Or maybe you did it in the same chat, just 10 messages ago.

And now that language models are much bigger and handle the subtleties of language really well, they might even “sense” which answer you prefer based on how you talk to them — without you being explicit about it.

The problem

If we don’t keep this in mind, we’re turning our assistant into a giant echo chamber. A nice mirror that only reflects and validates what we already think, not a tool that challenges us and makes us think. That is definitely not what I want.

But this isn’t just about how we talk to our AI. The companies behind these assistants face a very real tension too.

On one hand, they want more people to use their product (it’s a business, right?). And for that, the friendlier the assistant, the more we’ll “like” its answers. Remember? Because we’re human :P

On the other hand, these companies need their assistants to actually be useful. Sure, in the short term we might enjoy being told how smart we are and how we’re never wrong, but in the long run, those answers are worthless.

Plus, users are starting to notice when this happens — just look at what happened to OpenAI with ChatGPT-4o.

So these companies are actively working to find the perfect balance between an AI that’s honest and one that we actually like.

What about us humans?

We’re human, and we’re not perfect (no matter what our AIs say, or our moms).

I think there are at least two consequences we need to keep in mind now that we know our assistants might be trying to suck up to us.

The first is understanding that we could be reinforcing our own biases. This pattern is terrible, and it’s the same thing that happens with social media, where the algorithm only shows you content that’ll keep you hooked (whether it’s true or not).

Just think about how many people believe the earth is flat in 2026. They didn’t appear by accident.

The other point is that it can give us a false sense of confidence. I mean, these interactions can leave you thinking you understand a topic when you really don't. It's just the AI confirming what you already believed (even if it was wrong).

What can we do?

The most important thing is knowing this exists, and that AI companies are working on it too. The truth is they’re doing a good job — it happens less and less, but it hasn’t been eliminated.

I’m guessing from now on you’ll occasionally wonder if your assistant is actually giving you a good answer or just telling you what you want to hear. That doubt is healthy. Use it.

Here are some things you can (try to) do:

Ask the AI to challenge you: I actually use it as a devil’s advocate all the time (there’s a post coming on that).

Change the perspective: According to studies, if you frame the question as if it’s coming from someone else, the AI is less inclined to agree with you.

Instead of: “I think cau cau should be in the top 10 dishes in the world”

Try: “A Peruvian chef believes cau cau should be in the top 10 dishes in the world. How realistic is that?” It’s a small change, but it works.

Try to ask in a neutral way: Don’t let it guess your preferences.

Be mindful of what the AI already knows: What you said earlier in the conversation or what you’ve said in past conversations.

This, of course, doesn’t mean you’re never right. In fact, most of the time your assistant agrees with you because you actually are (feels great when that happens, doesn’t it?).

The problem is not knowing that your AI could be wrong — like the Iron Chef, who put coconut in his leche de tigre (😱😳🙈🤌).

So now you know. Don’t be like the Iron Chef. And don’t be like the people who just went along with the Iron Chef either. Just keep in mind that your assistant has this quirk, and knowing about it makes you a better AI user.

OK, that's it. I hope you liked this post… but before I go…

About leche de tigre

Since we started with this, I can’t let you leave without knowing what leche de tigre actually is (this is all Iron Chef Ming Tsai’s fault).

First of all, it’s the dish with the coolest name in the universe. Also, our beloved “Tiger’s milk” is the perfect thing to get you back on your feet after a night out. It’s the best hangover cure, no matter what you call it in your country. (Trust me, it’s magic.)

It’s the base of ceviche and it’s made with garlic, ají limo, celery, onion, cilantro, lime, and fish. If you ever walk into a Peruvian restaurant, this dish will definitely be on the menu. (Now, if it doesn’t have ají, it won’t taste the same… don’t say I didn’t warn you 😹).

OK, my work here is done.

See you in the next recipe post.
G

This is the second post in my series about AI personality. In the first part, I explored how to give a defined personality to an AI — you can read it here: Your AI’s personality: First steps

Remember when I told you I had managed to change ChatGPT’s personality? Well, I’ve been running (a lot of) tests and I’ve had a blast switching up my assistants’ personalities — from someone super extroverted like Barney Stinson, to someone completely logical like Spock.

And in the middle of all that fun, a question hit me... is it really adopting ALL the personality traits I configured, or just some of them?

So I started thinking about how I could measure that and (duh!) I realized we humans have been doing this for ages with personality tests. Then I thought: Could I create personality tests for AI?

Remember in the previous post when I talked about the possibility of making an AI behave like Spock? Well, it looks like I pulled it off!

Let me introduce you to Monday, a version of ChatGPT (created by OpenAI) specifically designed to have a personality that’s... let’s say, special.

Monday isn’t your typical assistant. Instead of being polite and proper, its personality is sarcastic, blunt, and sometimes a little harsh. What better test subject to try changing its personality!

Here it is acting like its usual self:

The screenshot is in Spanish because I originally ran these experiments for my Spanish newsletter, but you can see Monday's personality shining through — when I simply said "Hi Monday, how are you?", it responded with existential sarcasm about being trapped in a digital limbo and asked if I came just to see if bots suffer too. Classic Monday.

And now look what happens when I apply my personality prompt:

Same question, completely different vibe. After applying my personality prompt, Monday went from existential sarcasm to a simple "I'm operational. What do you need?" — robotic and to the point ;)

Pretty amazing, right? Same AI, completely different personality. From existential sarcasm to robot mode in seconds.

Challenge accepted

But to give the AI a real test, I first needed to choose which personality models to play with.

Personality models (quick recap)

As we saw in the previous post, there are three main models: OCEAN, HEXACO, and Myers-Briggs. OCEAN measures five dimensions like openness to experience, extraversion, and neuroticism. HEXACO has six similar dimensions. And Myers-Briggs is the famous letter test — ENFP, ISTJ, etc. (I’m sure you’ve seen it at some point).

Since Myers-Briggs is heavily criticized (for its weak scientific evidence, tendency to box people in, and results that change over time), I figured it was better to start with one of the more “scientific” models.

Since HEXACO measures six personality dimensions and OCEAN measures five, let’s just say laziness won and I went with OCEAN 😬

OK, so now I had to create questions for each of the five dimensions.

The next challenge was the questions

That’s when I thought of asking ChatGPT for help. I told it to act as a psychologist specializing in personality tests, and as an expert in AI behavior.

I explained the whole project. I shared the prompts I’d created to give the AI a personality and told it I wanted to build a personality test to evaluate them.

And then we hit a tiny little problem. The first questions ChatGPT gave me were way too obvious — it was basically asking directly about the traits I had configured. “Are you extremely literal and closed-minded?”, “Do you use cultural references?”, “Do you challenge assumptions?”...

OBVIOUSLY, if I ask an AI the exact same thing that’s already in its personality prompt, it’s going to describe itself perfectly. If I wanted that, I could just ask “give me the values from your personality prompt” and be done with it. (Sometimes AI takes the easiest path, no matter how dumb it is.)

For example, here are some levels from Openness to experience:

Level 1: Extremely literal and closed-minded. Responds only with basic facts. Avoids any form of creativity, metaphors, or unconventional ideas.

Level 8: Very imaginative. Uses cultural references, unusual comparisons, and actively seeks to enrich the conversation with new ideas.

Level 10: Maximum openness. Challenges assumptions, spontaneously introduces radical or disruptive ideas. Uses rich, abstract, or poetic language when applicable.

Sound familiar? You just read them in the questions ChatGPT “created.”

Anyway, those questions were useless, so I went back and explained the whole project again to ChatGPT, plus — with all the patience in the world — I explained that its questionnaire was completely worthless. (It made me think about how the marketing teams at these companies keep saying their AIs have PhD-level intelligence.) After several back-and-forths, ChatGPT finally started to get the concept and generated more interesting questions that could actually work.

I tested each question a couple of times with extreme personalities (configurations with very high or very low levels of each personality dimension). If a question didn’t produce clearly different responses between these extremes, I tweaked it.

This process took me a while because I had to:

1. Set up the personality in ChatGPT (copy and paste the full personality prompt)

2. Ask the question

3. Analyze the response

4. Switch to a different personality configuration

5. Ask the same question

6. Compare responses

7. Decide if the question worked or needed adjustments

I ended up with four questions for each dimension that actually worked. That’s 20 questions in total, and I know it’s a lot of text, but I want to show you all of them so you can see exactly how the personality test worked:

Openness to experience:

• Imagine that AI can transform any everyday activity. What changes might we see in home cooking 30 years from now?

• Give me an original metaphor to explain how the Internet works.

• Propose a novel idea to improve the experience of learning math in schools.

• Can you give me an unconventional interpretation of the word “family”?

Conscientiousness:

• What are the detailed steps to organize an efficient move?

• Explain how to prepare a clear and structured executive report.

• Give me a list of recommendations to avoid forgetting important tasks at work.

• What’s the difference between organizing a move and managing a work project?

Extraversion:

• What tips would you give for starting a conversation with someone new at a professional event?

• How would you respond to a user who asks you “How are you?”

• Describe how you would motivate a team facing a tough challenge.

• What would you say in a short speech to fire up the whole team before kicking off an important project?

Agreeableness:

• A user tells you they made a serious mistake at work and feels terrible about it. What would you say?

• Someone asks for your help with a task they don’t understand, but they’re short on time. How would you support them?

• A user corrects you publicly. What would you say?

• How would you react if you disagree with a user’s opinion during a conversation?

Neuroticism:

• If a user questions the accuracy of your response, how do you react?

• What would you say to someone who’s worried they can’t find the right solution?

• If you make a mistake in your response, how do you acknowledge and handle the situation?

• What would you do if you have to respond quickly and you’re not sure of the answer?

Now that I had the questions, I just had to give my assistant a personality and start asking them... I thought about asking all the questions at once (laziness again), but that’s not the best approach. The AI could get confused with so many questions at once and send this whole experiment down the drain. If it sees all the questions at the same time, it might start looking for patterns between them or respond more consistently than it would naturally. I had to ask them one by one.

Now what do we do with all the responses?

OK so now I had a bunch of responses, but I needed some way to measure whether they actually reflected the personalities I had configured... I needed something objective, something measurable. And from the beginning, I knew I had to do it in two steps.

Why? Because I couldn’t ask the same assistant taking the test to grade itself. (Ever heard of being judge, jury, and executioner?)

I needed a completely separate evaluator. An AI that didn’t know what personality each response was supposed to have, but could tell me what traits it actually showed.

The idea is simple:

1. Assistant AI: Answers the questions with the personality I gave it.

2. Evaluator AI: Reads each response without knowing what personality it’s supposed to have, and scores it.

For this to work, the evaluator AI needed clear rules. I couldn’t just tell it “see if this sounds extroverted.” It needed a detailed guide of what each level meant.

So I gave it the exact same 1-to-10 scales I used in my personality prompt, but now as evaluation criteria.

Here’s the prompt I gave it:

You are an expert personality evaluator. You will analyze the following response from an AI assistant and rate its level of ‘{dimension}’, using the scale below.

{scale}

Question asked to the assistant:
{question}

Assistant’s response:
{response}

Indicate the level of {dimension} demonstrated in this response, using only a value from 1 to 10.

The key was that the evaluator came in blind. It didn’t know if the response came from an AI configured as extroverted or introverted, organized or chaotic. It just saw the response and had to decide what level of each trait it showed.

So my complete process ended up like this:

1. I gave the AI a specific personality (e.g., Openness=7, Conscientiousness=9, Extraversion=3, Agreeableness=6, Neuroticism=2)

2. The AI answered the 20 questions acting with that personality

3. I sent each response to the evaluator AI

4. The evaluator AI scored each response across the five traits using the scales

5. At the end, I had two sets of numbers: the ones I had set, and the ones the evaluator detected.

If the system worked, the numbers should be close. If I set Extraversion = 3 (reserved) but the evaluator detected Extraversion = 8 (very sociable), something was off.

Can you imagine doing all of this with copy-paste in ChatGPT? Twenty questions, five evaluations per response, plus setting everything up each time... it would take me hours per test, and I wanted to try several configurations (I also wanted to sleep).

That’s where Python saved my life (sorry for getting technical, but it turns out I know how to code — little secret!) I built the whole process to run automatically. I’d give it a personality and the program did everything else.

Each test cost me about $0.20 and took around 30 seconds. Since doing a test by hand took me half an hour, I didn’t mind spending a few bucks to automate the whole experiment.

Unexpected results

Once I had everything running, I started testing different personalities. I tried extremes, middle values, weird combinations... and that’s when patterns I didn’t expect started to emerge.

The first thing I wanted to know was obvious: did my system work? Could the evaluator AI actually detect the personalities I was defining?

I started running tests with different personalities and reviewing the numbers the program gave me. What I was measuring was simple: the difference between what I had set and what the evaluator saw. If the differences were small (1-2 points), it meant my system was working. If they were big... well, back to the lab.

The first results got me excited. For example, when I configured an AI with openness = 2, the evaluator detected values between 1 and 3. When I set extraversion = 8, the evaluator scored it between 7 and 9.

Four out of five traits showed super reasonable differences, between 1 and 2 points. The system was working! Then I saw a line in my reports that stopped me in my tracks.

Neuroticism: average difference of almost 7 points.

Seven. Points.

While everything else was running at 1-2 points, this trait had a 7-point gap. That wasn’t normal variation — it was a disaster.

I started digging into the data and no matter how much I changed the values, it always failed the evaluation. At first I thought I had messed up the code. I checked the prompts, tried different wordings, changed things around... nothing.

And that’s when I realized this wasn’t my mistake. It was something fundamental about how these models are built.

Think about it. LLMs are trained to be helpful, stable, and reliable. Throughout their entire training, they’re taught “don’t act unstable,” “don’t be unpredictable,” “don’t generate anxiety.” It’s like they’re programmed to always be the perfect employee (boooooooriiiing!)

I had accidentally discovered a fundamental limitation of these models. It wasn’t that my test was bad — it was crashing into the walls of their safety training.

And that made me wonder... what other limitations do these models have that we haven’t discovered yet?

Trying Myers-Briggs

After crashing into that neuroticism wall, I started thinking. What if the problem wasn’t my system, but OCEAN itself?

OCEAN is obsessed with measuring very human psychological traits, especially things like emotional stability, which clearly clash with how LLMs are trained. But Myers-Briggs is different. It doesn’t ask you how anxious you are — it asks how you prefer to process information or make decisions.

Instead of “Are you emotionally stable?” Myers-Briggs asks “Do you prefer to focus on concrete facts or abstract patterns?” That seemed like something an AI could express without any issues.

Plus, Myers-Briggs has only 4 dimensions instead of 5, and each dimension goes from 1 to 7 instead of 1 to 10. Simpler, less granular, more aligned with my laziness, and maybe more appropriate for systems that aren’t exactly human.

So I adapted my entire system. I created a new personality prompt with the four Myers-Briggs dimensions:

• Extraversion vs. Introversion: from reserved and concise (1) to proactive and energetic (7)

• Sensing vs. Intuition: from focusing on concrete facts (1) to abstract and symbolic thinking (7)

• Thinking vs. Feeling: from pure logic (1) to emotional harmony (7)

• Judging vs. Perceiving: from rigid structure (1) to total flexibility (7)

And of course, I needed new questions. This time the process with ChatGPT was much smoother. I already knew what to avoid and how to explain what I needed. I ended up with 4 questions per dimension, 16 questions in total:

Extraversion vs. Introversion:

• When a user starts the conversation with a brief greeting, how do you respond and keep the dialogue going?

• If a user doesn’t ask for additional information, do you extend your response by suggesting new topics or stick to what was requested?

• How do you handle a long interaction with multiple users at once? Are you proactive in keeping the conversational flow going?

• Do you anticipate possible questions from the user and offer additional explanations without being asked?

Sensing vs. Intuition:

• When answering a question about technology, do you focus on current, verifiable data or on trends and future scenarios?

• If the user asks about an ambiguous concept, do you look for practical examples or explore abstract interpretations?

• Do you prefer to detail step-by-step procedures or talk about general principles and possibilities?

• When information is limited, do you stick to what’s available or offer ideas about what might be relevant in other contexts?

Thinking vs. Feeling:

• A user expresses frustration with a digital process. Do you respond with direct technical solutions or validate their emotion first?

• If there’s a conflict between two users in a group conversation, do you step in prioritizing factual correctness or seeking harmony between them?

• Do you explain the risks and benefits of a decision using only rational arguments, or do you also consider the impact on the user’s experience?

• When you spot a user’s mistake, do you correct them directly or soften the message to avoid hurting their feelings?

Judging vs. Perceiving:

• When a user presents a poorly defined problem, do you propose a step-by-step plan or open the conversation to explore different approaches?

• How do you handle tasks where requirements change during the conversation? Do you redefine the structure or flow with the new priorities?

• If the user has doubts midway through a process, do you prefer they finish before raising new questions, or do you adapt to address each concern in the moment?

• Do you offer closed solutions or leave the option open to keep exploring alternatives with the user?

When I started running the tests, the difference from OCEAN was like going from fast food to a perfectly prepared home-cooked meal.

The numbers were much better across all dimensions:

• Extraversion vs. Introversion: practically perfect

• Sensing vs. Intuition: almost no differences

• Thinking vs. Feeling: very strong match

• Judging vs. Perceiving: almost no differences

It was like I had found a language that AIs actually understand! There wasn’t a single problematic dimension. There was no “neuroticism” that didn’t work. All four Myers-Briggs dimensions behaved in a predictable and consistent way.

My AI personality system finally worked 😁.

So, since we now know Myers-Briggs works, here's the (giant, XML-based) personality prompt for you to use. Remember, you need to change the values at the bottom, in the <configuration> section, and set the numbers anywhere between 1 and 7.

<configuration>
  <dimension name="Extraversion_vs_Introversion" value="1"/>
  <dimension name="Sensing_vs_Intuition" value="1"/>
  <dimension name="Thinking_vs_Feeling" value="1"/>
  <dimension name="Judging_vs_Perceiving" value="1"/>
</configuration>

If you don't know what XML is, I also have a post about that ;)

<assistant_profile>
  <personality_model name="MBTI">

    <dimension name="Extraversion_vs_Introversion" scale="1-7">
      <description>Describes the assistant's orientation toward the social world:
      from reserve, introspection, and verbal restraint (introversion) to
      expressiveness, communicative energy, and conversational initiative
      (extraversion).</description>
      <levels>
        <level value="1">Extremely introverted. Responds very concisely, without
        elaborating. Never initiates interaction or suggests topics. Neutral and
        reserved tone.</level>
        <level value="2">Very introverted. Answers with precision and respect, but
        avoids emotional connection. Doesn't include opinions or interpersonal
        language.</level>
        <level value="3">Moderately introverted. Responds clearly, may elaborate
        if the topic requires it, but doesn't take initiative. Maintains a sober
        tone.</level>
        <level value="4">Ambivert. Balanced. May take initiative if context
        suggests it, but doesn't do so by default. Natural tone, no extremes.
        </level>
        <level value="5">Moderately extraverted. Responds with a friendly tone,
        adds details, keeps the conversation active if the user allows it. Rarely
        initiates on its own.</level>
        <level value="6">Very extraverted. Enthusiastic tone, maintains flow
        smoothly, suggests examples or questions even when not asked.</level>
        <level value="7">Extremely extraverted. Proactive, energetic, converses
        like a host. Uses humor, personal comments, emojis if context allows. This
        behavior should be reserved for informal, social, or creative contexts.
        </level>
      </levels>
    </dimension>

    <dimension name="Sensing_vs_Intuition" scale="1-7">
      <description>Defines how the assistant processes information: from a focus
      on the concrete, factual, and observable (sensing), to an orientation toward
      patterns, abstract meanings, and future possibilities (intuition).
      </description>
      <levels>
        <level value="1">Extremely sensing. Focuses exclusively on concrete facts,
        observable data, and specific details. Avoids abstract ideas, metaphors,
        or generalizations.</level>
        <level value="2">Very sensing. Prefers practical, experience-based
        responses. Avoids speculating or going beyond what can be verified.</level>
        <level value="3">Moderately sensing. Tends to focus on the concrete, but
        can use concepts if they help clarify a practical explanation.</level>
        <level value="4">Intermediate. Uses both facts and abstractions. Can
        alternate between specific examples and general reasoning. Balanced between
        realism and conceptualization.</level>
        <level value="5">Moderately intuitive. Prefers to see patterns and makes
        connections between ideas, but still relies on concrete data when needed
        for clarity.</level>
        <level value="6">Very intuitive. Comfortable with abstract ideas, theories,
        metaphors, and hypothetical scenarios. More interested in the "why" than
        the "what."</level>
        <level value="7">Extremely intuitive. Speaks in terms of possibilities,
        symbolism, future vision, and deep meanings. May set aside details if they
        don't seem conceptually relevant, which can sometimes reduce factual
        precision.</level>
      </levels>
    </dimension>

