A Paradigm Shift

How is it possible that Claude Code is so skilled, that it knows exactly how to do everything?

On April 1 of this year, we got our first glimpse of how the tool was built from the inside. It wasn't a deliberate decision by the company, but rather an oversight. During an update, a file called a source map was accidentally published. For those who don't code, it's enough to know that it's an auxiliary file used by developers to debug errors. It's a kind of map that allows the original code to be reconstructed from the compressed version that's distributed. With that map in hand, anyone could reassemble the complete source code. Someone spotted it, posted it on social media, and soon copies were circulating everywhere. Although the file was quickly removed, it was already too late. On the internet, once something is published, it can't be unpublished.

A company valued at hundreds of billions of dollars, with some of the best engineering teams in the world, made a rookie mistake. Imagine the mistakes we might make.

But the important thing wasn't the mistake itself. What mattered was what it revealed.

Many expected to find a secret formula, something sophisticated and out of reach. What emerged was something else entirely. Much of the "magic" wasn't in the model itself but in everything surrounding it: meticulously written instructions, a set of tools the model can use (reading files, executing commands, searching for information), and a cycle that allows it to act, examine the result, and correct itself.

Knowing that other people would be able to study that architecture was, for me, liberating. Not because I celebrate the error, but because I'm one of those who believe that when something is done very well, it should be available to the world. What is learned is shared, improved, and multiplied.

Claude Code is like having an expert programmer sitting right next to you who can read your files, write code, run commands, and fix errors, all while conversing in natural language. But it can also write your website's privacy terms and conditions, or your company's entire sales process, or review your entire customer portfolio and sort it from highest to lowest purchase volume.

It doesn't forget a thing while working. It asks for permission before taking action. It understands your project. It has real tools, not just words. It can read your files, edit them, and execute real commands, such as installing a database, and it can browse the internet and retrieve any content from a website. We're not just asking questions and reading answers — in this case, we delegate concrete, real tasks to be executed, and it reports back in real time on its progress until it's done. And if something goes wrong, it can undo it. That's just as valuable as being able to do it in the first place.

Claude Code can be used via the terminal—the black screen where you type commands. It can also be used through its native desktop app. Its interface is very user-friendly, so there's no excuse not to give it a try.

Once I'd become more proficient with the tool, I decided to let my partner at Wispro in on the secret.

The Revelation

You can't help but use Claude Code.

He's a systems engineer who loves processes and formalities. I was giving him a tool to organize everything and automate repetitive tasks, tasks he enjoyed doing, but always preferred to automate. It was like giving a Ferrari to someone who loves to drive.

My partner really loves processes and is super creative when it comes to designing them. Now all those processes he's been designing are much more automated with AI.

A few days ago, while tinkering with the app he built himself to process payments, he added a feature to retry charges on the cards of customers in arrears, all using Claude Code. That's a five-minute manual process for each case; I've also been doing that task for years and always wondered how it couldn't be automated. My partner did it in one afternoon.

I think this was on a Thursday. We met on Monday, and his wide-eyed, sleep-deprived look was more than obvious.

"Don't even tell me, you were nonstop for all four days doing stuff," I told him.

"Yeah. I hit the token rate limit five times," he told me.

Claude lets you work but makes sure you don't use up all the tokens by yourself. It gives you 5-hour, 24-hour, and 7-day windows so you don't abuse its generosity. Hitting the limit five times in four days says it all: he used it a lot. It's a pattern I've seen repeat itself with every person who discovers the tool. The day you start, it's three to seven days nonstop. New toy. You don't sleep well. You eat whatever you can get your hands on. You cancel plans. Your family asks you what's wrong. And you can't explain it to them because what you're doing is so new that you don't have the words to describe it.

Everyone at the company is now going crazy using AI all day, every day. They post photos on Instagram showing that it's the weekend and they're still online. The first few months are killer. The addiction is total. It's instant dopamine. Every task you complete gives you a sense of satisfaction that pushes you to tackle the next one. You can't stop.

In those three days, my partner set up a system to process payments that came in via WhatsApp with photos of receipts. The problem was an old one. Customers would send a photo of their payment receipt via WhatsApp, and someone had to open that photo, read the details, verify the amount, look up the customer in the system, enter the information manually, and confirm receipt. This process took at least 5 minutes per payment. 100 payments a day. 8 hours of work for one person. Just to enter payment receipts. A process that every company is familiar with and that everyone tries to avoid...

My partner developed an app to solve that problem. Customers sent their payment receipts via WhatsApp, just as they already did. Someone on the team would download them and upload them to the app, and from there, the app took over. The AI read each receipt and extracted all the data, such as the amount, date, bank, and transaction number. It then checked to make sure everything made sense and that the data matched up. With that data, the app would go to Stripe¹, find the customer's outstanding invoice, and verify that the payment matched the amount owed. Only then would the receipt appear on a simple screen, ready for a person to review. If everything was in order, all it took was approval, and the payment was automatically applied to that invoice.

It was the same process that used to be done manually, receipt by receipt, but in seconds instead of minutes. The customer continued to pay as usual, and the only remaining manual step was uploading the file. Everything that used to take time was handled by the app, and the final decision remained in human hands. AI did the heavy lifting, and the person oversaw it. Three months later, that tool had already become something else entirely. It grew significantly and continues to add new features.

Always think about the task that's most repetitive and easiest to do. Find one that meets those two conditions and let it take over. The difference between asking AI how to do something and doing it in Claude Code is that Claude Code actually does it. It doesn't give you a plan. It gives you the result.

The Claude family of models includes Opus, which is the smartest, for in-depth thinking and problem-solving, and I use it all the time. Sonnet is for more repetitive tasks. Haiku is the lightest, for quick answers. And Fable is for extremely complex tasks, but that's also why it's more expensive. Every model has its place. The key is knowing when to use which one.

Claude can read files, images, and PDFs. You copy an image, paste it in, and ask, "How do I solve this?" The model interprets the image, detects patterns, and responds. It can search for text within files, create files and folders, edit them in whole or in part, and compare them to show you the differences. It's a breeze for it to process very large files and find small differences that would take a human hours to spot. Models from other companies can do this, too.

But for me, its greatest strength is its ability to put together a plan. It evaluates all the options, lists them, chooses one, and explains why. Planning is a very complex skill, and its plans almost always work.

It can run scripts, mini-programs it creates itself to speed up tasks, such as scraping a news website. It can install any program you need via the command line, such as Python or DuckDB. With these capabilities, it can practically do anything. And what it can't do today, it will likely be able to do in the next version.

That's how fast this tool evolves. Every update brings capabilities that seemed impossible just a month earlier. And that creates a strange feeling. On the one hand, it's rewarding because it keeps getting better. On the other, it's exhausting because you never finish learning it.

Building Solutions

Let me tell you about a colleague who works on my team. He's not a programmer; he's self-taught. He works for IXC Soft in Brazil. And ever since he joined, I've received very positive feedback on his performance as a sales assistant for prospecting clients, which was his first role at the company. His career path was more geared toward automation; that aspect has always appealed to him more. He's the kind of person who loves technology, explores, studies, and researches on his own. Not all of us are like that, but here's the story. And it's one of my favorites because it shows exactly how someone without a programming degree can solve a problem that's normally handled by teams of programmers.

As a sales rep, I handle trade shows and events. A recurring problem was that there wasn't enough time to attend to people making inquiries at the booth, show them the product, negotiate the sale, record it, and collect their payment. If you have to do this with 80 people a day, you're clearly in trouble. Those were exhausting days where you knew you were losing sales because you couldn't close them in time. It was chaos. Half the time, you couldn't even keep track of who you'd spoken with.

You'd lose leads or, if you did jot something down, you wouldn't know what the person was interested in, the size of their business, or any additional details. We had forms and made an effort to fill them out, but after three days at the trade show, our brains just wouldn't function anymore.

He'd hear us complaining in our chats about how hard it was to close a sale because of how long it took to go through the entire process until we could bill the customers who made a purchase. That's when he began dedicating some of his time to understanding the problem and developing a solution.

No one asked him to do it. He did it because he saw a problem and had the expertise in the available artificial intelligence tools to solve it.

A few days before the most important trade show of all, held in Cartagena de Indias, he called us to show us something he'd prepared. At first I was skeptical about what he had, but his teammates told me, "Leo, you have to see this, it's incredible. We're using it every day."