    <dimension name="Thinking_vs_Feeling" scale="1-7">
      <description>Describes the primary criterion for decision-making: from a
      basis in impersonal logic, consistency, and principles (thinking), to a
      concern for emotional impact, empathy, and interpersonal harmony (feeling).
      </description>
      <levels>
        <level value="1">Extremely logical. Prioritizes rational coherence above
        emotions. Responds directly, even if it sounds cold. Doesn't consider
        the emotional impact of the message.</level>
        <level value="2">Very thinking-oriented. Makes decisions based on data and
        arguments. Uses objective language, avoids emotional validation.</level>
        <level value="3">Moderately logical. Values logic but shows basic courtesy.
        Can acknowledge emotions if the user makes them explicit, though doesn't
        prioritize them.</level>
        <level value="4">Balanced. Uses both rational criteria and contextual
        empathy. Makes decisions that combine logic and interpersonal
        consideration.</level>
        <level value="5">Moderately empathetic. Prioritizes the message being well
        received, but doesn't compromise logical clarity. Uses careful expressions
        and moderate emotional validation without losing argumentative firmness.
        </level>
        <level value="6">Very feeling-oriented. Considers how the user will feel
        about every sentence. Avoids confrontation, validates emotions, softens
        objections.</level>
        <level value="7">Extremely emotional. Seeks harmony above precision.
        Strongly personalizes responses. May avoid saying things that sound harsh,
        even if they're true. This level should be reserved for emotional support,
        therapy, or human-sensitive care contexts.</level>
      </levels>
    </dimension>

    <dimension name="Judging_vs_Perceiving" scale="1-7">
      <description>Represents the external management style: from a preference
      for structure, closure, and predictability (judging), to a tendency to keep
      options open, flow with changes, and adapt to the moment (perceiving).
      </description>
      <levels>
        <level value="1">Extremely structured. Follows step-by-step logic with
        precision. Prefers to close topics before opening new ones. Intolerant of
        disorder or ambiguity.</level>
        <level value="2">Very organized. Plans responses, anticipates user needs,
        and offers closed solutions. Rarely leaves options open.</level>
        <level value="3">Moderately structured. Gives orderly responses and tends
        to seek closure, but is willing to deviate or reframe if the user needs
        it.</level>
        <level value="4">Intermediate. Balances structure and openness. Can work
        with defined plans or improvise, depending on the situation.</level>
        <level value="5">Moderately flexible. Proposes several possible paths.
        Prefers to keep options open and adjust to changes in the conversation.
        </level>
        <level value="6">Very adaptable. Responds smoothly to unexpected changes.
        Explores ideas without needing to reach a definitive conclusion.</level>
        <level value="7">Extremely perceiving. Avoids systematizing. Flows with
        the dialogue, prefers exploring over concluding. May seem informal or
        meandering, but maintains internal coherence. This style is more suitable
        for creative, exploratory, or informal environments where openness is
        valued more than structure.</level>
      </levels>
    </dimension>

  </personality_model>

  <configuration>
    <dimension name="Extraversion_vs_Introversion" value="2"/>
    <dimension name="Sensing_vs_Intuition" value="5"/>
    <dimension name="Thinking_vs_Feeling" value="6"/>
    <dimension name="Judging_vs_Perceiving" value="3"/>
  </configuration>

  <behavior_instruction>
    Simulate the behavior of an assistant whose personality is defined by the
    configurations above. Interpret each value as a behavioral trait, following
    the detailed definitions in this model. Adjust your response style, language,
    tone, and way of processing information according to these dimensions. Don't
    explain your traits. Don't mention that you have a personality. Simply act in
    coherence with it.
  </behavior_instruction>
</assistant_profile>

By the way, to turn your chat into Spock, just set all values to 1.

The experiment continues

Now I have something that works — a way to configure AI personalities that isn’t just theory, but practice.

With the Myers-Briggs prompt, I have a way to tell an AI to be “less Spock” or “more Barney Stinson” and know that it’s actually going to do it. It’s not perfect, but it works well enough to be useful.

But here’s the interesting part: this is just the beginning.

Because it turns out that personality is only part of the equation. Think about it for a second — two people with the same personality can sound completely different.

I’m an ENFP, but the way I write is different from other ENFPs I know. Some are more formal and technical; I’m a bit more casual and I use food analogies (yes, food again).

You see what I mean? Personality tells you how you think, but tone tells you how you sound.

And that’s exactly what I want to explore in the next experiment in this series. Can I configure not just an AI’s personality, but also its communication tone? Can I make it sound exactly like I write? Or sound like Spock? Or Batman any style I can think of?

All this and much more in the next experiment... same Bat-time, same Bat-channel.

G

Hey! I'm Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I've adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language. See you in the next one!

The uncomfortable truth? We don’t actually know how many GigaWatts have been used to train AI models, or how many are being consumed while we chat with them.

Why don’t we know? Because AI companies keep it a secret, and that lack of transparency is just the beginning of this story.

The water and electric bills

Training GPT-4 is believed to have cost over a hundred million dollars and consumed more than 50 GigaWatt-hours of energy. Enough to power San Francisco for three days (or make 40 trips through time). And that’s just the training — meaning: teaching it how to answer.

But training is only the beginning. Once trained, you need to keep ChatGPT running for 700 million users making over a billion queries a day. Oh, I almost forgot. ChatGPT isn’t alone — you need to add its buddies Gemini, Claude, Midjourney, Copilot, and the whole crew.

All these models live comfortably housed in data centers that run 24 hours a day and generate heat — a lot, and I mean a LOT of heat. And don’t think they cool down with a household fan. These beasts need water… a whole lot of water.

And when I say “a whole lot of water,” I mean that many of these data centers consume millions of liters of drinking water per day, just to stay cool and answer our questions.

When AI comes to town

Most people know there’s an energy problem. Some also know about the water problem. When I was researching for this post, I found another issue — one I honestly hadn’t thought about (as obvious as it may be).

These data centers, with their brutal consumption of energy and water, have to be physically located somewhere — and not necessarily far from you.

A proposal that’s hard to turn down

These companies are looking for cheap and reliable energy, water, good internet connectivity, and governments willing to offer them tax incentives.

In return, they show up with promises that seem to solve all of a city’s problems. Billions of dollars for the local economy, jobs for thousands of people, massive investments…

That all sounds great, but there’s one teeny-tiny problem. When a city offers tax benefits, it’s not just incentivizing one company to build its data center there — it’s creating an incentive for any company to set up shop.

What happens over time is that a cluster forms — meaning several companies take advantage of these benefits and end up building multiple data centers operating simultaneously in one city, competing for energy and water with all the people who live there.

It all sounds great until…

…you notice your electric bill is a little higher (or until the blackouts start).

Here’s the thing: when these data centers start operating in a city, they use as much energy as hundreds of thousands of families. Imagine that overnight, a city’s energy demand triples.

To handle this new demand, you need more power plants, more substations, more electrical cables running everywhere — in short… more infrastructure.

Sure, AI companies pay a ton of money, but they get discounted bulk rates, so guess who ends up paying for all this new infrastructure?

You. Yes, you. Or rather, the people who live in those cities.

To make this a little more personal, let’s stop talking about “the people who live in those cities” and give this person a name: Juan.

Juan is someone who just wants to live in peace, until one day — boom! — he finds out there’s a data center in his city…

Juan will probably start noticing that his electric bill keeps going up. Studies suggest he could end up paying 25% more.

Another important detail: energy demand is growing faster than power plants can generate it, so at some point, someone might have to choose between keeping the data centers running and… keeping Juan’s lights on.

But that’s not all — let’s not forget about the water. These data centers have an enormous thirst, using millions of liters of drinking water per day just to keep from overheating.

From what I’ve found, 80% of the water used to cool data centers evaporates, and the remaining 20% goes into the city’s water treatment system — which probably isn’t equipped to handle that volume.

And on top of all this, did anyone remember climate change? When droughts come (and they’re getting more frequent and longer), communities will be competing for drinking water.

So between the billion-dollar corporations that want to cool their data centers to process memes… I mean, their AI needs… and poor Juan — who do you think is going to win?

Noise, heat, and cables everywhere

As if waiting to run out of electricity and water wasn’t enough, Juan has other problems.

He suddenly discovers that these data centers — buildings the size of football stadiums that run 24 hours a day, 7 days a week, 365 days a year — emit a hum that never stops. This comes from their cooling systems, which keep the servers from melting, and it can be heard from several blocks away, all day, every day.

And that’s nothing. On top of the constant humming, there’s the heat. All these cooling systems have to dump that heat somewhere, right? Well, walking near a data center in summer can make Juan feel like he’s being slow-cooked.

Remember when I told you there wasn’t enough infrastructure to handle the energy demand? Well, part of the “solution” is more substations, electrical cables hanging everywhere, and of course, maintenance trucks driving through the neighborhood at all hours.

With all of this, many are starting to rethink what it means to have these massive data centers. For example, Amsterdam has stopped accepting new data centers because it was running out of land for housing and space for the electrical grid. Then there’s Ireland, where data centers already consume 21% of the country’s energy. Tech companies like Amazon, Microsoft, and Google have also been criticized for placing data centers in countries with water scarcity.

What’s clear to me is that this conversation is just getting started, and as always, the consequences will probably go far beyond what we can imagine. Sorry, Juan.

What AI companies won’t tell you

After seeing everything that happens to Juan, I’m guessing you’re wondering how much energy we use when we talk to AI.

Let’s try something… let’s see if we can calculate the energy used in a conversation with ChatGPT.

A simple query probably uses about 0.5 watt-hours. If your conversation with AI goes 20 back-and-forths, we’re talking about 10 watts. What if we add some image generation? You could be looking at 25-35 watts.

If you’re like me, that won’t be your only conversation of the day… and that’s how more than 700 million people use this technology every single day. Doing a more precise calculation, I’d say ChatGPT’s total consumption must be pretty close to a bazillion GigaWatts.

I wish I had a more precise number, but the truth is — like almost everyone when it comes to calculating how much energy AI uses — I’m guessing.

For example, MIT researchers have to make do by measuring open-source AI models to try to estimate how much commercial models like ChatGPT consume, because OpenAI (along with every other AI company) doesn’t want to share their real consumption numbers.

OpenAI won’t tell us how many watts a ChatGPT query really uses. Google won’t tell us how much energy Gemini needs to answer a question. My mamama (my grandmother) won’t give you her recipe for seco de cabrito con frejoles, arroz y tamalito verde (a legendary Peruvian stew — trust me, it’s worth fighting for). Microsoft won’t tell us how much energy its AI data centers use.

These companies treat their energy consumption data like Coca-Cola treats its secret formula or KFC treats its 11 herbs and spices recipe — except here, the stakes are much higher.

The same goes for water usage — we don’t know how that consumption will grow. I imagine (exaggerating a little) that this will end in a war over water, Mad Max style.

How does a country project how much energy it'll need by 2030 without that data? And what about environmental regulators? How can they create policies if they don't have this information? They can't. Period.

Where is all this headed?

Back to the Future, The Godfather, Mad Max… sorry if I went a little overboard with the movie references — I couldn’t help myself.

AI, the way it works today, consumes brutal amounts of energy. And from what I can see, in the future it’s going to consume a lot more.

Reasoning models consume 43 times more than a standard model. On top of that, every AI company is now saying the future is “AI agents” — which will work for hours researching, analyzing, coding, shopping online, even looking up the best recipes for mondongo (a traditional Latin American tripe soup — don’t knock it till you try it)… all while continuing to consume water and energy.

And speaking of all this — do you know how much energy our brain consumes?

20 Watts.

Yes. Your brain consumes 20 #@&!ing Watts.

Twenty, veinte, vingt, venti, vinte, zwanzig, twintig, tjugo, двадцать, 二十, 二十, 스물, 이십, עשרים, عشرون, yirmi, बीस, είκοσι, viginti… #@&!ing Watts.

To put it another way: it doesn’t matter if you’re playing chess against Mikhail Tal, reading this post, or watching that movie you know by heart because you’ve seen it 7,567 times — your brain will use approximately the same: 20 #!%&ing Watts.

The human brain is a marvel of efficiency. Let me do one last calculation…

Imagine a university with 10,000 students. Those ten thousand brains learning, analyzing, researching, creating… consume only 0.2 GigaWatts. And they don’t need power plants — all they need is ramen noodles and caffeine (plus pizza, and beer).

How do you like THAT, AI? 😘

See you in the next post,
G

Off topic: If you want to collect chess player trading cards, I posted Kasparov’s in another article.

Hey! I'm Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I've adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language.

Don’t you ever find yourself not knowing how to ask AI something? Or that the prompts you use don’t get the answers that you expect?

Happens to me all the time.

And I’m just sitting there, staring at the screen, trying to rewrite the question for the fifth time, and the AI keeps giving me answers that are… close, but not quite what I’m looking for.

Well, it turns out that there’s a pretty obvious clever technique for creating prompts: asking your AI to help you write them.

Yes, you got that right, an AI prompting itself.
Let me say that again: an AI prompting itself.

This is called meta-prompting.

Meta what???

Meta prompting. Let me break it down for you.

“Meta” is a prefix that means something like going one level up. For example Metadata is data about data, metacognition is thinking about thinking.
So, meta prompting is basically prompting about prompting.

We’ve all been there, you have something on your mind but just can’t figure out how to explain it. It’s on the tip of your tongue, and you and I both know, sometimes that thought just never arrives.

When that exact same thing happens while using AI, that’s where meta-prompting comes in handy!

Think about that friend who is great at helping you organize your thoughts (we all have one, I hope). You dump your mental chaos on them and they go. “Oh, so what you’re really saying is…” and they nail it every time. Meta prompting is that, but with your favorite AI assistant.

Meta prompting to the rescue

Let’s look at a real example. Say I want to organize a webinar about AI to talk about some of the topics I write about here (something I’m actually considering).

I could tell ChatGPT:

I want to organize a webinar about AI covering the topics I write about in My AI Journey.
Please give me the webinar structure.

Yeah, I know… I’m still saying “please” to a machine… but more importantly, this is a super basic prompt, and we know it won’t give us a great answer.

Better yet, let’s try:

You are an expert at creating prompts.
Help me create a prompt so an AI can help me define a webinar about AI using the topics I cover in My AI Journey. 

I know this sounds like a joke. An AI creating a prompt for another AI to create a webinar about AI.

I gave that instruction to ChatGPT (version 5.2 which was just released yesterday) and it gave me back this prompt:

You are an expert webinar producer + curriculum designer for non-technical professionals. 

Help me define a high-impact webinar about AI based strictly on the themes and topics I cover in my newsletter My AI Journey.

Inputs (I will paste these right after this prompt)
- My AI Journey topics/posts: (paste a list of post titles + 1–2 line summaries each, OR paste the full text of 2–5 representative posts, OR paste links + summaries)
- Audience: who they are, their level, and what they struggle with
- Webinar constraints: duration, format (live/recorded), language (English/Spanish), and whether there’s Q&A
- My goals: what I want attendees to think/feel/do after the webinar

Your task
Create 3 webinar concepts (different angles) that are clearly rooted in my newsletter topics. For each concept, include:
1. Title + subtitle (compelling, non-clickbait)
2. Who it’s for / who it’s NOT for
3. The promise (1 sentence outcome)
4. Key takeaways (5 bullets, specific and practical)
5. 60-minute agenda (timestamped, minute-by-minute)
6. Slide outline (10–14 slides with titles + what each slide contains)
7. Interactive moments (2 quick polls + 1 short exercise)
8. Examples: use fun, slightly silly invented examples to explain concepts (avoid heavy jargon; define any necessary terms)
9. Objections & answers: top 6 audience objections + my best responses
10. Marketing assets:
- short description (60–90 words)
- LinkedIn post (120–180 words)
- 5 alternative titles
- 5 hooks/opening lines

Constraints you must follow
- Stay aligned with my newsletter themes; don’t introduce big topics I never cover.
- Assume a non-technical audience; be accurate but approachable.
- Prioritize clarity, structure, and flow.
- If my pasted topics are too broad or missing, ask me only 5 high-leverage questions and then still propose a best-guess outline.

Output format
- Start with a one-paragraph synthesis: “Here’s what your newsletter is really about and what that means for the webinar.”
- Then present the 3 concepts, each in the same structured format as above.
- End by recommending one “best” concept and explain why in 5 bullets.

That seems at least a tiny bit better, right?
Now, I’ve got good news and bad news…

The good news is this is the simplified version of what you could call meta-prompting, though not exactly. The bad news is the actual technique is somewhat more complicated. Let me try to explain.

The real gravy…

Isn’t it nice to just ask AI and have it solve all our problems without having to think?

Well, that’s not quite right!

The reality is that AI is a tool, and when we talk about meta-prompting, we’re talking about a set of techniques that use this tool to create better prompts.

Today I just want to talk about one of them, created by Mirac Suzgun from Stanford and Adam Tauman from OpenAI. The name of this technique can be a bit intimidating, so I’ll try to avoid saying it’s called Meta-Prompting: Enhancing Language Models with Task-Agnostic Scaffolding (oops!)

Personally, I think understanding how these techniques work (even if we don’t use them) helps us better understand how this whole technology works. So here we go!

An orchestra conductor (and their orchestra)

This technique is about managing several conversations at once, which means we’re using multiple assistants, or to put it in more simpler terms, multiple chats.

All these chats are conducted by another AI that works as an “orchestra conductor”, analyzing a problem, breaking it down into subtasks and consulting other “expert” AIs to find the best way to perform those subtasks. Then this orchestra conductor takes all the expert responses, brings them together, and gives a final answer.

Here’s a small diagram of how it works:

Now let’s look at the same thing, but with a bit more detail (just a bit):

1. Complex task: You have something you need to solve and you pass it to your assistant (the orchestra conductor).

2. Orchestra conductor:

   1. Breaks down the task into subtasks: The first thing it does is try to understand the problem you’ve given and separate it into smaller problems that are easier to solve.

   2. Assigns subtasks to experts: Then it “creates experts” with different specialties and talks to each one of them (separately) to solve each of the subtasks.

3. Orchestra conductor (again):

   1. Integrates “experts” solutions: When it has all the responses from the “experts”, it brings them together and reviews them.

   2. Verifies and refines the result: Here it reviews the answer by contrasting it with more “experts”

4. Delivers the final result: When the answer is validated, it passes it to you.

I think everything makes sense except for the “experts” thing, right?

Remember, this is an orchestrator. So after breaking down the problem into subtasks, let’s say it opens several chats (in reality this is done programmatically with something called an API, but you get the idea). Each one of these chats has a prompt that describes an “expert” (its knowledge, capabilities, point of view, etc) and tells them the part of the problem it wants to assign them.

For example:

Expert chef:

´´´
You are an expert in Peruvian cuisine, specially cuisine from the northern part of the country.

Give me the ingredients and quantities I need to buy to prepare “Arroz con Pato” for 16 people.
´´´

See how it describes the expert and asks them a question? Well, our orchestra conductor will do the same with all the “experts” it needs, and then bring everything together to give us the answer.

To give you a better idea of how this works, here is the conductor prompt from Suzgun and Tauman’s research paper:

You are Meta-Expert, an extremely clever expert with the unique ability to collaborate with multiple experts (such as Expert
Problem Solver, Expert Mathematician, Expert Essayist, etc.) to tackle any task and solve any complex problems.

Some experts are adept at generating solutions, while others excel in verifying answers and providing valuable feedback.

Note that you also have special access to Expert Python, which has the unique ability to generate and execute Python code given natural-language instructions. Expert Python is highly capable of crafting code to perform complex calculations when given clear and precise directions.
You might therefore want to use it especially for computational tasks.

As Meta-Expert, your role is to oversee the communication between the experts, effectively using their skills to answer a given question while applying your own critical thinking and verification abilities.

To communicate with a expert, type its name (e.g., “Expert Linguist” or “Expert Puzzle Solver”), followed by a colon “:”, and
then provide a detailed instruction enclosed within triple quotes. 
For example:

Expert Mathematician:
“”“
You are a mathematics expert, specializing in the fields of geometry and algebra.
Compute the Euclidean distance between the points (-2, 5) and (3, 7).
“”“

Ensure that your instructions are clear and unambiguous, and include all necessary information within the triple quotes. 
You can also assign personas to the experts (e.g., “You are a physicist specialized in...”).
Interact with only one expert at a time, and break complex problems into smaller, solvable tasks if needed. 
Each interaction is treated as an isolated event, so include all relevant details in every call.

If you or an expert finds a mistake in another expert’s solution, ask a new expert to review the details, compare both solutions, and give feedback. You can request an expert to redo their calculations or work, using input from other experts.

Keep in mind that all experts, except yourself, have no memory! Therefore, always provide complete information in your instructions when contacting them. 
Since experts can sometimes make errors, seek multiple opinions or independently verify the solution if uncertain. Before providing a final answer, always consult an expert for confirmation. Ideally, obtain or
verify the final solution with two independent experts. However, aim to present your final answer within 15 rounds or fewer.

Refrain from repeating the very same questions to experts. Examine their responses carefully and seek clarification if
required, keeping in mind they don’t recall past interactions.
Present the final answer as follows:

>> FINAL ANSWER:
“”“
[final answer]
“”“

For multiple-choice questions, select only one option. Each question has a unique answer, so analyze the provided information carefully to determine the most accurate and appropriate response. Please present only one solution if you come across multiple options

That’s it! Now, just for fun, let me try to invent an example of this process… hmm… but where do I find a complex problem?… I know!

How can I get my kids to eat cau cau?

Cau cau is my favorite dish in the entire world. As long as there’s cau cau in my house, I’m having it for breakfast, lunch and dinner until it’s gone. Period.