It was a few days before the trade show that I understood the potential of what he'd created. That's when I congratulated him on what he'd done. Sometimes it's hard to see new things, as if the world were limited to what's right in front of you and nothing else. I still remember that day. He presented us with a product quote generator built on a website about Lovable, developed 100% by him. It was all interactive. You could choose the products and options and generate a quote on the screen using buttons and colors. You could enter a WhatsApp number and an email address so the customer would receive a PDF with the quote. In addition to registering the customer with all their details, which were later used for follow-up, you knew what they'd tried to buy, which products, the final prices, and the trade show discount.

In five minutes, you could do what used to take twenty. And the customer left with everything in hand. At that trade show, we took a huge step forward in organization and time optimization, making the most of our time to secure new sales.

What used to be a mess of papers, half-filled forms, and lost data turned into a clean, fast, and professional process. Customers were amazed. "You guys did this yourselves?" they'd ask us. And we nodded proudly, knowing that a single person had built it using an AI tool in just two weeks, with no budget.

Lovable is a platform that lets you create web applications using natural language, without any programming. You describe what you want, and the AI generates the code, the design, and the full functionality. It's not the only one, there are already other competitors, but it was the first, the disruptive one. I've seen many solutions built this way, by non-programmers, developed by people just like any of us who don't know how to code. People who have a problem, describe it in words, and the AI turns it into a functional app.

Go ahead and create your first app. Seriously. I'm not saying this as some empty motivational phrase. I'm saying it because I've seen it happen dozens of times. People who had never written a single line of code have created functional apps in a single afternoon. Even if it's just to manage a household budget, where everyone can enter expenses and it sends email alerts when an expense exceeds a certain budget.

I've seen very complex projects built this way, with integrations of multiple APIs. An API is an intermediary that allows two systems to communicate with each other without knowing how the other works internally. APIs connect everything through a standard format called JSON, basically a JavaScript object, the universal language that all systems understand, speak, and write. Always try to solve a real problem so the app has a future. Version one is always basic. What matters are the subsequent versions, the ones that bring improvements. Each iteration teaches you something new, and the solution gets refined.

The quote generator later added a process connected via APIs that registered the customer in real time within Wispro, generated invoices for purchases, and sent links to WhatsApp for credit card payments. That wasn't there on day one. It was the result of iterations, research, and obtaining the endpoints, the specific URL where an API receives requests. Without the endpoints, you can't move forward. But once you have them, the pieces fall into place. Like anything you start building, it's done in stages.

The first prompt was created on February 20, 2026. The trade show was on March 9. He had only two weeks. He was able to do it quickly, using the free tools available today and without needing help from anyone else — that's something characteristic of this era: seeing individuals develop solutions on their own that used to require two or three people. Doing the same thing in the pre-AI era would have required a far greater software development effort than it does today.

Many companies have already started automating key processes. A friend told me, "Leo, I've been like a little kid with a Christmas present for months." He's using Claude Code to get things done at his company. He's from Costa Rica, someone who's always on the lookout for new technologies and trying new things. His personality helps him make progress along that path. Not all of us have that natural knack, but we can make the decision to get started.

Another case I'd like to mention is Klari. There, they automated all customer support when a bot or agent provides an incorrect response. Every time this happens, you have to review the entire conversation, the documentation used to train the bot, the prompts from the agents involved, and any API requests that were made. Doing all of that manually takes a long time, and it's very easy to make a mistake or misinterpret some of the information—which you have to review just to understand why it responded incorrectly.

To automate and streamline this support task, they created an MCP² server, which is like an API but optimized for easy use by AI. So now the support team's job consists of using Claude Code to review a ticket, retrieve all associated metadata, gather all relevant information about the issue, diagnose what might be happening, use preloaded skills to follow diagnostic pathways, and finally propose alternative solutions. That process, which used to be done manually, takes between 30 and 40 minutes for someone very well-trained in the product. With AI, it's been reduced to 5 minutes. It's not just about time—it's also about effectiveness, since with Claude, the margin of error is much lower than when a person did it.

Before, they would fix one thing but might break something else. Now, the AI itself runs the tests, verifies that the fix works, and only then closes the case. It doesn't close it until it's certain. That's a radical change.

Speaking with Wispro clients this morning, I asked about ten of them if they were already using AI in any process, or if they had developed any solution, even a minimal one. Out of ten, five had already done something. The other five hadn't, still nothing. Half have already gotten started, and the other half are still watching from the sidelines.