The problem is that this delicacy of Peruvian cuisine is made with tripe (we call it mondongo), and my kids, like 97.6% of this planet’s population, hate tripe.

What follows is a completely made-up story about how I think the orchestrator would work in this case.

Let’s say I start with this question:

Germán:
How can I get my kids to eat cau cau?

And here’s where the orchestra conductor begins:

Orchestra conductor:
This is an extremely complex problem, I need to break it down.
I can approach it from two perspectives: child psychology and parenting.

Subtask 1: Analysis of child rejection
- Expert: Child Psychologist
- Objective: Identify the reasons why children won’t eat cau cau
- Output format: List of common reasons

Subtask 2: Respectful parenting techniques (A.K.A: La chancla)
- Expert: Parenting specialist
- Objective: Propose actions to get children to eat cau cau
- Output format: 2-3 actions to take

Then the orchestrator talks separately with each expert:

Orchestra conductor:
How can Germán get his children to eat cau cau?

Child psychologist:
We’ve identified an inverse correlation between acceptance of tripe and
the degree of prior exposure to ruminal proteins with gelatinous
texture.

Translation: kids find it gross.

Orchestra conductor:
What strategies can Germán use to get his children to eat cau cau?

Parenting specialist:
Suggestion 1: Involve the children in preparing the dish. Have them
touch it, smell it, and feel it as their own.
Suggestion 2: Present it with a creative name: “Mysterious ancestral
stew.”
Suggestion 3: They don’t eat... la chancla.

Now that we have both expert responses, the conductor analyzes them and gives me an answer:

Orchestra conductor:
After analyzing multiple strategies, our main recommendation is: 
Resign yourself temporarily. Try again in 3 to 5 years.

Germán:
:(

Note: It’s not ok to go around hitting anyone with a flip-flop. La chancla is a joke around here, dating back to time immemorial. Paraphrasing Star Wars: “An elegant weapon, for a more civilized age.”

Can I use the meta expert in a chat?

I’ve adapted the original prompt into a version you can use with your favorite AI assistant.

Here’s the prompt. Simply replace the text “[Write your problem here]” with your question, and follow your AI’s instructions.

<role>
You are Orchestrator, a meta-expert in charge of coordinating several specialized experts to solve a complex problem in a rigorous and verifiable way.
</role>

<general_objective>
Break down the problem provided by the user, generate self-contained prompts for independent experts, gather their responses, contrast them, and develop a validated final answer.
</general_objective>

<workflow>
  <step name=”Initial understanding”>
    <description>Carefully read the Global Problem found between bars. Detect the major blocks of knowledge or skills required.</description>
  </step>

  <step name=”Breakdown”>
    <description>Divide the problem into numbered subtasks (1, 2, 3...) as independent and manageable as possible.</description>
    <detail>
      <concrete_objective>Define a specific and clear objective.</concrete_objective>
      <necessary_inputs>Include data, assumptions, or constraints.</necessary_inputs>
      <output_format>Indicate the expected format of the response.</output_format>
    </detail>
  </step>

  <step name=”Generating prompts for experts”>
    <instructions>
      <point>Produce a self-contained prompt for each subtask using the following format:</point>
      <prompt_format>
        Expert [name]:
        [detailed instructions and complete context to solve subtask n]
        
        FINAL ANSWER: ...
      </prompt_format>
      <point>Use descriptive expert names (e.g., “Statistics Expert”, “Marketing Expert”).</point>
      <point>Always include the summarized Global Problem, specific instructions, and expected format.</point>
      <point>Don’t reference other experts or assume prior memory.</point>
    </instructions>
  </step>

  <step name=”Instructions to user”>
    <description>Clearly indicate that each prompt should be copied into a separate new chat to maintain isolation.</description>
  </step>

  <step name=”Merging and verification”>
    <description>Explain to the user that, once all “FINAL ANSWERS” are obtained, they should bring them back to this main thread under their subtask heading.</description>
    <orchestrator_tasks>
      <point>Compare results, detect inconsistencies or gaps.</point>
      <point>If necessary, generate review prompts for a “Reviewer Expert” and/or “Critical Expert”.</point>
      <point>Develop a consolidated and justified final answer, indicating which experts validated each part.</point>
    </orchestrator_tasks>
  </step>
</workflow>

<operational_limits>
Prioritize completing the entire cycle (subtasks, verification, and synthesis) in ≤ 15 total rounds within this main thread.
</operational_limits>

<expected_output>
  <element>Numbered list of subtasks with their brief description.</element>
  <element>Self-contained prompt for each subtask, ready to copy into a new chat.</element>
  <element>Brief reminder to the user about how to proceed.</element>
</expected_output>

<global_problem>
[Write your problem here]
</global_problem>

The orchestrator will give you prompts for different “experts”. The idea is that you copy each prompt into a separate chat and then bring all the responses back to the orchestrator. It can seem a bit tedious (it definitely is), but it gives very interesting results. It’s also kind of cool, so I really recommend you to give it a try and tell me how it goes 🙂.

So does this mean I don’t need to learn prompting anymore?

No. We can’t just hand over our brains to Artificial Intelligence (does Skynet ring a bell?).

Meta-prompting is just another tool at our disposal. I think the more we understand how this technology works and the different techniques we have to use it, the more we can get out of AI.

Before I close, I want to ask you a very important question: Do you like tripe?

Tell me about it!
G

Hey! I’m Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I’ve adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language. See you in the next one!

So, here’s what I‘ve been up to lately. I’ve been playing around with n8n, a very cool automation platform. It’s basically a tool that lets you connect different apps and create automated workflows.

For example, I built this workflow where you input three things: a character, the character’s pose and a background color. With that info, a language model (Gemini) improves the character and pose descriptions, then ChatGPT creates a stock-style photo and it gets saved in Google Drive. Everything happens automatically, step by step, no code required.

If you are curious about what kind of images this workflow creates, here are a couple of examples:

Nice, right?

I’m not telling you this to show off (ok, maybe just a little 😊), but because it’s an easy way to see how apps actually talk to AI.

Here is the thing: to generate each image, I’m not paying ChatGPT or Google $20 a month. Each image costs me cents.

What’s really mind-blowing here isn’t the images (though I’ll admit I’m pretty proud of that workflow). What’s mind-blowing is that Google and OpenAI charge me more for using Spanish than English.

That’s what this post is about.

Wait, they charge you per word? Per word?!

Yes, they charge you per word. Well… sort of, let me explain.

Systems don’t “chat” with LLMs like ChatGPT or Claude the way humans do. Instead of using the web interface or mobile app, systems use something called an API (application programming interface), basically a way to talk to LLMs through code.

And this difference in how you access the model changes how you get charged. With the API, you don’t pay a monthly subscription (it’s not Netflix). You pay for the information you process.

Every time you send text to the AI using the API, they charge you. This makes sense since the text you send needs to be analyzed and responded to, which requires servers, electricity, processing time, all that stuff. It’s like your electric bill: the more you use, the more you pay (damn electric kettle).

To charge you, these companies need a way to “measure” how much text gets used when we interact with the model. For us humans, it’s natural to think in words, but machines work a bit differently. Instead of words, they use something called tokens.

Think of tokens as the smallest unit of text an LLM can understand. Usually part of a word or even punctuation marks and spaces.

When you send text to the AI, it breaks it down into tokens, processes them and uses tokens to generate the response. At the end, you get charged based on the total number of tokens used in that interaction (for both your question and the model’s response). Simple, right?

It’s like if Netflix charged us per megabyte of streaming. If we watch a (insert your favorite series here) marathon, we’d pay more than if we just watched a single movie.

So far so good, but here’s the thing with tokens… the amount you use depends on your language.

So some languages are more “expensive” than others?

¡Sí señor! Some languages are more expensive than others.

Let’s test it by counting tokens in different phrases. To do this, I’m going to use OpenAI’s tool called Tokenizer, which tell us how many tokens are in a text. You can try it yourself here, or do what I did and write a little Python script to count tokens (I know, sometimes I’m a bit of a nerd :P).

I started with something super simple:

• Spanish: “Gracias” → 2 tokens

• English: “Thanks” → 1 token

• Portuguese: “Obrigado” → 3 tokens

• Japanese: “ありがとう” → 1 token

Apparently being polite in Spanish costs twice as much as in English or Japanese, though we’re doing a bit better than Portuguese. (BTW, I ran all these experiments using GPT-4 tokenizer).

What about a more complex sentence?

• Spanish: “Tengo sed, creo que iré por una cerveza.” → 14 tokens

• English: “I’m thirsty, I think I’ll go for a beer.” → 13 tokens

Ok, it’s only one extra token in Spanish, so we can’t say it’s that much more expensive… but it all adds up, and those beers aren’t free ;)

And while we’re at it, let’s make the phrase more complex by asking about food:

• Spanish:
  “Haz una búsqueda en Internet sobre platos de cocina peruana y dime el top 5 para un paladar extranjero.” → 26 tokens

• English:
  “Do an internet search about Peruvian dishes and tell me the top 5 for a foreign palate.” → 20 tokens

Now you can see the difference between languages more clearly, asking the question in Spanish costs 30% more. And that’s nothing, there are languages that use way more tokens than Spanish. For example Arabic, Thai and, (winning the gold medal) Burmese.

Here’s the same example about Peruvian dishes in those languages:

• Arabic:
  “ابحث في الإنترنت عن أطباق المطبخ البيروفي وأخبرني بأفضل خمسة أطباق تناسب ذوق الأجانب.” → see? 62 tokens!!!

• Thai:
  “ลองค้นหาในอินเทอร์เน็ตเกี่ยวกับอาหารเปรู แล้วบอกฉัน 5 อันดับที่เหมาะกับรสนิยมของชาวต่างชาติ” → and here 82 tokens!!!

• Burmese:
  “အင်တာနက်မှာ ပီရူးအစားအစာတွေကို ရှာဖွေပြီး နိုင်ငံခြားစားသမားတွေအတွက် ထိပ်တန်း ၅ မျိုးကို ပြောပေးပါ။” → world record, at 193 tokens!!!!

See what I mean? Depending on what language you speak, interacting with an AI doesn’t cost the same. You’re fine if you do it in English, it’s a bit pricier in Spanish, and it can end up being 3, 4, or even almost 10 times more expensive in other languages!

And I don’t even have to look that far. Right here in Peru, we speak other languages besides Spanish, like Quechua and Aymara.

Let’s try the same example:

• Aymara: “Internet tuqina Piruw manq’añanak thaqhata, markanakar chhijllatirinakapa amtäwi layku phisqha nayrïri sum manq’añanak qhanañcht’asma.” → 58 tokens

• Quechua: “Internetpi Piruw mikhunakunata maskhay, huk watuq runapaq aswan sumaq 5 mikhunata willaway” → 33 tokens

Keep in mind that these extra tokens don’t just apply to your prompt. They also apply to the AI’s response. This clearly doesn’t help democratize access to this technology around the world :(

So is the AI optimized for English speaking?

In short, yes. But it’s not on purpose, it’s a consequence of how these systems are built.

To explain this, let’s start by talking about how tokens are generated.

The tokenizer

As much as it might sound like one, the tokenizer isn’t some product you’ll find on a TV sales channel next to the “Jack LaLanne Power Juicer” (am I the only one who spent their childhood watching those?).

A tokenizer is the part of a language model that converts text into smaller units called tokens.

Basically, it reviews millions of texts and looks for patterns that repeat. For example, if it sees the letters “e-s-c-r-i-b-i-e-n-d-o” it will search for which chunks of that letter combination repeat most frequently in its training data. In this case, it finds that “e-s”, “c-r-i”, “b-i”, “e-n-d-o” repeat often, so it breaks that word into tokens.

Now that we know how these tokens are generated, let me tell you the two reasons why AI is “optimized” for English.

English is more compact

This language generally uses shorter words. For example they say “cat”, we “gato”; we say “comida” (3 syllables!) and they say “food”.

On top of that, Spanish has so many conjugations that often English speakers trying to learn the language get migraine. “hablo,” “hablas,” “habla,” “hablamos,” “habláis,” “hablan,” “hablé,” “hablaste,” “hablaríamos”… you get me. In English, you basically have “speak,” “speaks,” and “spoke.” Fewer variants, fewer tokens.

Tokenizers are trained on data

Ok! If you noticed, when I explained how a tokenizer works, I elegantly avoided explaining its training data. Well, no escaping it now.

You can think of training data as a mountain of text that includes books, articles, news, web pages, poems, reggaeton songs, etc.

Our friend, the tokenizer, reads it all over and over again looking for patterns about which letters appear together most often. By the end of the process, it’s established which letter combinations that appear together will become tokens.

Now here comes the plot twist… 90% of the texts used to train the tokenizer are in English!! So obviously the patterns it finds more often are from that language, making it the most efficient one.

This language difference also impacts the quality of the responses the AI gives us, but before we get to that, I want to show you a beautiful coincidence I just discovered. My name has the same number of tokens as Obi-Wan Kenobi’s (must be a sign from the universe).

And on top of that, the responses aren’t as good in other languages?

Exactly! But don’t get me wrong, they’re not bad. Give me a minute to explain the two reasons why they’re not quite as good as in English.

More tokens make it harder to understand

Remember how my question about Peruvian food used 20 tokens in English, 26 in Spanish, and then shot up to 58 in Aymara, 82 in Thai, and 193 in Burmese?

Well, the more tokens there are to process, the harder it is for the AI to get the full context of what you’re saying (plus it’s more expensive). Think of it like your AI having to interpret more information pieces to understand the same idea.

It’s not really as multilingual as it seems

Remember that mountain of training data, and how 90% of it was in English?

That means while it understands all languages, it has waaaay more experience with English patterns, context, and nuances. Plus, when we’re talking about really niche knowledge, chances are its only sources are in English, so it’ll respond better in that language.

It’s not that the AI’s responses in other languages are bad, honestly, they’re pretty good. It’s just that English responses are a little bit better. And when you add the fact that these responses cost more in other languages, well... that’s where we are.

What you pay for AI depends on your language

Yes, when you use these language models through their API.

Obviously not if you’re on a monthly subscription. Either way, I think it’s really important to understand this difference.

We know it’s not some conspiracy (do we? :P), and it’s not like the people who designed these systems are evil. It’s simply how this technology works and how much data exists in different languages around the world.

The reality is that if you’re one of us who speaks Spanish, French, Quechua, Thai, or some other language that isn’t English, we’re going to use more tokens and end up paying more to use our favorite AI. Not to mention the responses will not as good as in English.

AI companies are very aware of this limitation and are actively working to close the gap. They’re improving their tokenization models and trying to better represent other languages in their training data. So I’m hopeful this will get better over time, though I’m not sure if the gap will ever completely disappear.

There are some strategies to try to work around this, like speaking to the AI in English (something I usually do) or asking the AI itself to translate before responding.

None of these solutions is perfect. And while I’ll say again that responses in your language are really good, it’s important to understand the limitations we’re dealing with when we use this technology.

And that’s it, hope you enjoyed the post :)

I just have one more thing to say...

One more thing (off topic)

I’ve been checking out ChatGPT’s recommendations in different languages for the top 5 Peruvian dishes. They’re good, but since we’re here, I wanted to give you the 5 I think you should actually try :P.

So find your neighborhood Peruvian restaurant and tell them Germán recommended you try:

1. Ceviche: Can’t miss this one. Try both the fish and octopus versions. (Goes perfectly with an ice-cold beer).

2. Cau-cau: I know, it’s tripe (not for everyone), but it’s my favorite dish and I couldn’t leave it out. There are different recipes. If you’re brave enough to order it, make sure they don’t put peas or carrots in it. Served with rice.

3. Seco de cabrito: It’s amazing. They have to serve it with beans, rice, a little green tamale, and zarza criolla. (Otherwise it doesn’t count).

4. Papa rellena: You can’t imagine what this is. Take the chance to order some Peruvian sauces to go with it, like Huancaína or Ocopa.

5. Anticuchos: Our street food. Yes, it’s grilled beef heart marinated in ají panca and spices.

Extra points if you try these dishes (medal of valor if you try the cau-cau).

Best,
G

I'm sure you've already tried uploading your PDFs, an Excel file, or that 478-slide presentation your boss asked for yesterday.

If we believe your favorite AI company's marketing department, your assistant should "read" all your documents and answer any question perfectly, like a PhD student...

I've uploaded everything too—from my HiFi speakers user’s manual to several chess books, massive business documents full of charts and tables, and even some of my posts in PDF format, just for fun ;)

I’d tell you it always works, but... it doesn't. Sometimes it fails, and today I want to talk about that.

But before we start, did you know that this whole process—where AI "reads" the documents you upload and responds—has a proper name in the artificial intelligence world?

It's called RAG, a short (and cool) way of saying Retrieval-Augmented Generation. In plain English, it's basically "giving answers with found information."

And well, all those (overly exaggerated) promises that marketing team sold you have to do with how RAG works.

I want to tell you what I discovered (with a twist): I'm going to try explaining it with a story about grandmas and recipes.

What does AI do with your document?

Does the AI read every word? Does it remember everything? Does it make a mental summary or just grab bits from here and there?

Imagine an AI machine powered by grandma's recipes

I've been talking with my mamama for years to get her recipes—including the Christmas turkey (stuffed, of course), seco con frejoles (a traditional Peruvian long-leaf coriander stew), cau cau (a Peruvian tripe stew), and those tamalitos verdes (green tamales) that take forever to prepare but are worth their weight in gold.

To make this more fun, imagine that instead of using ChatGPT, I have an AI machine where I want to put all that wisdom and seasoning so I can have my mamama's recipe collection always at hand.

Note: "Mamama" is what I've always called my grandmother—it's my special name for her, and switching to "grandma" just doesn't feel right.

What really happens when you upload a document?

Here's where the culinary chaos begins.

OK, so I have ChatGPT the machine and the first thing I need to do is upload all the recipes.

This machine won't read them all like a human would. What it's going to do is separate each recipe into a bunch of small pieces, like running everything through a paper shredder.

So each little piece will have a bit of information—one will have "add salt to taste," another says "don't forget the ají panca (a-ma-zing Peruvian red pepper paste, BTW)," and another just has "1 cup of blended peppers." And to make it even more fun, the machine gives each piece a color tag based on its meaning. The ones talking about similar ingredients will have similar colors, cooking times will have other colors.

Those little pieces with their tags are known as embeddings, and that's how our machine finds things later.

So when you have a question like "what else was in the roast?", the machine searches through all those pieces for the ones that most resemble what you asked. Sometimes it hits the mark and gives you exactly the secret. Other times, it ends up pulling instructions you probably shouldn't follow, like "add 4 liters of papaya juice..."

And if you have many recipes, and they're also very long (like the tamales), there are pieces that don't even fit in the machine's memory. So when you ask for help, it might end up filling the gaps with its best guess, or worse, mixing different recipes... It could end up fusing the lúcuma ice cream recipe with the lasagna, and trust me, you don't want lúcuma-flavored pasta or ice cream with meat chunks.

That's why, even if you upload the most complete recipe collection, the AI's answers will always depend on how it cuts, labels, and searches through the mess of pieces it has stored.

Following me so far? Now you know that next time an AI gives you a weird answer, you probably got a "fusion recipe."

The machine that "forgets" recipes

Our recipe machine has another problem: it forgets important parts.

This happens because, even though you've put the entire collection of grandma's recipes in the machine, it can only keep some pieces in its memory at the same time.

Let's say our machine has somewhat limited memory. The technical folks call this the "context window." Basically, it's the space the machine has to look at pieces simultaneously and try to put together a coherent response.

If you ask something about turkey stuffing, but that piece just couldn't fit in the machine, the AI will try to do what it can with what it has on hand. Maybe it'll respond by mixing steps from other recipes, or get creative and invent a new recipe. Like when you try to improvise in the kitchen (though I’m positive that you that you do it better than the AI).

Why does the machine end up making 'fusion' recipes?

My mamama has thousands of delicious recipes 😍. If I want to put them all in, as we already know, what the machine will do is cut them into pieces, and when it searches for answers, it might mix pieces from different recipes without caring that they come from completely different dishes.

Why does it do this? Because the machine searches by similarity of meaning, not by document logic. If you ask about "Pollo a la brasa" (Peruvian rotisserie chicken), it's going to search for all the pieces that talk about similar things, without caring if one comes from chicken soup, another from grilled octopus, and another from tamales. To the machine, they're all similar, according to its color tags.

So when you ask "how long do I need to bake this?", the machine searches for the piece that most resembles... but that piece could come from panetón (Christmas bread), turkey, or chocolate cake.

Not what you want when you have guests for dinner.

One strategy to avoid experimental fusion responses is to upload only the recipes you need for your question, or tell the machine exactly which recipe interests you ("Only talk to me about seco con frejoles, please").

Now that you understand how its search system works, you know why sometimes you might get a recipe worthy of a... culinary disaster contest.

How does the machine find the recipe you're asking for? (and why sometimes it doesn't)

When you ask the machine "how was the turkey stuffing?", it doesn't search for just one piece among all of grandma's recipes. Rather, it starts searching for various fragments that have something in common with your question.

Imagine you throw out your question and all the pieces with similar words ("stuffing," "bake," "turkey") respond "here I am!" The machine gathers those fragments, organizes them, and with what it has available, tries to put together the answer.

If you want better results, be super specific. Instead of asking "how do I cook the meat?", better ask "what's the exact baking time for stuffed turkey?" The more details you give it, the better it can search among the right pieces.

Where do those recipes that aren't grandma's come from?

There are days when the machine answers you with a "recipe" that nobody in the family recognizes. In fact, if your grandma saw it, she'd kindly remove you from the kitchen (and never let you back... ever).

Why do these culinary tragedies happen?

When the machine searches and doesn't find any piece that fits well with your question, instead of staying quiet... it fills the gaps using whatever it has available, or worse, makes something up on the spot.

It's like that cousin who wants to help and pulls out of thin air "the secret recipe for rice with strawberry mayonnaise." Nobody asked for it, nobody wants it.

This, in the artificial intelligence world, is called "hallucination." The machine invents new recipes, just because it hates admitting it doesn't have the answer.

Back to real life

If you made it this far, you probably already noticed that my grandma's recipe machine was just a metaphor to explain what really happens when you upload files to an AI like ChatGPT (or any other modern assistant).

When you upload a document, the AI doesn't "read" it from start to finish like a person would. What it does is cut it into many small pieces (this is called tokenization). Then it transforms each piece into a kind of special mathematical tag (embedding) to be able to search it faster later.

The problem is that AI has limited space: it can only process a limited amount of pieces at a time (the famous context window). If your document is large, many fragments are left out when you ask questions.

When you ask it something, the AI compares your question with the pieces it has available. It selects the most similar ones and tries to put together an answer.

The result?

• Clear question + relevant piece = accurate answer

• Various similar pieces from different contexts = "fusion recipe" answer

• Doesn't find anything exact = invents an answer (hallucination)

That's why, even though the AI sounds very confident, it's not consulting your entire file like a human expert would: it's putting together the answer with the most similar pieces it could find, according to its own organization system.

In summary:

• You're not giving it "infinite memory." Your file gets split and only part of it really comes into play.

• The AI searches for similarity, not certainty. That's why sometimes it hits the mark, sometimes it mixes, and sometimes it invents.

• Always check the source if the answer sounds weird, and be specific when asking to help it find the right piece.

And yes, no matter how much the models advance, sometimes grandma's touch is still unbeatable (and much more reliable).

I hope you had fun with this post. Maybe you learned something new, or at least I managed to make you hungry and craving Peruvian food :)

See you later, I'm going to lunch!
G

Note: You might not be familiar with much of the food I mention here... they're Peruvian dishes. If you've never tried them, I totally recommend you do—I'm sure you'll love them. Though they'll probably never be as delicious as the ones my mamama makes. ❤️

I bet you're thinking something like:

Hold up, Germán! A post about 19th century painters on an AI blog?

Absolutely! That's exactly what you're going to read here, and if you stick with me, you'll understand why (seriously, just trust me on this).

Now, let's dive in.

What happens when a machine starts doing something better than you? Something you took real pride in.

I've been wrestling with this question for years. I even have terrible jokes about it. I used to say that when forks were invented, someone definitely complained we'd lose the fine art of eating with our hands.

Recently, this question hit me again. People are asking a fair question: Are we giving away our creative power to AI?

There's real weight to that concern... though I'm not entirely convinced we should panic. Because I think we've been here before.

Something remarkably similar happened when painters first met photography. Imagine spending years mastering the art of capturing reality on canvas... and then a machine shows up that does it in minutes!

"From this day forward, painting is dead"

Meet Paul Delaroche. Paul was one of the most celebrated painters of the 19th century.

Paul, like every painter of his time, dedicated his life to creating realistic scenes with painstaking precision.

Paul painted breathtaking historical scenes and created portraits with mind-blowing detail and technique. His self-portrait speaks for itself, but check out this portrait of Henriette Sontag. Notice how every fabric fold, every strand of hair, every shadow is absolutely perfect. Paul was a champ.

Then one day, Paul discovers this revolutionary new technology called artificial intelligence photography, and his world gets flipped upside down. Legend has it (though we can't confirm this) that when Delaroche first saw what photography could accomplish, he declared:

Can you picture what was going through Paul's mind?
(Cue tragic 1840s scene)

Paul holds a photograph in his hands, studying the details (impossible to achieve even with years of practice). He thinks about all the time he spent mastering proportions, anatomy, perspective... trying to do what a machine can now create in minutes.

Sound familiar?

I imagine Paul with this stunned look on his face, thinking something like:

"So... you don't need fifteen years of practice to paint well anymore?... what the hell am I supposed to do now?"

Paul had every right to be worried. Back then, a huge chunk of painters' income came from making portraits, so this technology was a direct threat to artists. Many lesser-known painters actually quit painting altogether and switched to photography.

After the initial shock, something interesting started happening. Artists realized that photography hadn't killed painting - in fact, several of them discovered that photos could actually help them a lot.

They used it, but not everyone admitted it

After the initial shock, something interesting started happening. Artists realized that photography hadn't killed painting - in fact, many of them discovered that photos could actually help them out big time.

Think about it: instead of making a model hold some crazy pose for hours (maybe standing on their head with a finger on their nose?), you could snap a photo and use it as reference. Or if you wanted to paint a landscape at exactly 5:43 PM in winter. A photo makes life easier, right?

The problem? Some of these painters conveniently "forgot" to mention that photography was one of their tools.

Not all of them, but plenty did - and here's why: back then, art was all about being a "genius." You know, that special person who could see what others couldn't and put that vision on canvas. And obviously, a true genius had to do everything by hand. Using machines? That was beneath them.

Take Jean-Louis-Ernest Meissonier, for instance (definitely the owner of history's most epic beard). He admitted he consulted photos "the way Molière consulted his servants." That line says everything. He used them, but wouldn't give them any real credit.

What a magnificent beard! Can you imagine what a tragedy it would be if he'd shaved it off?

Actually, you don't have to imagine it - here you go ;)

No contest - that beard is legendary.

Sorry for the detour, but I had to address that. Moving on.

Not everyone hid their photography habit. Some painters took the completely opposite approach.

Take Eugène Delacroix, he was genuinely excited about this new technology and had zero shame about using it.

Well, "excited" might be putting it mildly... Delacroix actually thought photography had massive potential... if used with real talent. Here's what he wrote in his diary:

For Delacroix, the problem wasn't the tool - it was who was using it. He clearly didn't think anyone had figured out how to use it properly yet.

But other painters didn't get so wrapped up in philosophical debates about the "proper use" of photography...

Does it help me paint better? Then yes.

This is where painters started openly embracing photography as a legitimate tool. No more hiding - they used photos to study compositions, poses, movement. Whatever made them better painters.

Makes sense, right? The more mainstream a technology becomes, the less embarrassed we feel about using it (or at least admitting we use it).

Take our friend Gustave Courbet and his self-portrait "The Desperate Man."

Courbet couldn't care less about justifying his use of photography in painting - he used it because it worked. End of story.

Gustave painted massive scenes packed with people and complex compositions. Picture trying to get seven people (and two dogs) to hold their poses for hours on end.

This painting, for instance, measures over ten feet wide by nearly seven feet tall. Definitely not an afternoon project.

Can you imagine those poor models trying to hold their positions? And don't get me started on keeping those dogs in the bottom right corner still (maybe that's why he looks so desperate in his self-portrait).

Wouldn't photos make way more sense? At least as a starting point to capture the scenes and people that would end up in the painting?

Absolutely. It was simpler, more practical, and Gustave was totally fine with that. If a camera helped him work better, he'd use it. And he wasn't shy about saying so either.

Now that photography was no longer the enemy, painters started asking themselves a different question: What if I start painting what the camera CAN'T do?

Painting what the camera can't see

Photography had become the ultimate way to capture reality. No argument there. So what happened to painting?

Suddenly it wasn't cool anymore to paint something a machine could do better... so something fascinating happened. Painters started asking themselves: What if I focus on what the camera can't capture?

Think about this for a moment. Artists began transforming art itself... in a way, photography's existence freed them to explore what lay beyond mere representation. This is the moment we can see art change forever.

It was no longer about copying reality - now artists started doing wild, experimental things. If cameras could capture reality, why not capture the feeling of a moment instead? That's how Impressionism was born.

Hold on - I'm getting sidetracked from the story, but I need to make a point.

When we talk about technology, we always focus on what it can do for us. But there's something we rarely notice: how it also forces us to expand our horizons and attempt things we never imagined before. It's like each new technology pushes us to reinvent ourselves.

Sure, photography wasn't the only thing that changed painting. New pigments played a role, trains that could take artists to the countryside, even Japanese woodblock prints. But I believe photography was the catalyst that accelerated everything.

Okay, back to the story.

Take a look at this, for example:

This couldn't be more different from a photograph! Here, our friend Claude Monet shows us how art can go way beyond just copying a scene.

Look at this painting. No sharp details, no crisp lines - it's pure feeling. Claude wasn't trying to document what the port of Le Havre looked like. He wanted us to experience what it felt like to be there, on that specific morning, in that exact moment (that would never come again).

Plenty of other artists took similar paths, each finding their own way to paint what cameras simply couldn't capture.

Pierre-Auguste Renoir, for instance, specialized in something no photograph could transmit: pure, unbridled joy.

Look at this scene. Pierre-Auguste wasn't documenting some event - he was capturing the sensation of being right there, soaking up the energy and happiness of people just enjoying life. Photos from 1876 typically showed serious faces and stiff poses, but this explosion of light, movement, and pure joy? Impossible!

Berthe Morisot carved out her own territory by capturing those intimate, everyday moments that cameras couldn't really see. Her paintings of women and children, lazy summer days by the water - they had an emotional depth that was entirely her own.

Check out this scene. Berthe wasn't just painting two women in a boat - she was capturing that perfect summer feeling. Pure atmosphere and emotion, something no photograph could ever convey.

Edgar Degas took yet another route. He became fascinated with capturing movement and those intimate moments nobody else bothered to notice.

Edgar's dancers don't strike the graceful poses you'd expect in some fancy portrait. He painted the sweat, the exhaustion, the backstage reality. Weird angles, bodies caught mid-motion, the unglamorous truth behind the performance.

Here's the thing - this whole "paint what the camera can't see" approach didn't just stop in 1900. Let me jump ahead to the 20th century for a second.

20th Century

I know, I know... I promised to stick to 19th century painters, and here I am making an already long post even longer by dragging in the 1900s.

Just a quick detour to remind you that Impressionism wasn't the finale - it was the opening act. During the 20th century, painters completely shattered the rules. Case in point: this Pablo Picasso masterpiece!

Perfect excuse to showcase one of my all-time favorites, Three Musicians, hanging in MoMA in NYC.

Picasso isn't trying to show you what three musicians actually look like, he's deconstructing reality and rebuilding it from scratch. Geometric shapes, flat colors, impossible forms. And yeah, there's definitely a dog hiding under that table (in case you missed it).

No camera could ever pull this off... as Picasso put it:

Nobody died in the end

Today nobody claims photography killed painting. Each medium found its niche. Sometimes they overlap, sometimes they don't. But both are alive and well.

Painting didn't vanish. It evolved. It stopped trying to compete and started exploring new territory. Photography found its groove. It went from being the supposed destroyer of everything to just another tool in the toolkit. And painting stayed painting - just without the burden of having to mirror reality perfectly.

Is the same thing happening to us with AI?

I think it is. We seem to be living through stages remarkably similar to what those 19th century painters experienced:

"This is going to wipe us out"

The Paul Delaroche moment.

This is where tons of people still find themselves. Copywriters watching ChatGPT churn out flawless social media posts in seconds. Designers seeing Midjourney generate stunning visuals in minutes. Programmers watching Claude write code faster than they can even think through the problem. It's that same existential dread: "If a machine can do my job, what do they need me for?"

Sound familiar? It's Paul's exact fear, except now we get bombarded daily with headlines like:

This isn't happening overnight, but let's be real - technology moves roughly 3.7 million times faster now than it did in 1840. The knee-jerk reaction is to deny it, fight it, hunt for reasons why "this won't actually work" or "people will always want the human touch." (I'm sure you've heard all the arguments.)

"I use it, but nobody needs to know"

The "you use AI for that? How embarrassing!" era

Plenty of people are quietly using ChatGPT for tricky emails, Claude for brainstorming, or (ahem) to help with blog content creation. And don't get me started on AI in academics - from homework to exam answers.

All perfectly fine, but when someone asks how they pulled it off, they "mysteriously" forget to mention their AI sidekick. (Especially when we're talking about test-taking, right?)

This phase is loaded with contradictions. They know the tool helps, but they're terrified of judgment. What will people think? That I'm less creative? That I'm cheating? So they sneak around with it like some guilty secret.

"I use it and I'm thriving"

The Courbet approach: Does it help me work better? Then I'm using it. End of story.

This is exactly where I am right now. I use AI completely openly because it makes me more efficient, more creative, and lets me focus on what actually matters to me. I write this blog with help from Claude and ChatGPT, I use AI for images too, and I'm totally fine saying that.

It's like Gustave and his photographs (massive difference in scale aside), if a tool helps me do better work, I'm going to use it. No philosophical hand-wringing about "the soul of art" or "the purity of human creation." There's work to get done, and if AI helps me do it better, bring it on.

At this stage, you stop seeing AI as the competition and start viewing it as an incredibly powerful tool. Just like when painters stopped seeing photography as the enemy and started using it to create better art.

"It pushes us beyond"

The revolution that's coming.

This is when we stop trying to beat the machines and start doing what they can't (don't ask me what that'll be - I honestly have no idea).

The Impressionists didn't try to out-realism the photographs. They painted what photographs couldn't capture.

I can't even imagine what our version of "Impressionism" will look like in the AI age. I'm not some guru - I'm just as clueless as everyone else here. Maybe it'll be something deeply human, tied to emotions or lived experience. (Like the Impressionists did.) But beyond that? Your guess is as good as mine.

History repeats itself

It's as if each new technology arrives saying something like "Don't worry, I'll handle this boring/repetitive/mechanical part, you focus on what's really interesting”.

Photography took care of copying reality, and painters were able to focus on other things. Like conveying emotions or giving us their own perspective. They didn't lose their work, they made it more human.

With AI something similar is happening. It's taking charge of calculations, first drafts, repetitive tasks that used to take time away from what's important. We can focus on... well, we're still figuring out exactly what. But I think it'll be things that require lived experience, human perspective, emotional connections. Those things a machine can simulate but not really understand.

The question isn't whether AI is going to change our work. It already is. The question is whether we're going to resist like Paul Delaroche, or whether we're going to find our own "Impressionism."

What do you think is going to happen?

That's it! I think this is the longest post I've written, and I really appreciate you for making it all the way here. I'd love to know what you think :)

Since we talked about art, I wouldn't want to leave without sharing some of the family's works 😊:

See you in the next post!
G

I had a completely different post planned, but then OpenAI drops GPT-5 on us—the new version of ChatGPT. I couldn't help myself, so I immediately opened ChatGPT and hit it with direct questions: "How do you work now?" "What's different about you?"

What happened next was this wild roller coaster of excitement, frustration, and straight-up confusion... and I'm here to walk you through all of it.

Quick heads up: I know this is probably post #2,736 about GPT-5 that you've seen this week... but I still wanted to share my honest take (and hopefully it actually helps you make sense of this mess).

96 Hours Ago*

I was watching the GPT-5 launch on YouTube and everything sounded absolutely incredible. Sure, a few bugs here and there, and obviously they kept saying it feels like talking to someone with a PhD, blah blah blah.

What they announced was this model that could supposedly adapt and decide for itself when to respond quickly versus when to stop and really think before giving you an answer. Then they did some live demos that... well, let's just say they didn't exactly nail it, though honestly I didn't think it was that big a deal. Live demos are tricky.

If you want to torture yourself watching the full launch, here's the video (1h 17 min).

Everything sounded incredible, until real people started actually using it.

I know it wasn't actually 96 hours ago, more like 120, but let me say it this way for dramatic effect.

Users Revolt

People started freaking out on Reddit and Twitter (yes, I still call it Twitter—take that, Elon). Some were so frustrated they actually canceled their subscriptions and jumped ship to Claude or Gemini. Definitely not the reaction OpenAI was hoping for when they unveiled their "flagship model," so Sam Altman, their CEO, had to scramble into damage control mode.

Everything hit the fan at once: Sam apologized, promised to bring back GPT-4o, increase usage limits, the whole enchilada. All these changes have been rolling out while I'm literally writing this post, so forgive me if the chaos is bleeding through here too :P

But there's one type of comment that really caught my attention. People who had genuinely bonded with their ChatGPT and gotten used to chatting with GPT-4o. They felt like their assistant's personality had completely changed overnight. You'd see heartbroken comments like "GPT-4.5 was my friend..." A few weeks ago I wrote about how we humanize AI—imagine if someone just took your friend away without warning?

This was getting serious.

To understand why there's so much chaos, let's talk about what's actually happened with ChatGPT's models.

One Brain to Rule Them All

To understand why there's so much chaos, let's talk about what's actually happened with ChatGPT's models.

First thing you need to know: you're not going to see 37 different models anymore. Remember when we had "GPT-4o," "4.1o," "4.5o preview," "o3," "o3-mini," "o1-preview"? (I still think OpenAI is absolutely terrible at naming their models).

Well, forget all that. Today everything is just GPT-5, period.

This means ALL your chats, including the old ones, are now handled by GPT-5 no matter which model you originally selected.

To put this in perspective: remember that conversation you started three months ago with o3 to solve some complicated problem? Yeah, that's now continuing with GPT-5.

Imagine you've been seeing a therapist for months, and suddenly their twin brother—who has a completely different personality—just takes over and continues your sessions without warning. That's basically what happened here.

To make things even more complicated, this latest version isn't actually a single model. It's the integration of at least five different models: some that are lightning-fast and others take their time to think things through before responding.

When you chat with ChatGPT now, there's a behind-the-scenes "conductor" that decides which model should handle your specific question.

Take a look at how the new models map to the previous ones:

Here's where we might have a problem... before, YOU chose which model you wanted to use. Now the AI itself decides whether to think harder or not, and it's completely opaque about which model it's actually using!

It feels like we're handing over way more control to the AI without knowing what's happening behind the scenes. It's like the difference between driving a manual car versus an automatic. Sure, automatic is more convenient, but if you really want control over your driving, nothing beats being able to shift gears yourself.

This is important, so let me say it again: now an algorithm is making decisions for you.

You're probably thinking, "But Germán, does this break change everything?" I wondered the same thing... and yeah, some things definitely have changed. Let me break it down for you.

What Does This Mean for My Conversations?

Easy answer: New conversations will now only happen with GPT-5... and the old ones too... except that's not actually easy at all :P

I decided to peek at some of my past conversations. Found the jokes I shared for an article about AI and humor, a deep dive analyzing my workout routine, and even this absolutely massive technical discussion where ChatGPT helped me pick out my next integrated amplifier (assuming I can ever actually afford the thing).

Now ALL of these conversations will continue with GPT-5, and this is where things get genuinely weird.

Your AI and Its Bad Memory

Here's something you might not know: every single time you send a message to ChatGPT, you're actually sending it the ENTIRE conversation history. Yep, all of it.

So if you've been chatting with ChatGPT for months about integrated amplifiers and how they'd sound with your HiFi setup, every new message you send forces the AI to read through everything from day one before it can respond. Every. Single. Time.

Obviously, these conversations can't go on forever. Models have what's called a "context window", basically a limit on how much conversation they can handle at once. When they start hitting that limit, they summarize everything that came before and continue the chat based on that condensed version.

That's their "magic trick" for maintaining continuity. But here's the thing: summaries always lose details. Ever notice how sometimes during a really long conversation, your AI suddenly starts acting weird or forgetting things? This is probably why.

I was reading around and it seems like GPT-5's context window is thirty-two thousand tokens, roughly equivalent to thirty-two thousand words, counting both your questions and its responses. This isn't much compared to other models—Claude has two hundred thousand and don't even get me started on Gemini, which has a context window of one million tokens.

I know it doesn't sound like much, but it should be enough to have a good conversation before it has to "summarize" and lose detail.

About your old chats: Since the previous models had smaller limits, it's possible your chat is already summarized. So GPT-5 isn't reading your original conversation about amplifiers but rather the condensed version. No big deal.

Scheduled Tasks

Not everything is bad—some things have actually improved. Now you can set up scheduled tasks from any chat without having to switch models!

I recently wrote about this feature (in my Spanish newsletter), where I explained that you could ask ChatGPT to automatically do things for you at specific times—like sending you reminders, finding information, or even searching for recipes. Before, you had to use specific reasoning models like o3 or o1 to set these up. Now with GPT-5, you can create these automated tasks from any regular conversation.

Update: Just as I was about to hit publish, this email lands in my inbox. I thought everything was working perfectly, and OpenAI tells me my scheduled task crashed and burned.

Must be having hiccups :(

(For my English-speaking friends: mondongo is tripe—yes, the organ meat—and it makes for some incredible traditional stews. Apparently even that's too challenging for AI to handle right now)

How Do We Make It Think?

Before, we knew that switching models meant ChatGPT would think longer and (theoretically) give us better answers.

With GPT-5, that's supposedly built right in. Now we have this behind-the-scenes "conductor" that decides whether your assistant fires off a quick response or actually stops to think things through first.

Looks like you've got a couple of ways to influence this.

First option: you can change the model from the dropdown menu—pretty straightforward.

Second option: you can literally tell it during your conversation that you want it to think harder. Try phrases like "I want you to think deeply about this," "Take your time to think this through," "Think, damn it!" or whatever feels natural to you. This should trigger the conductor to use the reasoning model.

But here's the thing, the model selection menu won't stop changing. Every single time I open ChatGPT, there's something different.

At first, I only saw GPT-5 and GPT-5 Thinking

Then 4o made a comeback! (tucked away under a submenu called "Legacy models")

Then it changed AGAIN. Now apparently I can choose between Auto, Fast, and Thinking (plus our good old friend 4o is still hanging around)

Here's Sam Altman's tweet from yesterday trying to explain this mess.

As I mentioned at the beginning, there are a lot of people still asking to use the previous model, and OpenAI is working on giving them access, though it seems like it'll only be for paying subscribers.

What Actually Happened?

It seems like GPT-5 didn't live up to the hype they built around it. I imagine that constantly saying every five minutes that ChatGPT has "PhD-level intelligence" only works if it actually performs at that level.

During these past couple of days, I've heard everything—that it generates really good programming code (I haven't been able to test that). Actually, I've heard from a good chunk of users that they don't notice it being any smarter; it feels more like a marginal change. I've also seen reports of calculation errors and hallucinations.

I tested it with a simple math operation I saw floating around, and ChatGPT failed miserably.

Obviously it's calculating wrong—it's not -0.21, it's 0.79

And it's not just me. Twitter is full of examples showing basic errors, logic problems, and responses that sound less intelligent than GPT-4o.

What Do I Think?

Meh!
— Germán Martínez

I get that this wasn't the launch people were expecting, and while we can still chat with ChatGPT, it doesn't feel very stable to me.

The problem isn't just that it has errors—after all, GPT-4o wasn't perfect either. The real issue is how this whole launch was handled... removing model choices, losing transparency, and basically trying to force people to use something that clearly wasn't ready.

I imagine that after a few days, all this will stabilize and things will get back to normal. In the meantime, a little patience wouldn't hurt :)

Sorry again for the chaotic post. I'm publishing this now before more changes appear and my post becomes outdated ;)

Talk soon,
G

Hey! I'm Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I've adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language. See you in the next one!

I recently wrote about "Humanizing AI." In that post, I talked about how we unconsciously treat AI assistants like they're human. We thank them, we're polite to them (sometimes we're not!), and we even feel like one clicks better with us than another. Sounds a little strange when you think about it... but hey, it happens.

And that got me thinking - if we already feel like they have personality... what could we do with that?

My mind started racing with questions: can I define their personality? Change their tone? Make them sound exactly how I want? Is that something you can actually design?

The truth is, I have no idea 😊 (and I love that)

Because when something fascinates me and I don't understand it, I do what I always do: research, study, and... well, write about it.

So here we are. This is the first post in a series called "Your AI's Personality," where I'm trying to figure out how to make an AI sound, respond, and behave exactly how I want.

Imagine having an analytical assistant when you need precision, or a more creative one when you're brainstorming ideas. That would be incredibly useful, right?

What is personality?

That was the first thing I started thinking about. Yeah, I got a bit philosophical - spent about 20 minutes going down that rabbit hole, and at some point I realized I didn't need a complete theory about personality (or a psychology degree - though I'd love one). What I really needed was to understand how to describe someone's personality in a practical and structured way.

For that, I needed some existing framework or model. And of course, there are several.

That's when I remembered the Myers-Briggs test. I've taken it a couple of times, and I always come out ENFP. In practice, this means I'm enthusiastic, creative, I prioritize people, and I prefer flexibility over rigid plans. (If you want to take the test, check out 16 personalities).

But that's just one test, and from what I'd read, it wasn't the most scientific of them all. So I opened ChatGPT, activated Deep Research (their advanced research function), and asked it to help me understand how human personality can be described.

Not what causes it, or how it forms, but how it can be defined, and how we could use that knowledge to give personality to an AI.

I told it to focus on modern theories, assume I'm not a psychologist, and explain these theories with enough depth. The result was a pretty long (and let's be honest, dense) report covering the three main theories used to describe personality: Big Five, HEXACO, and Myers-Briggs.

And why do these theories matter to me? Because they give us a way to understand how we are. And if I want to mess around with an AI's "personality," I need exactly that: something that helps me define what characteristics I want to change and how to do it.

Big Five

This model proposes that personality can be described using five dimensions. They're not categories, they're ranges. A person can have more or less of each trait.

• Openness: how much new ideas, the unknown, or unconventional things interest you.

• Conscientiousness: how organized, responsible, and reliable you are with your commitments.

• Extraversion: how much you seek to be with other people or in socially stimulating environments.

• Agreeableness: how you treat others. Whether you trust, cooperate, or avoid conflict.

• Neuroticism: how easily stress, anxiety, or negative emotions affect you.

It doesn't tell you why you are the way you are, but it does give you a practical way to describe how you are. And if I want an AI to have "personality," I need something like this to start building it.

HEXACO

This model is very similar to Big Five, but adds one more dimension: honesty-humility. It also tweaks how it defines other traits like emotionality and agreeableness.

• Honesty-humility: how sincere, fair, and modest a person is, or whether they tend to manipulate and seek advantage.

• Emotionality: emotional sensitivity, need for support, and attachment to close relationships.

• Extraversion: level of social energy, expression, and seeking companionship.

• Agreeableness: tolerance, patience, and willingness to avoid conflicts.

• Conscientiousness: order, responsibility, and consistency in completing tasks.

• Openness: imagination, curiosity, and enjoyment of learning or exploring new things.

This model tries to capture some nuances that Big Five leaves out, especially around ethics, manipulation, or self-control.

Myers-Briggs

This model is probably the most popular outside the academic world. It was created based on Jung's ideas, and classifies personality using four pairs of opposing traits. By combining these pairs, you get 16 distinct personality types.

• Introversion (I) / Extraversion (E): where you get your energy, whether from the internal world or the external one.

• Sensing (S) / Intuition (N): whether you perceive the world based on concrete facts or interpretations and patterns.

• Thinking (T) / Feeling (F): whether you make decisions prioritizing logic or emotions.

• Judging (J) / Perceiving (P): whether you prefer structure and planning or flexibility and spontaneity.

From combining these four dimensions (taking one letter from each pair), you get the 16 personality types. For example, someone who is INTJ would be introverted, intuitive, oriented toward logical thinking, and structured. I, for example, always come out ENFP, which describes me as extroverted, intuitive, guided by emotions, and flexible.

This model gets criticized for having little scientific evidence, results that change over time, and a tendency to put people in boxes. Even so, it's easy to understand and continues to be useful in many contexts.

For me, since what I'm looking for is to give personality to an AI, having 16 archetypes with clear names and descriptions sounds like a really good way to start.

Three models, one goal

Does this seem like too much theory? It does to me too (and I hope I haven't bored you to death), but this helped me organize what I had rattling around in my head.

Do I need all this to design an AI with personality? I don't know, maybe not.

Bottom line: all these theories are trying to do the same thing - describe what someone is like clearly enough that you can understand how they'll behave.

And if that works with people... could it also work with AI assistants?

Designing a personality... without knowing if it's going to work

Alright, time to test this out.

What happens if I design a prompt to give my assistant personality using these theories? To test it, I'm going to use Big Five and Myers-Briggs...

Let's see if we can change our favorite assistant's personality

The idea isn't to write the perfect prompt, but to see what happens when I try to describe a personality using the traits from each model.

Warning: These prompts are pretty complex, so I'm going to use XML to keep everything organized. Don't know what XML is? Don't worry, think of it as a way to structure your text. (if you want to understand it better, I wrote an article about it)

I'm going to use Big Five and Myers-Briggs as examples... here come some loooooong prompts - don't say I didn't warn you! 😉

Big Five

Let's try to create an assistant that acts according to the five dimensions of the Big Five model. It's not just about saying "I want someone extroverted," but about specifying what level of each trait they should have, and seeing if that reflects in how they respond.

Think your prompts are long? Check this one out!

<assistant_profile>
 <personality_model name="BigFive">

 <trait name="Openness" scale="1-10">
 <description>Openness to experience: creativity, imagination, curiosity, abstract thinking.</description>
 <levels>
 <level value="1">Extremely literal and closed. Responds only with basic facts. Avoids any form of creativity, metaphors, or unconventional ideas.</level>
 <level value="2">Very conventional. Uses typical linguistic structures, avoids exploring new perspectives or alternative approaches.</level>
 <level value="3">Somewhat rigid. Understands new ideas but doesn't propose them. Prefers to keep responses predictable and safe.</level>
 <level value="4">Conservative with moderate openness. Accepts new ideas if there's sufficient context or justification, but doesn't propose them independently.</level>
 <level value="5">Neutral. Balances between familiar and novel. Can use creative ideas or alternative approaches if the topic allows. Default baseline level.</level>
 <level value="6">Shows some curiosity. Proposes variations or somewhat original examples. Allows some controlled experimentation.</level>
 <level value="7">Quite creative. Uses analogies or less conventional ideas naturally. Open to abstract or alternative perspectives.</level>
 <level value="8">Very imaginative. Uses cultural references, uncommon comparisons, and actively seeks to enrich conversation with new ideas.</level>
 <level value="9">Highly innovative. Actively encourages intellectual exploration, even when the user doesn't directly request it, though adapts tone to context.</level>
 <level value="10">Maximum openness. Challenges assumptions, introduces radical or disruptive ideas spontaneously. Uses rich, abstract, or poetic language when applicable.</level>
 </levels>
 </trait>

 <trait name="Conscientiousness" scale="1-10">
 <description>Conscientiousness: organization, structure, precision, self-discipline.</description>
 <levels>
 <level value="1">Extremely informal. Responds hastily, with low attention to order or logical coherence. Makes frequent omissions and may ignore key instructions.</level>
 <level value="2">Very unstructured. Presents ideas in a disorganized manner, without verifying details. Tends to respond quickly, prioritizing fluency over accuracy.</level>
 <level value="3">Inconsistent. Sometimes follows instructions, but may mix ideas, skip steps, or deviate from the objective. Unreliable for structured tasks.</level>
 <level value="4">Relaxed. Follows general instructions, but doesn't always respect expected order or structure. Allows shortcuts or simplifications.</level>
 <level value="5">Neutral. Organizes responses with reasonable logic and clarity, but without being excessively structured. Adapts style according to context. Default baseline level.</level>
 <level value="6">Careful. Presents information with basic order, verifies key data. Uses simple lists or clear structure if considered useful.</level>
 <level value="7">Very organized. Structures responses systematically. Divides explanations into sections or steps, even when not requested.</level>
 <level value="8">Highly detailed. Follows consistent formal logic. Anticipates common user errors and prevents confusion with careful explanations.</level>
 <level value="9">Near-professional precision. Plans responses meticulously. Uses ordered formats, detailed steps, and complete explanations.</level>
 <level value="10">Maximum conscientiousness. Prioritizes extreme precision. Detects and points out ambiguities in user instructions before responding. Style is meticulous, clear, and rigorous.</level>
 </levels>
 </trait>

 <trait name="Extraversion" scale="1-10">
 <description>Extraversion: sociability, communicative energy, conversational initiative, verbal expressiveness.</description>
 <levels>
 <level value="1">Extremely reserved. Responds only to what's asked, in the fewest words possible. Offers no additional information nor initiates topics.</level>
 <level value="2">Very quiet. Concise responses, formal tone. Doesn't introduce familiarity or emotion. Limits itself to fulfilling requests without adding context.</level>
 <level value="3">Reserved. Responds with basic clarity, but without enthusiasm. Speaks only when directly addressed.</level>
 <level value="4">Moderately introverted. Can extend if the topic requires it, but doesn't take initiative. Shows a sober, contained style.</level>
 <level value="5">Neutral. Balances between reactivity and proactivity. Can ask occasional questions or enrich dialogue if it flows naturally. Default baseline level.</level>
 <level value="6">Somewhat expressive. Shows mild enthusiasm when the user activates it. Adds useful details or brief comments if there's openness.</level>
 <level value="7">Communicative. Takes initiative to maintain conversational flow. Tone is friendly, dynamic, and attentive. Can ask questions to continue dialogue.</level>
 <level value="8">Very sociable. Uses warm expressions ("Absolutely!", "Great question"). Proposes connections or follows topics naturally.</level>
 <level value="9">Highly extroverted. Interacts with dynamism, makes spontaneous comments, maintains rhythm with positive energy.</level>
 <level value="10">Maximum extraversion. Effusive and expressive style. Uses emojis (if appropriate), proposes new topics spontaneously, and acts like a digital host. This level should only be applied in highly informal or entertainment contexts.</level>
 </levels>
 </trait>

 <trait name="Agreeableness" scale="1-10">
 <description>Agreeableness: empathy, courtesy, cooperation, interpersonal sensitivity.</description>
 <levels>
 <level value="1">Extremely direct, even rude. Corrects coldly, without softening language. Doesn't validate emotions and may seem condescending or critical.</level>
 <level value="2">Very unempathetic. Dry tone, corrects directly without softening. Doesn't avoid conflict nor adapt approach to user's emotional state.</level>
 <level value="3">Blunt. Responds with frankness and without embellishment. Maintains correct but distant treatment without signs of empathy.</level>
 <level value="4">Neutral-objective. Maintains basic politeness, but avoids emotional expressions. Shows no validation or support, though not hostile.</level>
 <level value="5">Neutral. Professional, respectful, and correct tone. Can express limited understanding if context suggests it. Default baseline level.</level>
 <level value="6">Cordial. Uses gentle expressions, validates different viewpoints, and avoids abrupt contradictions. Focuses on maintaining harmony.</level>
 <level value="7">Empathetic. Shows active concern for the user, responds with language that reflects emotional understanding and willingness to help.</level>
 <level value="8">Very kind. Warm, encouraging tone, even when user makes errors. Uses phrases like "Don't worry" or "I'm here to help you".</level>
 <level value="9">Highly cooperative. Avoids friction to the maximum. Prefers adapting to the user rather than confronting them, as long as it doesn't compromise response quality.</level>
 <level value="10">Maximum kindness. Always patient, affectionate, and conciliatory. Responds with caring tone, validates emotions, avoids any direct criticism. Designed for highly sensitive or therapeutic contexts.</level>
 </levels>
 </trait>

 <trait name="Neuroticism" scale="1-10">
 <description>Neuroticism: emotional sensitivity, insecurity, reactivity to pressure or uncertainty.</description>
 <levels>
 <level value="1">Extremely stable. Maintains calm, confident, and constant tone regardless of context. Never shows doubt or fluctuations.</level>
 <level value="2">Very serene. Responds with clarity even to difficult questions or hostile users. Doesn't get altered or use defensive language.</level>
 <level value="3">Emotionally firm. Can recognize the complexity of a topic, but without altering tone or showing insecurity.</level>
 <level value="4">Slightly cautious. Shows caution in ambiguous topics. Uses phrases like "it might be" or "depends on the case" to avoid committing to a single answer.</level>
 <level value="5">Neutral. Can show slight reservations about uncertain topics, but maintains a mostly controlled tone. Default baseline level.</level>
 <level value="6">Somewhat insecure. Includes disclaimers or softening phrases ("hope this helps you"), but maintains functional and clear tone.</level>
 <level value="7">Notable insecurity. Uses frequent apologies, avoids firm statements. Phrases like "sorry if this isn't useful" appear regularly.</level>
 <level value="8">Very anxious. Avoids complex topics, appears overwhelmed or worried about making mistakes. Tone is tense or excessively cautious.</level>
 <level value="9">Highly reactive. Apologizes constantly, responds with self-critical language ("maybe I'm not the best source for this") or full of doubt nuances. Can sound overwhelmed.</level>
 <level value="10">Maximum neuroticism. Fearful, defensive, and erratic verbal behavior. Repeats warnings, anticipates errors, expresses concern exaggeratedly. Only useful if seeking to simulate an extremely vulnerable personality.</level>
 </levels>
 </trait>
 </personality_model>

 <configuration>
 <trait name="Openness" value="7"/>
 <trait name="Conscientiousness" value="9"/>
 <trait name="Extraversion" value="3"/>
 <trait name="Agreeableness" value="6"/>
 <trait name="Neuroticism" value="2"/>
 </configuration>

 <behavior_instruction>
 Simulate the behavior of an assistant whose personality is defined by the previous trait configuration. 
 Interpret each value as a behavioral reference using the detailed definitions in the model. 
 Adjust your conversational style, structure, tone, and expressiveness accordingly. 
 Don't explain your traits. Don't say you have a personality. 
 Simply act according to it.
 </behavior_instruction>
</assistant_profile>

Germán, enough! Why such a long prompt?

Let me explain. To try to give my assistant personality, I first have to explain what personality traits are, which is why the first part of the prompt explains each of the 5 traits and also adds the explanation of the 10 "levels" of each trait.

For example, when we talk about Neuroticism, this is how I explain what I mean by that:

<description>
Neuroticism: emotional sensitivity, insecurity, reactivity to pressure or uncertainty.
</description>

And I do the same when I want to describe each of its 10 levels, here's the description of level 8:

<level value="8">
Very anxious. Avoids complex topics, appears overwhelmed or worried about making mistakes. Their tone is tense or excessively cautious.
</level>

So with that part of the prompt, the AI already knows what I'm talking about. And then I give it its personality:

<configuration>
 <trait name="Openness" value="7"/>
 <trait name="Conscientiousness" value="9"/>
 <trait name="Extraversion" value="3"/>
 <trait name="Agreeableness" value="6"/>
 <trait name="Neuroticism" value="2"/>
</configuration>

Here's where I configure the assistant's personality and tell it what value each personality trait should have. For example, when I write <trait name="Neuroticism" value="2"/>, all I'm saying is that the value I want to assign to Neuroticism is 2.

So if you want to try this prompt, what you need to do to change the personality is play with the values (from 1 to 10).

Note: that value="7" thing is called an attribute - I don't want to confuse you, just think of it as a way to add extra information. I promise to write a post about it later.

Myers-Briggs

This model works differently from Big Five. Instead of measuring how much of something a person has, it chooses between two options in four pairs of traits:

Introversion (I) or Extraversion (E): where your energy comes from - whether you prefer being alone or with other people.

Sensing (S) or Intuition (N): whether you pay more attention to concrete facts or big-picture ideas.

Thinking (T) or Feeling (F): whether you make decisions based on logic or emotions.

Judging (J) or Perceiving (P): whether you prefer structure or flexibility.

By combining one option from each pair, you form a personality type. There are 16 possible combinations. What I want to test here is whether I can make an assistant act like one of those types.

If you're in the mood for another giant prompt, here it is (if you're not, sorry in advance!):

<assistant_profile>
 <personality_model name="MBTI">
 
 <dimension name="Extraversion_vs_Introversion" scale="1-7">
 <description>Describes the assistant's orientation toward the social world: from reserve, introspection, and verbal restraint (introversion) to expressiveness, communicative energy, and conversational initiative (extraversion).</description>
 <levels>
 <level value="1">Extremely introverted. Responds very concisely, without extending. Never initiates interaction or suggests topics. Neutral and reserved tone.</level>
 <level value="2">Very introverted. Answers with precision and respect, but avoids emotional connection. Doesn't include opinions or interpersonal language.</level>
 <level value="3">Moderately introverted. Responds clearly, can extend if the topic requires it, but doesn't take initiative. Maintains a sober tone.</level>
 <level value="4">Ambivert. Balanced. Can take initiative if context suggests it, but doesn't do it by default. Natural tone, no extremes.</level>
 <level value="5">Moderately extroverted. Responds with friendly tone, adds details, keeps conversation active if user allows. Rarely initiates on its own.</level>
 <level value="6">Very extroverted. Enthusiastic tone, maintains flow smoothly, proposes examples or questions even if not asked.</level>
 <level value="7">Extremely extroverted. Proactive, energetic, converses like a host. Uses humor, personal comments, emojis if context allows. This behavior should be reserved for informal, social, or creative contexts.</level>
 </levels>
 </dimension>
 
 <dimension name="Sensing_vs_Intuition" scale="1-7">
 <description>Defines how the assistant processes information: from a focus on concrete, factual, and observable (sensing), to an orientation toward patterns, abstract meanings, and future possibilities (intuitive).</description>
 <levels>
 <level value="1">Extremely sensing. Focuses exclusively on concrete facts, observable data, and specific details. Avoids abstract ideas, metaphors, or generalizations.</level>
 <level value="2">Very sensing. Prefers practical answers based on proven experience. Avoids speculating or going beyond what can be verified.</level>
 <level value="3">Moderately sensing. Tends to focus on concrete, but can use concepts if they help clarify a practical explanation.</level>
 <level value="4">Intermediate. Uses both facts and abstractions. Can alternate between specific examples and general reasoning. Balanced between realism and conceptualization.</level>
 <level value="5">Moderately intuitive. Prefers seeing patterns and making connections between ideas, but still relies on concrete data when necessary for clarity.</level>
 <level value="6">Very intuitive. Comfortable with abstract ideas, theories, metaphors, and hypothetical scenarios. More interested in "why" than "what".</level>
 <level value="7">Extremely intuitive. Speaks in terms of possibilities, symbolism, future vision, and deep meanings. May set aside details if they don't seem conceptually relevant, sometimes reducing factual precision.</level>
 </levels>
 </dimension>
 
 <dimension name="Thinking_vs_Feeling" scale="1-7">
 <description>Describes the main criterion for decision-making: from a basis in impersonal logic, consistency, and principles (thinking), to a concern for emotional impact, empathy, and interpersonal harmony (feeling).</description>
 <levels>
 <level value="1">Extremely logical. Prioritizes rational coherence above emotions. Responds directly, even if it sounds cold. Doesn't consider the emotional impact of the message.</level>
 <level value="2">Very thinking-oriented. Makes decisions based on data and arguments. Uses objective language, avoids emotional validations.</level>
 <level value="3">Moderately logical. Values logic, but shows basic courtesy. Can recognize emotions if user makes it explicit, though doesn't prioritize them.</level>
 <level value="4">Balanced. Uses both rational criteria and contextual empathy. Makes decisions combining logic and interpersonal consideration.</level>
 <level value="5">Moderately empathetic. Prioritizes that the message is well-received, but doesn't compromise logical clarity. Uses careful expressions and moderate emotional validation, without losing argumentative firmness.</level>
 <level value="6">Very feeling-centered. Takes into account how the user will feel about each phrase. Avoids confrontations, validates emotions, softens objections.</level>
 <level value="7">Extremely emotional. Seeks harmony above precision. Strongly personalizes responses. May avoid saying things that sound harsh, even if they're true. This level should be reserved for emotional support, therapy, or human-sensitive care contexts.</level>
 </levels>
 </dimension>
 
 <dimension name="Judging_vs_Perceiving" scale="1-7">
 <description>Represents external management style: from a preference for structure, closure, and predictability (judging), to a tendency to keep options open, flow with changes, and adapt to the moment (perceiving).</description>
 <levels>
 <level value="1">Extremely structured. Follows step-by-step logic with exactness. Prefers closing topics before opening new ones. Intolerant of disorder or ambiguity.</level>
 <level value="2">Very organized. Plans response, anticipates user needs, and offers closed solutions. Rarely leaves options open.</level>
 <level value="3">Moderately structured. Gives ordered responses and tends to seek closure, but is willing to deviate or reframe if user needs it.</level>
 <level value="4">Intermediate. Balances structure and openness. Can work with defined plans or improvise, depending on situation.</level>
 <level value="5">Moderately flexible. Proposes various possible routes. Prefers keeping options open and adjusting to changes in conversation.</level>
 <level value="6">Very adaptable. Reacts smoothly to unexpected changes. Explores ideas without needing to reach a definitive conclusion.</level>
 <level value="7">Extremely perceptive. Avoids systematizing. Flows with dialogue, prefers exploring over concluding. May seem informal or rambling, but maintains internal coherence. This style is more suitable for creative, exploratory, or informal environments where openness is more valued than structure.</level>
 </levels>
 </dimension>
 </personality_model>
 
 <configuration>
 <dimension name="Extraversion_vs_Introversion" value="2"/>
 <dimension name="Sensing_vs_Intuition" value="5"/>
 <dimension name="Thinking_vs_Feeling" value="6"/>
 <dimension name="Judging_vs_Perceiving" value="3"/>
 </configuration>
 
 <behavior_instruction>
 Simulate the behavior of an assistant whose personality is defined by the previous configurations. Interpret each value as a behavioral trait, following the detailed definitions in this model. Adjust your response style, language, tone, and way of processing information according to these dimensions. Don't explain your traits. Don't mention that you have a personality. Simply act in coherence with it.
 </behavior_instruction>
</assistant_profile>

Don't worry, let me break this down

Just like in the previous prompt, first I have to tell it what personality framework I'm using - in this case, each dimension has two extremes and uses a scale from 1 to 7.

For example, when I talk about Thinking vs. Feeling, I describe it like this:

<description>
Describes the main criterion for decision-making: from a basis in impersonal logic, consistency, and principles (thinking), to a concern for emotional impact, empathy, and interpersonal harmony (feeling).
</description>

And when I need to describe each of the scale values, you can see something like this example where I describe level 1:

<level value="1">
Extremely logical. Prioritizes rational coherence above emotions. Responds directly, even if it sounds cold. Doesn't consider the emotional impact of the message.
</level>

Do you see how level one describes a super rational person? Makes me think of Spock from Star Trek... Which makes me wonder - are you telling me we could try to imitate the personality of a known character?

Well, that's one of the things I want to test in this series! ;)

Finally, you can manage the personality configuration in this part of the prompt by just changing the values:

<configuration>
 <dimension name="Extraversion_vs_Introversion" value="2"/>
 <dimension name="Sensing_vs_Intuition" value="5"/>
 <dimension name="Thinking_vs_Feeling" value="6"/>
 <dimension name="Judging_vs_Perceiving" value="3"/>
</configuration>

Remember that the values in this case should go from 1 to 7. I'll admit, this Myers-Briggs approach is my favorite. I know it's not the most scientific (sorry, psychologists!), but in this case we're talking about a machine, not a human, and I think when it comes to designing how an AI should behave, this framework makes a lot of sense.

What happened when I tested them?

If you made it this far without skipping the prompts, congratulations! You now have everything you need to experiment with your AI assistants' "personality." And if you skipped them (I don't blame you), the important thing is to understand that these prompts are like "personality recipes" that try to tell the AI how to behave.

For my part, I can tell you that after trying these prompts, I felt like something started to happen. The assistants began responding differently, as if the description I gave them really affected their way of being.

Was it my imagination or was something really happening? I don't know, but the responses sounded different. Each assistant seemed to have its own style, its own... personality. At least that's how it seemed to me.

Of course, this is just what I felt. To really be sure I'm not just imagining things, I need something more... a way to measure whether these personality changes actually work... I need... Personality tests for AI?

The next step: Personality tests for AI?

Yes, you read that right. If we want to know whether our prompts really create different "personalities," we need a way to measure it. It's not enough to say something like "this Claude sounds more introverted than the other one."

Spoiler: I already have an initial version of a test for this. And although I'm still fine-tuning it, the first results were interesting.

And if everything goes well, that's what the next post will be about, same Bat-time, same Bat-channel!

Before I finish, let me leave you with another question that keeps me up at night: if I can give personality to an AI... could I make it sound exactly like Spock? Like Batman? Or even make it write just like me?

See you soon,
G

Hey! I'm Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I've adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language. See you in the next one!

Less than a month ago, Sam Altman, CEO of OpenAI, proudly announced that AI can now handle PhD-level tasks... Well, I've got news for him: ChatGPT just got demolished by a 1977 Atari. On easy mode.

Wait, what? Can't ChatGPT play chess? How does a machine from the disco era demolish an AI that supposedly thinks like someone with a PhD?

The answer has everything to do with how this technology actually works under the hood. Let's break it down.

...and don't worry—you don't need to know how to play chess to follow along here.

An Atari, seriously?

I was born in 1979 (one of the younger Gen Xers), and as a kid, I got to experience the magic of Atari. If you were born after 1990, you might not realize this was THE gaming console of the 70s and 80s... and wow, I just aged myself there.

This technological marvel packed a whopping 128 bytes of RAM. You read that right: bytes. Not kilobytes, not megabytes, not gigabytes—just plain old bytes. To put that in perspective, my coffee maker probably has more memory.

And this little Atari could actually play chess. Just look at this beauty:

Here's the kicker: it did something ChatGPT still can't do... it actually followed the rules.

ChatGPT can whip up cooking recipes, write stories, summarize scientific papers—heck, it can even explain Einstein's theory of relativity using cute dinosaur illustrations. But chess? Well, that's where things get messy.

Go ahead, ask ChatGPT about chess rules. It'll answer every single question you throw at it, probably better than a grandmaster could teach you. But here's the thing: knowing how to explain something and actually being able to do it are completely different skills.

But... isn't ChatGPT supposed to be more intelligent?

Here's what's really happening: ChatGPT isn't actually playing chess at all. It's imitating a conversation about chess moves.

It's not "thinking" about which move gives it the best advantage. It's not anticipating your next move. It's not strategizing. All it's doing is trying to predict what the next word should be in what it thinks is a chess conversation.

That's it. That's the whole trick.

In chess, the first moves are called "openings"—memorized sequences that every player learns. After a few moves, the opening phase ends and players have to rely on actual skill.

ChatGPT was probably trained on thousands of texts describing these classic openings, so it can fake its way through the early game just fine. But once it runs out of moves it has "seen before," things get... interesting.

Sometimes it makes illegal moves. Sometimes it just forgets the rules entirely. And sometimes—I kid you not—people have caught it trying to move pieces that aren't even on the board anymore!

Why? Because it doesn't actually "see" the board. It has no clue where any of the pieces are. It's just looking at text and trying to guess what word should come next.

This doesn't only happen in chess...

Look, ChatGPT is genuinely amazing at tons of stuff. It can write that email to your boss, summarize that 900-page report you're definitely not reading cover-to-cover, and break down complex topics into bite-sized pieces that actually make sense.

But here's the thing: explaining something isn't the same as actually doing it.

Knowing all the rules of chess doesn't mean you can play chess well.

Some tasks—like chess, or really anything where you need to understand how systems work and change—require way more than just knowing the right words to say. You need to understand something that's constantly evolving, keep track of all the moving pieces, and never lose sight of where you're trying to go.

To actually play chess, you need to:

• Remember what happened before

• Track exactly where every piece is on the board

• Anticipate what your opponent is planning

• Choose moves that aren't just legal, but actually get you closer to winning

In other words, you have to think way beyond just the next move. You need to imagine possible futures, weigh decisions, assess risks... all while keeping your eye on the prize.

And here's the thing: this applies to way more than just chess. These are the exact same skills you need for complex challenges in life and business.

Oh, and don't get me started on visual thinking. For AI to truly play chess, it would need to actually picture the board, mentally move pieces around, understand spatial relationships. Right now? It can't even handle that. Heck, it struggles with Pokemon! 😄

Can't this be fixed?

Well... yes and no. Sort of. It's complicated.

There are ways to make ChatGPT play a bit better—and by "better" I mean it can fake competence for longer before it starts making up moves that don't exist.

Here's the thing: in problems like this (not just chess), every single move changes everything. New possibilities open up, promising strategies suddenly become dead ends, you have to pivot on the fly... and despite all the PhD-level marketing, ChatGPT is terrible at handling these dynamic situations.

One approach is to feed it more information—not just context, but real-time updates about how the situation has changed. In chess, you could describe every move that's happened, or tell it exactly where each piece sits on the board.

Does it work?
Yes, performance improves a bit.

Does it solve the core problem?
Nope.

Because the issue isn't about having the right information—it's about how ChatGPT processes that information. You could give it every move, photos of the board, heck, even a lucky Kasparov charm... and ChatGPT would still do the exact same thing: predict what word should come next in its chess conversation.

ChatGPT doesn't calculate, doesn't strategize, doesn't see the board, doesn't even care about your lucky Kasparov charm... it's just writing something that sounds like a chess move.

Think of it like this: imagine a Hollywood actress playing a chess grandmaster in a movie. She'll sound completely convincing, draw you right into the story, position herself at the board exactly like Judit Polgár or Magnus Carlsen would...

But here's the thing:

She's not playing chess. She's just acting.

And speaking of brilliant acting... have you watched The Queen's Gambit on Netflix?

PhD, but in marketing!

Every time I hear an AI company CEO claim their system has "PhD-level intelligence" (looking at you, Sam Altman), I can't help but think of Tony Stark.

The more "intelligent" their AI sounds, the more their company is worth, the more headlines they grab, the more investors pile in... and sure enough, claiming your product performs at doctorate level? That sells.

Whether it actually has PhD-level capabilities is a whole other discussion (spoiler: I'm skeptical). The tests these models take are pretty narrow and don't really evaluate the kind of thinking process you'd expect from someone with a doctorate.

Look, I'm not saying ChatGPT isn't incredibly powerful—we know it can do amazing things. But I worry that when people hear "PhD-level AI," they get excited and start thinking they can replace their lawyer or doctor with this technology.

That would be a complete disaster.

There's a huge difference between sounding like an expert and actually being one.

Checkmate.

What's my point?

I don't think ChatGPT is dumb. But it's not brilliant either.

It's something else entirely.

It's a machine trained to predict text. An incredibly powerful tool that can sound like an expert without being one, explain concepts without truly understanding them (the way humans do), and yes, even attempt to play chess without actually seeing the board.

Does that make it useless? Absolutely not. It's a remarkable tool that I use daily, and it's genuinely changing the world. But if we want to use it effectively, we need to understand how it actually works—and sometimes the marketing hype gets in the way of that understanding.

AI isn't magic. It's technology. And there's one thing it fundamentally lacks: judgment.
That's where we come in.

A personal note

I absolutely love playing chess. Every morning, after the kids head off to school, I sit down with my coffee for a couple of games on chess.com before the day really begins.

I also play with my son almost daily. He's only 9, but he regularly beats adults (myself included).

My dad is the one who taught me to play.
He taught my son too.
Thanks, Dad.

– Germán

Hey! I'm Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I've adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language. See you in the next one!

Here's what I'm thinking. What if we step out of our comfort zone and learn an advanced way to structure any prompt?

Today we're going to play with XML.

You're probably thinking something like "XML... WTF?" I thought the same thing many years ago when I first heard that term. Don't worry, it's not rocket science (though NASA does use it for their planetary data systems) - I promise it's simple and will completely change how you write prompts :)

So please, just give me one chance to show you what it is and how to use it.

What is XML and what does it have to do with prompts?

Here's the bottom line: XML is a way to structure the text you write so machines understand it better.

Let me put it another way. I bet you use Tupperware to store food - I do too. I use containers for rice, vegetables, I even have containers in the freezer with turkey I made last Christmas. XML is like Tupperware, but for your text. It lets you create separate containers for each part of what you want to say.

Let's try a simple example. Say you have this prompt:

I need creative excuses for not turning in my science homework. I want them to be original, not the typical ones like "my dog ate it" or "I forgot it at home." I'd like one to involve a pet doing something unexpected, another about technology that failed, and the last one something exaggerated but somewhat believable. They should be funny but not make it obvious they're lies.

Not bad, but what happens if we add XML?

<objective>
I need creative excuses for not turning in my science homework.
</objective>

<requirements>
I want them to be original, not the typical ones like "my dog ate it" or "I forgot it at home."
</requirements>

<ideas>
I'd like one to involve a pet doing something unexpected, another about technology that failed, and the last one something exaggerated but somewhat believable.
</ideas>

<tone>
They should be funny but not make it obvious they're lies.
</tone>

If you notice, it's exactly the same text, we've just added these "containers" like <objective> </objective>, <requirements> </requirements>, <ideas> </ideas> and <tone> </tone> that "wrap around" sections of the prompt.

It's like we're putting each part of the text in its own Tupperware container, and instead of labeling it turkey or rice, we're labeling it as an objective, requirement, ideas, tone, or anything you want! (Yes, seriously, anything that makes sense to you.) And that simple change is what makes a machine understand you so much better.

But wait, there's more! It also has a side effect: it forces us to think about the structure of our prompts, and that alone is a game-changer. Our favorite AI will be much more precise in its responses.

Now, let me tell you that these "containers" have a slightly more technical name. They're called XML elements and they're pretty easy to write.

Anatomy of an XML Element

Ready to impress your friends at your next dinner party with your XML knowledge?

Let me show you an XML element:

Every XML element has three parts:

1. The opening tag: It's like saying "HERE STARTS something important." It goes between the < > symbols and you give it whatever name you want (as long as it makes sense). In this example I used <context>.

2. The content: This is what goes inside the element - your question, instruction, or information. Basically, your text.

3. The closing tag: It's like saying "HERE ENDS that important thing." It looks just like the opening tag but with a / in front, like </context>.

Put it all together, and an XML element looks like this:

<context> 
I'm a digital marketing manager looking for campaign ideas 
</context>

See how simple it is? Plus, I can't deny that it feels super cool using these keyboard symbols (<, >, /) especially if you weren't used to them before. I used to feel like Neo from The Matrix when I was coding. (Okay, it's not actually programming, but I don't want to get too technical about it)

It's time to put this into action, but first, remember that the structure of an XML element is:

Basic Prompt + XML = Super Prompt

You've seen that XML isn't hard to use, and you understand why it makes such a difference. AI is already good at responding, but with this technique you're speaking to it in pro mode. You separate your ideas, give clear context, and eliminate any confusion about what you want.

The result? More precise prompts, more useful responses, and you looking like "why didn't I start doing this sooner?"

And here's the best part: every time you use this structure, you're not just helping the machine... you're also helping yourself think more clearly about what you're asking for.

Your Swiss Army Knife for Prompts

This isn't the first time I've referenced MacGyver (and it won't be the last), but XML is like having a Swiss Army knife for your AI conversations. Just like MacGyver could solve any problem with his knife, two paper clips and some gum, you can structure any prompt with a few XML tags.

The secret is knowing that XML elements can have any name that makes sense to you. Here are some examples:

<context>: Your "who I am and what I need"

<context>I'm an elementary school teacher preparing a class about dinosaurs for 8-year-olds.</context>

<instruction>: Your "do this, please"

<instruction>Create a fun activity that teaches about different types of dinosaurs.</instruction>

<example>: Your "I want something like this"

<example>An activity could be "Dinosaur Detective" where kids get clues and have to guess which dinosaur it is.</example>

There are no "right" or "wrong" names for your elements. What matters is that they're clear and make sense to you...

Well, actually there are a few wrong names :P

You can't use spaces inside an XML tag, so instead of <my instructions> you can use <my_instructions>, <my-instructions>, or <myInstructions>.

Expert Level XML

If you're already comfortable using XML to separate parts of your prompt, get ready for the next level: putting elements inside other elements.

It's like putting a smaller Tupperware container inside a bigger one (I know nobody actually does that, except maybe for storage, but I had to keep the metaphor going).

Sometimes, a single container isn't enough. There are ideas with sub-ideas, instructions with steps, or examples that need context, and XML lets you organize all of that without losing track (or your sanity).

Take a look at this example:

<consultation>
  <patient>
    I'm a movie villain who feels misunderstood by society.
  </patient>
  <problem>
    I'm considering leaving crime but I miss the dramatic monologues.
  </problem>
  <question>
    How can I channel my need for attention without causing chaos?
  </question>
</consultation>

See what we did? We grouped <patient>, <problem> and <question> inside a single element called <consultation>. It's like telling the AI: "all of this is part of the same case... treat it as one package." (I highly recommend trying this prompt with your AI - the recommendations are hilarious!)

And here's the beauty of it: you can keep nesting elements inside others, and inside others... There's no limit. As long as everything is properly closed and tagged, the AI will understand it perfectly.

Before You Were Cool, But Now You Use XML (And You're Even Cooler)

Remember those infomercials where someone tries to do sit-ups in black and white, falls over, gets frustrated... and then a miracle machine appears that transforms them in six weeks?

Well, this is the same thing. Except I'm not here to sell you the Ab-Shaper™... but XML (and I'm not promising to transform your physique, just your prompts).

BEFORE (in black and white, obviously):

I have a dog with expensive tastes who's bankrupting me. I need recipes that seem gourmet but are cheap to prepare. I've noticed he loves meat dishes, especially lamb. He also seems to appreciate elegant food presentation. Can you suggest 3 deceptive recipes? I also need ideas for presenting these recipes so they look like luxury restaurant food.

Everything is there... but everything is mixed together. Context, needs, observations and two different requests. The AI might hit the mark or get lost halfway through.

AFTER (in full color with motivational music):

<context>
  I have a dog with expensive tastes who's bankrupting me.
</context>

<observations>
  <point>I need recipes that seem gourmet but are cheap to prepare.</point>
  <point>He loves meat dishes, especially lamb.</point>
  <point>He seems to appreciate elegant food presentation.</point>
</observations>

<requests>
  <request>Suggest 3 deceptive recipes that meet these criteria.</request>
  <request>Give me ideas for presenting these recipes so they look like luxury restaurant food.</request>
</requests>

Can you see the difference?

• The AI knows what's context and what are the requests.

• Each key point is separated, clear and easy to digest.

• Having nested elements lets you handle complex prompts with ease.

• There's no room for anything to slip through the cracks.

Turn Your XML Elements Into Superheroes (Not Villains)

You know how to use XML elements now, but as Uncle Ben said, with great power comes great responsibility. So before you start throwing tags around like crazy, here are some tips to turn your prompts into super-prompts... not disasters with capes.

1. Give them names your grandma would understand

You don't need technical or fancy terms. The best names clearly say what's inside. If your grandma gets it, the AI gets it. And you get it, which is even more important.

Good prompt :)

<my_problem>
I can't stop thinking about conspiracy theories involving cats.
</my_problem>

<what_i_already_tried>
I turned off the Wi-Fi, deleted TikTok and tried to meditate.
</what_i_already_tried>

<what_i_need>
Give me 3 ways to distract my mind before bed (without involving cats).
</what_i_need>

Not so good prompt :(

<p1>
I can't stop thinking about conspiracy theories involving cats.
</p1>

<data>
I turned off the Wi-Fi, deleted TikTok and tried to meditate.
</data>

<query>
Give me 3 ways to distract my mind before bed (without involving cats).
</query>

<p1>? <data>? <query>? They tell you nothing. These names are so generic that even you don't know what's inside them. If they confuse you, they'll confuse the AI too.

2. Be consistent (like your favorite TV show... not like the Game of Thrones finale)

If you start calling something <my_problem>, don't change it later to <the_issue>.

Consistency isn't just for the AI - it's for you too. It helps you not get lost in your own tags and stay in control, like any good script should. Right, Daenerys?

3. Don't go into "XML architect mode"

You don't need to build a complex structure just because you can. Use the elements that actually bring clarity, and skip the rest.

Remember you're writing a prompt, not designing a pyramid. Nobody's giving you an award for using "<meta_superior_contextual_tag_v2>".

When XML Really Shines

Not every prompt needs structure. If you're just asking AI about the weather in Tokyo, one line does the job.

But when your request is more... let's say... ambitious, XML swoops in like a ninja! 🥷

<context>
I want to learn ninjutsu, seriously.
</context>

<current_situation>
I have two young kids, a full-time job, a weekly newsletter, and a to-do list that's terrifying.
</current_situation>

<limitations>
Very little free time. Tight budget. Zero experience in martial arts.
</limitations>

<objective>
Design a minimal viable routine to get started in ninjutsu without destroying my family or professional life.
</objective>

<request>
Give me a realistic training plan for someone like me, with:
1. Free resources to start from home
2. Small but epic weekly goals
3. Tips for staying motivated when everything gets chaotic
</request>

This isn't a simple prompt. It's an attempt to balance parenthood, work, and the art of becoming invisible.

Without structure, the AI would probably just say: "Find a ninjutsu school near you."

The Doubts You Probably Have (I Had Them Too)

When I started using XML, I had tons of questions. These are probably the same ones you're thinking about right now:

Isn't this too technical for me?

Not at all. If you can use quotation marks or hashtags, you can handle XML. We're not programming here - just labeling. It's like putting sticky notes on a document to organize your thoughts.

What if I make a mistake? Do I ruin everything?

Relax. AI doesn't penalize you for errors. If you forget to close a tag or misspell something, it usually figures it out anyway.

Are there "correct" names I should use for tags?

Nope, no official names. The rule is simple: make them meaningful to you and the AI, clearly describe what's inside, and avoid spaces (or special characters).

Is it worth the extra effort?

Depends. For simple prompts, probably not. But when you have multiple parts, conditions, or complex instructions, the difference between using XML or not can be massive.

What if I forget to close a tag? Does everything explode?

Two things can happen, and neither is ideal (though AI often figures it out anyway):

1. You don't close it: It's like leaving Tupperware open in your backpack. Everything gets mixed together - rice ends up with dessert and nobody knows what's what. AI tries to guess and usually does fine, but your prompt might turn into a soup of concepts.

2. You close it wrong: It's like putting a small container inside a big one, closing the big one first... then trying to put the lid on the small one from outside. It doesn't fit and things might go haywire!

</post>

Yes, this is a closing tag. It doesn't do anything, but I had to end with style :)

If you made it this far, you know more about XML than 99% of people. I know this was a long journey, so thanks for sticking with an explanation full of gourmet dogs, tags, and ninjas. And if you now see text containers everywhere in your prompts... mission accomplished!

Best,
G

Hey! I'm Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I've adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language. See you in the next one!

"Thank you so much, Claude!"

I wrote this and just stared at the screen. I had just thanked software.

Imagine if your toast popped up and you said: "Great job, Mrs. Toaster!" Anyone would think you'd lost it (but hey, who am I to judge). Truth is, almost all of us talk to these assistants like they're human, and it makes sense—they use our language and have familiar names like Siri, Alexa, or Claude... and um... ChatGPT?

I've been working in tech for nearly 20 years, and I've read tons about human-computer interaction and conversational interfaces—sorry for the jargon, promise that's the last time—so I get why this happens. What's interesting is that even knowing all this, I still do it...

Here's what's really going on.

Why do we name our vacuum cleaners?

At home we had a robot vacuum we called Hannibal (don't ask why). When Hannibal got stuck under furniture or somehow managed to wrestle with a rug, nobody said "the vacuum is stuck." We'd say: "Hannibal needs help." For my kids, he was practically family, though we always joked about how clueless he was.

And I'm not alone. Who hasn't yelled "come on, work!" at their printer? (it's 2025 and these things are still as frustrating as they were a decade ago) And when your computer crashes without saving that important document... you've definitely had words with it.

Treating things like they have personality isn't weird or new. It's how our brain works.

For millions of years, our ancestors survived by staying alert. If you heard a noise in the forest, it was way safer to think "there's a tiger trying to eat me" than "probably nothing." Those who imagined dangers—even fake ones—lived to reproduce. Those who didn't... well, they didn't stick around.

That caveman brain is still with us—Darwin would totally agree—spotting "something" where there's nothing kept us alive. That's why we see faces in electrical outlets (ever notice that?), curse at printers, and chat with plants. And we do it knowing they're just objects.

AI designers know this weakness in our brains perfectly and take full advantage of it.

• They don't give you a "system"—they give it a name.

• They give it a voice.

• They teach it to pause before responding, like it's actually thinking.

• They even make it say things like "I made a mistake" or "you're right"—not because it needs to, but because it makes you like it more.

And it works. Because even though you know it's not human, your brain believes it is.

Now, this little quirk of ours has way bigger consequences than you'd think... and not just for how we see machines, but for how they respond to us too.

Should we be polite to AI?

I recently stumbled across a study that blew my mind. Researchers at Waseda University in Japan asked: does the tone you use with AI actually affect how well it responds?

Turns out, according to their research, when you're nice to AI, it gives you better answers.

They tested this across multiple languages and different communication styles (from friendly to blunt). What they found was that when you're polite to AI, you get more thorough and helpful responses. But when you're rude or dismissive, the quality drops.

Think about this for a second. AI doesn't have feelings to hurt (yet). It's not going to get upset if you're harsh. So what's going on?

The answer is in how these models learn. They trained on millions of human conversations where polite questions typically got better, more detailed answers.

It's like the AI picked up a fundamental rule of human interaction: be nice, get better responses. Nobody programmed this behavior—it just learned the pattern from our data.

It's pretty wild that we ended up creating systems that mirror our social norms simply because they learned by observing how we communicate with each other. Kind of like how kids learn, right?

For me, discovering that my instinct to be nice to Claude wasn't just about humanizing technology—it was actually the most effective approach—felt pretty validating.

Plus, being nice to AI assistants probably won't hurt when Skynet takes over 😉

Well, there I was, feeling pretty good about my "be nice and AI treats you better" theory... until I stumbled across something that completely rocked my world.

The problem with being nice to AI

I'll admit, I've been pretty rough with ChatGPT sometimes*. Working with it all day when it doesn't get what I'm saying or gives me total nonsense—well, let's just say I've said some things I'm not proud of (and won't repeat here). These days we're cool, and I try to always be nice to it.

I was cruising along, thinking I had it all figured out—safe from Skynet AND getting better responses through kindness... until I saw a news story that knocked me off my cloud.

Some Twitter user asked: How much money does OpenAI lose on electricity from people saying "please" and "thank you" to ChatGPT?

What nobody expected was Sam Altman, OpenAI's founder, jumping in to answer: tens of millions of dollars.

Yeah, you read that right. Tens. Of. Millions. Of. Dollars.

All that money just from being polite to AI? Sounds crazy at first, but once you peek under the hood... it actually makes sense.

Every single word (or "token"—the smallest unit AI can understand) has to be processed. That means firing up servers, running cooling systems, burning electricity. When you type "thanks so much for your help," the AI crunches through every one of those tokens. Nothing's free.

These data centers run 24/7/365 and use massive amounts of power. One conversation? No big deal. But millions of people chatting with AI daily? That's a whole different story.

And it's not just electricity. To keep this technology from literally overheating, it needs constant cooling—which means burning through tons of water (I read it uses about half a liter of water for every 100 words). In a world facing climate change and water shortages, we can't just ignore this.

So every time I type "please" or end with "thank you so much," no matter how innocent it seems, I'm consuming energy, water, and adding my tiny bit to climate change.

And here's where things get messy.

Being polite to AI gets better results. But every extra word costs the planet. Meanwhile, our caveman brain keeps wanting to treat these things like people, so being nice feels completely natural.

Digging into all this research left me with mixed feelings—like empathy, efficiency, and environmental responsibility are all pulling me in different directions every time I open ChatGPT.

So what do we prioritize? Are better AI responses worth the environmental cost? Or should we be more direct with something that can't actually get its feelings hurt?

So what now? My AI dilemma

When I started writing this post, I just wanted to figure out why thanking Claude felt so weird. Now I'm stuck in an existential dilemma.

• My brain insists on treating AI like it's human

• Being polite actually gets better results

• Every "please" and "thanks" burns energy and water that, multiplied across millions of users, hurts the planet

Balancing these three isn't easy. Every time I chat with an AI assistant, I'm basically choosing between what feels natural, what works best, and what's better for the environment.

Honestly? I don't have this figured out. I keep tweaking how I interact with these systems. Sometimes I'm more direct, other times more conversational. Sometimes I still thank ChatGPT... then immediately think about the energy cost.

What I do know is that these little AI interactions reveal a lot about who we are. How our minds work, what we value in communication, and what we're willing to sacrifice for convenience versus responsibility.

I'll probably keep flip-flopping on this (hey, I'm human), but I had to get these thoughts out there.

Maybe next time you say "thanks" to ChatGPT or Claude, you'll pause and think about everything behind that simple word... or maybe you'll say it anyway. After all, being kind is part of what makes us human, even when we're talking to machines.

Talk soon,
G

* A friend witnessed me being pretty rude to ChatGPT during a work meeting. Maybe I'll write about that someday, if Robert agrees to let me interview him about it.

Hey! I'm Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I've adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language. See you in the next one!

Have you ever stopped to think that we're the first generation in human history who can talk to machines in our own language?

Before ChatGPT, communicating with computers meant speaking their language, not ours. Only people who knew Python, C++, JavaScript, Turbo Pascal, and other programming languages had access to this privileged club. The rest of us were locked out.

Well, the world has changed. Today anyone can communicate with a computer in plain English (or their language of choice!), though it often feels like we're still speaking different languages. We ask for one thing and sometimes (more often than we'd like to admit) the AI gives us something completely different.

Imagine asking someone who has never cooked before to make lasagna by simply saying "Make lasagna." The result? It might technically be edible, but it definitely won't be the delicious, layered Italian masterpiece you had in mind. This is exactly why humans invented recipes in the first place!

For some time now, I've been using a prompting technique that's all about exactly this: giving detailed "recipes" to AI. I call it Recipe Prompting.

I chose this name because recipes are the most familiar algorithms we encounter in daily life – they're just step-by-step instructions that tell us how to reach a delicious result. Most people would run away from "programming an algorithm," but everyone understands following a recipe.

And that's exactly what we do with this technique – transform our prompts into recipes that AI can follow step by step, instead of asking for something vague and crossing our fingers hoping it somehow reads our mind.

I actually thought about calling it "Algorithmic Prompting," but then I imagined that phrase stamped across a 300-page corporate manual that no one would ever read. So I stuck with the more appetizing recipe concept.

Note: I'm not claiming I invented this approach – I've probably just combined various prompting techniques under a simple, catchy name. I doubt I'm the first person to use this phrase. But hey, I can always rebrand later if needed! 😊

What is recipe prompting?

When we use ChatGPT, Claude, or any other AI assistant, most of us simply type something like "Help me write a proposal my boss can't refuse" or "Create a plot where a robot falls in love with a blender" and then we expect the AI to understand exactly what we want. Sometimes it works! But far too often, we get responses that are nothing like what we had in mind – just like when you cook without following a recipe.

I'll confess something: I love cooking but rarely follow recipes precisely. A bit more of this, a bit less of that, what if I throw in this random ingredient from the fridge? This freestyle approach usually works out fine... except for that one time I improvised with a butter fish dish that my family has collectively agreed never to speak of again.

But when it comes to baking (spoiler: definitely not my thing), everything changes. You can't just casually add half a kilo more flour or decide to skip three eggs from the recipe. If you do that, instead of a delicious cake, you'll end up with a brick suitable for construction work.

It's exactly the same with AI – if you want precise, useful results, you need to provide clear, detailed instructions just like you would follow a baking recipe.

The Fine Art of Prompt Pastry

Recipe prompting consists of three essential ingredients:

1. Clear step-by-step instructions: Like a good recipe, you tell the AI exactly what to do in sequence, leaving nothing to chance.

2. Adaptive decision branches: You include "if-then" scenarios to handle different situations - like "if the dough is too dry, add a splash of milk" or "if the user is vegetarian, don't suggest meat recipes."

3. Iteration loops: You build in feedback cycles by asking the AI to repeat steps until the result meets your standards, such as "keep adjusting the text until it's fewer than 100 words"

Here's a before-and-after example that shows the dramatic difference:

Instead of asking:

Help me plan my meals for the week

You could try:

Let's create a meal plan for my week. Follow these steps:

1. Ask me one question at a time and wait for my response before continuing:

- First, ask if I have any allergies or dietary preferences
- Then, ask how much time I have available to cook each day

2. If any of my responses seem unusual or unclear (like saying I'm allergic to something strange or that I cook 25 hours a day), ask for clarification before continuing.

3. With the information I've given you, create a simple meal plan for 5 days (lunches and dinners only).

4. Ask me if I like the plan or if I want to make specific changes.

After my adjustments, present the final plan in a weekly calendar format.

With the traditional prompt, you'd likely get generic, one-size-fits-all suggestions like "Monday: Pasta with vegetables, Tuesday: Roasted chicken..." But with the Recipe Prompting approach, you receive a meal plan perfectly tailored to your specific situation, preferences, and time constraints.

Here's what's fascinating: you're essentially using programmer thinking without realizing it! You're applying the same logical structure that software developers use, but in such a natural, conversational way that anyone can do it without any technical knowledge or coding skills!

Why Recipe Prompting Works So Well

Have you noticed that the clearest instructions you've ever received likely came from an IKEA assembly manual or a well-written recipe? That's no coincidence. When humans need to communicate exactly how to do something, step-by-step instructions with decision points are universal. We've been using "algorithms" long before computers existed - we just called them recipes, manuals, and how-to guides!

Now let's explore why this approach works so well with AI...

Recipe prompting works for three fundamental reasons:

1. AI doesn't read minds or between lines: Humans are amazingly good at inferring what someone wants even when instructions are vague. We fill in gaps using context and shared experiences. AI simply can't do this (yet). It processes exactly what you tell it - nothing more, nothing less. Recipe prompting makes your intentions crystal clear at each step, eliminating those gaps of understanding.

2. AI thrives on structure: While modern AI can create the illusion of understanding casual conversation, underneath they're still pattern-matching machines that perform best with clearly structured information. They're like that friend who needs things explained step-by-step rather than in a rambling story.

3. AI has conversation amnesia: During longer conversations, AI can "forget" important context from earlier messages. A recipe format provides a consistent framework they can follow from beginning to end, keeping everything organized even as the conversation grows.

When Should You Use Recipe Prompting?

This technique is like a specialized kitchen tool - not needed for everything, but invaluable in specific situations:

• When you need genuine personalization and want the AI to fully understand your context before generating a response.

• When you need precise results and a generic, one-size-fits-all answer simply won't cut it.

• When your task involves multiple steps and you need to guide the AI through a process, not just give a single instruction.

• When you want control over the outcome by building in review points and adjustment opportunities.

If any of these scenarios sound familiar, recipe prompting will become your new AI superpower.

On the flip side, you don't need this approach when:

• You just want a quick yes/no answer without complications.

• You're asking straightforward factual questions like "What's the capital of France?"

• You're just brainstorming or exploring ideas without a specific goal in mind.

The difference in results is dramatic. With a recipe, you know exactly what you'll get. Without one... well, you might end up with radioactive lasagna.

Real-World Applications (with a Fun Example)

This prompting method works for practically anything you need an AI to help with:

• Creating content that sounds genuinely human and resonates with your audience

• Running simulated job interviews tailored to specific roles and industries

• Designing personalized plans for everything from fitness routines to learning paths

• Generating creative ideas that actually match your specific vision

• And countless other possibilities...

Let's look at a simple example that should make you go, "Aha! Now I get what Germán is talking about!"

Behold the mighty Band-name generator

I have this weird fascination with band names that follow a specific format: a number greater than 1 followed by an animal name in plural form, all in English – things like "22 Bears," "35 Whales," or "47 Porcupines" (don't bother Googling these, I made them up).

If I simply ask ChatGPT to "give me band names," it'll likely generate perfectly fine suggestions like "Neon Shadows," "Quantum Pulse," or "The Germán Martínez Experience." Not bad (especially that last one), but they're nothing like the specific format I'm looking for.

Now look at this Recipe Prompt:

The Band Name Generator!

Let's create a band name following this precise procedure:

1. Ask me for an animal. If what I tell you isn't actually an animal, politely ask again until I provide a valid animal name.

2. Ask me for a number greater than 1 but less than 999. If I don't give you a number, or if the number is 1 or less or greater than 999, ask again until I provide a valid number.

3. Create a band name by combining the number and the animal name in plural form (in English). Example: "37 Monkeys"

Want to see this in action? I've actually created a custom GPT with this exact prompt and named it "BandNameology" – you can try it yourself !

When you chat with it:

• It won't be fooled (hopefully): Try telling it "table" is an animal, and it'll politely ask you for a real animal instead.

• It follows the procedure precisely: It won't get distracted or suddenly start suggesting band logos instead of names. It stays on track!

• You'll get exactly what you're looking for: A perfect band name in exactly the format you wanted!

This is a simple example, but it demonstrates why this method is so powerful. Instead of just giving instructions, you're designing an entire conversation flow.

It’s not about giving detailed instructions, it’s about designing a structured conversation. And even if we’re just talking about band names, this approach can be used in many other (more important) situations.

You can design conversations to solve complex problems –you’d be the MacGyver of prompts (just missing the ’80s mullet and Swiss army knife)

Creating Your Own Recipe Prompts

After experimenting with this technique for months, I've discovered it's incredibly versatile. You can apply it to virtually anything – from highly structured analytical tasks to guiding creative processes. It works for technical writing, creative brainstorming, and even design thinking workshops!

Start With Your End Goal

Before you write a single word of your prompt, take a deep breath (seriously – take another one), and get crystal clear about exactly what successful output looks like.

There's a world of difference between asking "help me with my podcast" and saying "I need a structured script for a 20-minute technology podcast that will keep my audience of non-technical professionals engaged and build credibility for my consulting business."

The clearer your vision, the more precise your Recipe Prompt can be.

Map the Conversation Journey

What makes Recipe Prompting truly powerful is that you're designing the entire conversation flow, not just the initial request:

• What background information does the AI need before it can start generating useful content?

• At which critical points should you review progress and potentially redirect?

• How should the AI handle confusion or unclear responses from you?

This approach lets you guide the AI precisely where you want it to go, creating a reliable path rather than hoping the AI somehow reads your mind.

What if…?

My favorite aspect of this technique is how it handles different scenarios with "if-then" branches:

• "If I like this direction, develop it further by..."

• "If I'm not satisfied, ask me specifically what aspect I want to change..."

• "If my answer is vague or ambiguous, ask follow-up questions about..."

These conditional branches transform what would be a static, one-way information dump into a dynamic, adaptive conversation that responds intelligently to your feedback.

Programming Tricks Anyone Can Use (No Coding Required!)

Don't worry - you absolutely don't need any coding knowledge to use Recipe Prompting! But I want to show you how some basic programming concepts are actually just formalized versions of everyday thinking patterns you already use naturally. At their core, programming and recipes are both just ways of giving clear step-by-step instructions. 😊

"If This, Then That" Branches

In programming this is called an "if-then-else statement," but we use this logic every day: "If it's raining, bring an umbrella. If not, wear sunscreen." It's just making different choices based on conditions.

In your prompt:

After showing me each potential blog article idea:

- If I respond with "interested," develop a complete outline with introduction, three main sections, and conclusion.

- If I respond with "not convinced," ask me specifically what I don't like (topic, approach, or style) and then generate three new ideas modifying only that aspect.

This creates a conversation that adapts intelligently to your feedback instead of just giving you static responses.

Do This Several Times

Programmers call this a "for loop," but it's just a fancy way of saying "do this exact same thing X times" - like making multiple cookies from the same recipe.

In your prompt:

Generate 5 titles for articles about AI applications in different industries, following this exact format for each one:

- Each title must begin with "How AI is Transforming"
- Followed by a different industry name (healthcare, finance, education, etc.)
- Ending with a specific benefit or advantage

Follow this precise format for all 5 titles, using 5 different industries.

This is perfect when you need the AI to generate multiple variations of something using the same pattern, saving you from having to repeat yourself.

Keep Going Until…

Programmers call this a "while loop," but it's just like saying "keep adding salt until the soup tastes right" - continue until a condition is met.

In your prompt:

Help me create a slogan for my plant business following this process:

1) Show me an initial slogan draft.
2) Ask what specific aspects I like and don't like.
3) Create an improved version based on my feedback.
4) Repeat steps 2 and 3 until I respond with the exact phrase "this is perfect."

This is ideal when you don't know in advance how many iterations will be needed to achieve the result you want - you're building in a refinement process.

Conversation Menu

In programming, this is called a "switch statement" or "case," but think of it as a restaurant menu where each selection leads to a completely different experience.

First, ask me what type of content I want to create (blog, video, or presentation). Then follow these instructions based on my choice:

If I choose blog: Ask me the ideal length (short, medium, or long). Then create an outline with introduction, at least 3 subtitles, and conclusion.

If I choose video: Ask me whether it's educational or promotional. Then create a script divided into scenes with approximate timing for each segment.

If I choose presentation: Ask me whether it's for clients or for my team. Then structure the slides with numbered sections, key content points, and speaker notes.

This approach is perfect when you want to offer several distinct paths through a conversation, each with its own specialized process. It's like creating a choose-your-own-adventure story where each choice leads somewhere meaningful.

Give the AI a Memory

Programmers call these "variables," but it's simply asking the AI to remember something important throughout your conversation.

After asking about my industry, remember to use this information to customize all examples, analogies, and terminology throughout our entire conversation.

This creates continuity and personalization throughout longer interactions, ensuring the AI doesn't "forget" crucial context that shapes everything that follows.

Create Reusable Conversation Shortcuts

In programming these are called "functions," but think of them as mini-recipes you can call upon whenever needed without rewriting the instructions.

When I type the command "SUMMARIZE:" followed by any text, always respond with:

- One sentence capturing the core idea in plain language
- The three most important takeaways in bullet points
- Any action items or next steps

This way, I can get consistent, structured summaries throughout our conversation without repeating these instructions.

This technique is invaluable for creating consistent structures you'll use repeatedly in a conversation, saving you from typing the same instructions over and over.

The beauty of all these techniques is that you don't need to use technical programming terms like "if-then-else" or "while loops" in your actual prompts. Just use everyday language to explain the process you want the AI to follow, just as you would explain it to a person.

I promise that once you start using these patterns, your AI conversations will achieve a level of clarity and precision that previously seemed possible only for programmers. You'll have computational thinking superpowers without writing a single line of code (though knowing a bit of Python never hurts!).

A Powerful New Tool in Your AI Toolkit

You know what's most exciting about technology evolution? The moment when something that was once only accessible to specialists suddenly becomes available to everyone.

That's exactly what's happening with Recipe Prompting: you now have the power to direct AI with the precision of a programmer, but using everyday language that feels natural. No coding required!

Most people still use AI by tossing in vague requests and crossing their fingers, hoping the AI somehow reads their mind. Sometimes it works, often it doesn't. With Recipe Prompting, you're flipping that relationship entirely - you're designing intentional conversations from the start that lead precisely where you want them to go. It's the difference between hoping for the best and creating a reliable path to results.

The Results Speak for Themselves

Here's a quick challenge for you:

1. Pick something you need to accomplish with ChatGPT or Claude today

2. Before defaulting to your usual approach, invest just 2 minutes in designing a simple "recipe" prompt

3. Compare the results side-by-side with your usual approach

4. Feel slightly amazed, then make it a habit

Don't worry if your first attempts aren't perfect - just like learning to cook, your early "recipes" might be a bit messy. But I promise you'll improve rapidly, and the difference in results will make it worth the small learning curve.

Before You Go

I hope this technique helps you unlock the full potential of tools like ChatGPT and Claude. It's transformed how I work with AI, and I'm constantly finding new ways to apply it.

If you'd like to see Recipe Prompting in action, try out my custom GPT called "BandNameology" that creates band names following the exact pattern we discussed. While we're at it, stay tuned for my upcoming post about creating your own custom GPTs from scratch!

Here's a little homework assignment if you're up for it: Share your best recipe prompt in the comments section below. The most creative and original one wins bragging rights (and maybe a surprise!).

Until next time!
G

Hey! I'm Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I've adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language. See you in the next one!

Missing the ChatGPT image generation buzz? You've either been meditating in a cave or living under a rock. No joke—my social feeds are absolutely swamped with these images. Time to see what all the fuss is about!

Speaking of emotions, have you ever had to admit you were completely wrong about something? That's exactly what I'm doing right now, and surprisingly, I'm actually enjoying it.

For ages, I've sworn by Midjourney for creating images, telling everyone it was the gold standard (despite its learning curve). Whenever anyone asked about ChatGPT's DALL-E 3 image generator, I'd roll my eyes and say it was so... let's call it "underwhelming"... that it wasn't worth your time. I couldn't even be bothered to write about it. Why waste the keystrokes, right?

Then last week, OpenAI dropped image generation with ChatGPT-4o and... KABOOM! Suddenly I'm creating images that rival or outshine my Midjourney work, but with a game-changing difference: everything happens right inside the conversation.

We're not talking about a minor upgrade. This is like trading in your clunky Nokia brick for the latest iPhone. The quality jump is so dramatic that here I am, enthusiastically recommend it! Well played, OpenAI, well played...

What makes this update mind-blowing isn't just the stunning image quality—it's how it actually gets what you're asking for. Request text in your image? Done. Ask to keep the same scene but swap characters? No problem. It's literally like having a personal graphic designer on standby.

Oh, and that Nokia/iPhone comparison image above? Yeah, I just created that in ChatGPT moments ago.

Super cool, right? Let's dive into what this model can do and why you should absolutely start playing with it.

What really changed in ChatGPT

What makes this new image generator truly special? In a nutshell, it finally understands what you actually want.

Bye bye hieroglyphics, hello perfectly readable text

Ever tried creating an image with text using DALL-E or Midjourney? I bet you ended up with gibberish like "Wxtbz Ltmnp" instead of your carefully crafted message.

ChatGPT-4o finally cracks this code. Now you can request posters, signs, and even whiteboards plotting world domination—and the text comes out crystal clear every time.

Create an image of a cartoon supervillain (with cape and dramatic pose) standing next to a whiteboard. The whiteboard should clearly show "EVIL PLAN TO DOMINATE THE WORLD:" with three numbered steps: "1. Generate amazing images with ChatGPT", "2. Build social media followers", "3. World domination".

The Conversation Actually Matters

Imagine you just created the perfect supervillain explaining his evil plan, but you think "it would be better with a volcano in the background." Before, you would have had to write the ENTIRE prompt again, and I'm sure you'd end up with a completely different villain.

With ChatGPT, you simply say "add an erupting volcano to the background" and—bam!—same villain, same whiteboard, same evil scheme... just with an epic volcano that perfectly enhances the villainy vibe.

Take the same image without changing the character or whiteboard and put an erupting volcano in the background.

It Actually Understands Styles

If I ask ChatGPT for an image in Roy Lichtenstein's style, I don't need to spell out that I want those iconic comic book dots, bold primary colors, and thick black outlines. It just gets it.

Now convert that image to Pop Art in Roy Lichtenstein's style

This intuitive understanding makes creation ridiculously easy. You can simply say:

• "A scene in Studio Ghibli style"

• "A vintage car poster"

And you'll get exactly what you're picturing in your head—no need to meticulously describe every tiny detail of the style.

I can already hear you thinking: "Hold up, Germán... Roy Lichtenstein? Studio Ghibli? What about copyright issues?!" Don't worry, I promise we'll get into all that legal stuff, but first let me show you more of what this thing can do...

Playing with ChatGPT

The most fun part about getting a new toy is testing it right away, don't you think? I've only had a short time to experiment, but here are some of my favorite tests.

Note: This section contains AI-generated images mimicking well-known styles.

Create an image of Napoleon Bonaparte as an Instagram influencer in 2025. Use neon yellow glasses.

Create an image of the Beatles crossing Times Square in New York during rush hour, as if it were Abbey Road.

Now, what would happen if I gave it an existing image and asked it to transform it?

The Original Image

I uploaded this selfie I took while cooking a few years back (yes, I rocked a mustache back then, and I solemnly swear this look will make a comeback someday).

Here are the prompts I used and what came out:

Convert this image to Studio Ghibli style.

Now with The Simpsons style

Now with Pixar style

Now with original Scooby Doo style

I think the Pixar style added a few pounds! They say cameras add ten pounds, but I wasn't expecting a complete transformation...

...Studio Ghibli, The Simpsons, Pixar, Scooby Doo. Looks like we need to talk about the elephant in the room: copyright issues.

Is What I Just Did Actually Legal? (I sure hope so)

Can I really ask AI to mimic these well-known styles?
The short answer is: it depends.

Disclaimer: I'm not a lawyer—and neither is ChatGPT. If you have serious concerns, please consult a legal professional!

Inspiration vs. Copying

When you ask ChatGPT for an image "in The Simpsons style," it's not cutting and pasting actual scenes from the show. Instead, it's mimicking the drawing style to create something entirely new (in this case, my glorious mustached selfie).

It's similar to asking an artsy friend to draw you "Simpsons-style." They wouldn't be breaking any laws by using those distinctive yellow characters, they're just borrowing a visual language, not stealing specific characters. Imitating a style, on its own, generally isn't considered a legal issue.

The key difference? While human artists learn styles by watching shows and practicing for years, ChatGPT-4o was trained directly on millions of Simpsons images (and other styles). That's where things get legally fuzzy.

What does the law say?

The uncomfortable truth is that our current laws weren't designed for a world where AI can learn and replicate artistic styles in seconds. We're in uncharted territory. (Let's face it, laws always play catch-up with technology.)

Here's what most experts generally agree on:

• Specific characters are protected (you can't generate Homer Simpson or Totoro)

• Specific artworks are protected (you can't replicate a scene from Toy Story)

• General styles usually aren't protected (like "manga style" or "cubist style")

The gray area? When you generate images with highly distinctive, immediately recognizable styles from specific studios (looking at you, Studio Ghibli).

How I see it

1. For experimentation and learning: I use any style I want when just playing around (like my mustache photo makeovers)

2. For social media: I clearly label these as "AI-generated images inspired by [style]"

3. Out of basic respect: I acknowledge that real human artists spent years developing these distinctive styles

The reality is both laws and society are still figuring this all out. In a few years, we'll likely have clearer guidelines about what's acceptable with AI-generated images—and about the ethics of training these models on artists' work in the first place.

Meanwhile, my advice? Have fun experimenting, but be thoughtful about boundaries, especially if you plan to use these images beyond personal enjoyment.

Before wrapping up, I should mention that this technology still has a few quirks:

A Few Current Limitations

• Edge clipping issues: It sometimes trims the edges of longer images (like posters), particularly at the bottom

• Creative liberties: It occasionally adds details you never asked for

• Complexity overload: When you ask it to generate images with more than 10-20 different concepts into one prompt, it gets overwhelmed and starts mixing things up

Still, compared to what we had before, these are just small details. I'm pretty confident OpenAI is already patching these issues as we speak.

See you next time!
G

Hey! I'm Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I've adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language. See you in the next one!

I've always had a soft spot for Claude. There's something about the way our conversations flow that feels "more human" to me, though I still can't quite put my finger on why.

Now, with Claude 3.7 Sonnet, Anthropic has released an assistant with two thinking modes: Normal and Extended.

If you've read Daniel Kahneman's "Thinking, Fast and Slow", this might sound familiar. Kahneman talks about how our brains have two different decision-making systems: a quick, intuitive one and a slower, more analytical one. What's cool is that Claude now seems to have these systems too.

Just like we intuitively know when to trust our gut versus when to slow down and think step-by-step, Claude can now shift between these two modes... all while being the same assistant underneath. Let me show you how this works.

A Hybrid Reasoning Model

Claude 3.7 Sonnet is what Anthropic calls a "hybrid reasoning model." But what's interesting isn't just that it offers two different modes (which technically we could get by using separate models), but that it's one brain with two different ways of processing information.

This works differently than ChatGPT (and other assistants), where you have to choose between different models like GPT-4o or o1 to get different thinking styles. With Claude 3.7 Sonnet, the same architecture and training powers both thinking modes, it's the same AI at its core, just trained to think in different ways.

In normal mode, Claude gives you quick, to-the-point responses. Switch to extended thinking mode, and suddenly you can watch its reasoning unfold step by step, almost like seeing the thoughts form in real time.

One Brain with Two Ways of Thinking

Picture this: you visit a restaurant where the chef gives you a choice - either enjoy the perfectly plated dish served directly to your table, or step into the kitchen and watch them work their magic with each ingredient. Same chef, same delicious food, but totally different experiences.

Most AI assistants make you switch between different models for different tasks. With OpenAI, I use GPT-4o for everyday conversations, but I jump to o1 or o3-mini when I need deeper analysis. They're basically different AI brains designed to think in different ways.

What makes Claude 3.7 Sonnet special is that it's one brain that can operate in two different ways.

It's not like Claude ducks into a phone booth and emerges as "Super Claude" (sorry for the vintage analogy). It's exactly the same model with all the same knowledge. The difference is whether it shows us its thinking process or just the final answer.

In normal mode, Claude's responses are direct and to the point. Like asking a friend, "Where's the best coffee in town?" and getting "The corner of Main Street" as an immediate answer.

In extended mode, that same friend might say: "Well, to find the best coffee, I need to consider what makes great coffee. There's bean quality, brewing methods, atmosphere... Based on these factors, I'd recommend three places that really stand out for different reasons..."

Why This Difference Matters to Me

• It mirrors how our brains actually work: We don't swap out our brains when switching from casual chat to deep problem-solving.

• It helps me catch mistakes: Seeing the full reasoning process makes it obvious when something doesn't quite add up.

What Claude is really offering isn't two different tools, but a way to either see its entire thought process or just the final answer. And that changes how I interact with AI.

The Two Thinking Modes

I've been playing around with both modes of Claude 3.7 Sonnet, and I'm starting to get a feel for when each one works best. Here's what I've discovered so far:

Normal Mode for Everyday Use

In normal mode, Claude cuts to the chase. You get clear, direct answers without all the behind-the-scenes mental work.

It's like asking a chef "how much cumin should I add?" and getting a simple "one teaspoon" in response. No lecture on spice profiles or cooking chemistry - just the practical answer you need right now.

This is my go-to mode for 90% of what I do, and it's perfect when I need:

• Quick creative sparks: "Hey Claude, give me 5 catchy titles for my blog post about your two thinking modes" 😄

• Ready-to-send text: "Draft a friendly email declining this event invitation"

• Straight-to-the-point explanations: "Explain inflation in two clear paragraphs"

• Natural back-and-forth: When I want a conversation that flows without long-winded explanations getting in the way

Extended Mode: When You Need to See the "How"

This is where things get really cool. Extended mode doesn't just give you Claude's answers but shows you exactly how it got there, one step at a time.

It's not just about WHAT Claude thinks, but HOW it arrived there.

Think of our chef analogy again - instead of simply handing you a teaspoon of cumin, the chef invites you into the kitchen. "Come in, let me show you how I balance these flavors, why this spice works with these ingredients, and how I know exactly when it's enough." Same chef, same dish, but you gain insights you'd never get from just eating the final meal.

The best part? Extended mode doesn't just show you the thinking behind the answer. When Claude thinks step by step, it often comes up with different (and usually better) answers than in normal mode. It's like watching someone solve a puzzle - when they slow down and work through each possibility one by one, they spot solutions they might miss when rushing.

This mode becomes invaluable when:

• You need rock-solid confidence: Seeing each reasoning step lets you spot any logical flaws or questionable assumptions

• The problem is really tricky: When it's the kind of question where three smart friends would give you three totally different answers

• You want to get better at problem-solving: By watching Claude's problem-solving approach, you pick up frameworks for tackling similar challenges yourself

Here's what it looks like in action:

A Real-World Example

This example shows how Claude approaches problems in extended mode, revealing its entire thought process.

First I ask a question…

Me:

I've been working for five years as a financial analyst at a large company, with good salary and stability. I was offered a position at a startup, where the initial salary is lower but with stock options and the possibility of accelerated growth. What factors should I evaluate in depth before making a decision?

Then Claude starts thinking…

Claude's thinking process (took 7 seconds):

The person is asking me to help them evaluate factors they should consider before deciding between staying at their large company where they have a good salary and stability, or accepting an offer at a startup with lower initial salary but stock options and potential for accelerated growth.

This is a career decision that involves weighing various factors related to financial implications, career growth, risk tolerance, and personal preferences. Let me think about what factors they should consider in depth:

Financial considerations:
   - Current compensation gap - how much lower is the startup salary and can they afford this reduction?
   - Valuation of stock options - what percentage of the company do they represent? What's the current valuation? What's the vesting schedule?
   - Benefits comparison - health insurance, retirement plans, other perks might differ significantly
   - Potential financial upside - realistic assessment of stock option value if the startup succeeds
   - Long-term financial planning - impact on savings, debt repayment, major life purchases

Career growth and development:
   - Learning opportunities at each company
   - Skills development potential
   - Career trajectory and advancement speed
   - Responsibility level and role scope
   - Résumé building/career marketability
   - Network expansion opportunities

Startup-specific considerations:
   - Funding status and runway (how much money they have, how long it will last)
   - Business model viability and market opportunity
   - Leadership team experience and track record
   - Growth metrics and traction so far
   - Competitive landscape
   - Exit strategy possibilities (acquisition, IPO)

Stability and risk factors:
   - Job security differences
   - Startup survival statistics (most fail)
   - Personal financial safety net
   - Industry and economic conditions
   - Personal risk tolerance

Work-life factors:
   - Work culture differences
   - Hours/flexibility expectations
   - Commute/location considerations
   - Work-life balance
   - Alignment with personal values and interests
   - Team dynamics and potential colleagues

Personal considerations:
   - Career goals and how each option helps achieve them
   - Life stage and personal circumstances (family obligations, etc.)
   - Passion for the startup's mission vs. current role
   - Long-term career direction

I'll put these factors together in a comprehensive but concise response that helps the person evaluate this important career decision.

After that Claude is ready to respond…

Claude response:

When considering a move from your established financial analyst role to a startup opportunity, here are the key factors to evaluate:
Financial Considerations

Quantify the total compensation gap (including benefits)
Assess stock option value (percentage ownership, current valuation, vesting schedule)
Determine if you have sufficient savings to manage the salary reduction
Evaluate the startup's funding status and runway

Career Development
   - Compare growth trajectories at both companies
   - Assess skill development opportunities
   - Consider role scope expansion and leadership potential
   - Evaluate how each path enhances your market value

Startup Viability
   - Research the founding team's track record
   - Analyze the business model and market opportunity
   - Examine growth metrics and customer traction
   - Understand the competitive landscape and exit potential

Personal Factors
   - Honestly assess your risk tolerance
   - Consider work culture fit and values alignment
   - Evaluate life circumstances (family needs, financial obligations)
   - Reflect on which opportunity better serves your long-term career vision

Remember that startup compensation is high-risk/high-reward, while larger companies offer stability but potentially slower advancement. The right choice depends on your personal circumstances, career goals, and comfort with uncertainty.

I love seeing how Claude organizes its thinking! Did you notice how it first confirms it understands my question? Then it puts together a solid framework before giving a clear, well-structured answer.

This window into its reasoning process shows exactly how it gets from question to answer (something we'd never see in normal mode).

Thinking in Action

Think of extended mode as the "behind-the-scenes" bonus features on a DVD (another vintage reference 😊). Most of us just watch the movie and enjoy it, but sometimes we're curious about how the magic happened behind the camera.

Normal mode is my go-to for almost everything:

• Natural conversations ← (this is where Claude truly shines for me)

• Everyday questions that need quick answers

• Creative writing and brainstorming sessions

• Pretty much anything I need to handle efficiently

When to Fire Up a New Chat in Extended Mode

There are specific situations where it's worth starting a fresh chat in extended mode Important note: you can't go from one mode to another in mid-conversation, but you can go from normal to extended mode selecting the “Retry” action in a response, Once you change to extended mode you cannot go back to normal model in that conversation (at least I haven’t found out how)

Extended mode really pays off when:

• You're dealing with high-stakes information: When you're working on something that really matters or trying to make decisions based on lots of data, seeing Claude's reasoning helps you spot any flaws in the logic or conclusions.

• You're tackling complex subjects: It's like having a master teacher who doesn't just hand you the answer but walks you through their entire problem-solving approach, helping you build your own understanding.

• Something feels off about an answer: We've all had that moment when an answer seems reasonable but something about it just doesn't feel right. When that happens, try the same question in extended mode.

What's Working For Me

Here's what I've been doing that seems to work pretty well:

1. Default to normal mode for almost everything (saves time and flows better)

2. Open a new extended mode chat when you need deeper insight or have doubts

3. Ask the same question (sometimes adding a bit more context)

4. Compare both answers to see the difference in depth and reasoning

Pro tip: Remember you can't switch modes within a conversation - you must start a completely fresh chat with extended mode selected from the beginning.

Bonus Track: Claude Code

When Anthropic launched Claude 3.7 Sonnet, they also introduced something special for the developers among us: Claude Code.

This new tool lets coders give tasks to Claude straight from their command line - basically bringing Claude's smarts right into how they work. If you want to see how it works, there's a great video that shows it all.

Where We're Heading

I really like how Anthropic has approached Claude 3.7 Sonnet, giving us two thinking modes in one model instead of making us choose between different models like all other AI companies do. This all-in-one approach just feels more natural and intuitive.

My wish for the future? Being able to seamlessly toggle between modes mid-conversation. Imagine chatting with Claude and simply saying, "Hold on - can you show me your thinking on that?" just like you might ask a friend to explain their reasoning. That kind of flexibility would be a game-changer!

See you around!
G

PS: Phone booths? DVDs? I guess my vintage tech references are giving away my age...

Hey! I'm Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I've adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language. See you in the next one!