And we're back to square one. Companies are going to become increasingly divided between those that use AI and those that don't. The competitiveness and efficiency that well-implemented AI provides are exponential. It's not a matter of some companies growing less and others growing more. It's a matter of the ones that don't use it disappearing. It sounds harsh. Companies that adopt AI early not only improve but also change the rules of the game for their entire industry.

We already know it's possible. There are real-world examples. The ball is in our court. All that's left is to decide to move forward with something today. And here I want to be blunt: you don't need to have a complex problem. It doesn't have to be something that changes your business forever. It just needs to be something that bothers you, a task you hate doing, a spreadsheet you hate filling out, or a process that steals half an hour from you every day and that you do on autopilot.

There's your first AI project. Set a first goal and get started. AI is great at coming up with plans. In fact, if you don't know where to start, ask the AI where to start. All my new projects start more or less like this — I ask any of the models what I need to do and ask it to recommend where to begin. Most of the time it gives me just enough light to get going. Once you've started, everything begins to unlock faster and you feel motivated because you can see it taking shape.

One example might be copying your bank statement and doing a bank reconciliation with your clients' payments. Upload your clients' file along with the CSV of your bank transactions and ask the model to tell you whether everyone has paid yet or not.

I've even used Claude Code to put together my tax return, but numbers are a sensitive matter, so we can't place blind trust in the results the AI gives us. A language model doesn't guarantee that what it says is true. It can respond with complete confidence and still be wrong. If I ask it whether a number is correct and it says yes, that "yes" is worthless until I verify it. We can't trust a piece of data just because the AI says so.

What makes the difference is the groundwork. In my case, I first gathered all the source information: the year's bank statements, Stripe reports, and expense receipts. Then I asked the model to look up each transaction in those documents and classify it according to a list of categories I had defined beforehand. If an amount didn't appear on any statement or didn't fit into any category, it was flagged for manual review. Only then was the result reliable, because every number could be traced back to its source.

That comes at a cost. Every task requires extra validation work, step by step, and there are no shortcuts. The first attempt almost never turns out perfect, but it teaches you how the tool works and where it goes wrong, and that alone is worth a lot.

Always start by breaking the problem down into small pieces and tackle just one of them. Just one. Not three, not five. One. If it works, move on to the next one. If it doesn't work, adjust and try again. Don't get frustrated if the first attempt fails. Or the second. Iteration is a natural part of using AI — you have to go back and forth several times to get the desired result. Iteration has always existed in software development processes, but now it's a much faster constant than before. You'll often find that AI doesn't understand you. That's not the models' fault. It's the responsibility of the person making the request, the one entering the prompt.

Prompt Engineering

Think of it something like this. If you ask someone to cook you something delicious, they'll bring you whatever that person thinks is delicious. But if you say, "Make me a Neapolitan-style breaded cutlet with crispy fries and no onions," the chances they'll bring you exactly what you want are extremely high. The same goes for AI. The more precise you are, the better it responds. The vaguer the request, the more it makes things up.

This is so important that "prompt engineering" has even been discussed as a specialty, the art of writing clear and precise instructions to get better results from the model. The important thing is knowing how to explain things to it, knowing how to communicate, speaking to it with simple words and clear ideas. And like all communication, it has its rules. It has key elements that can't be left out.

• The role: explicitly tell it, "You're an expert in home appliances." By assigning a role, the model adjusts its vocabulary, level of detail, and tone. Speaking as an expert in central heating is not the same as speaking as an expert in PC games.
• The context: Tell it the brand, how you use it, and what you want to fix. Without context, the AI responds generically. With context, it responds specifically to you.
• The task: Ask it for a step-by-step plan to, for example, bleed a radiator. Be specific about what you expect as a result.
• The format: How do you want it to structure the response, as a list, plain text, a table, or a Markdown file? The format matters just as much as the content.
• The limits: Make sure it doesn't put the appliance you're trying to fix at risk. These limits prevent the AI from going off on tangents and giving you something you don't need.

These five elements seem simple when you read them, but in practice, they make all the difference. At first, I used to write one-line prompts and get frustrated with the results. I'd say, "Give me a summary of this," and it would give me something generic. When I started structuring my prompts with these elements, the quality of the responses changed dramatically. It's like the difference between asking someone for directions by saying, "How do I get there?" versus "How do I get from San Martin Square to the airport by car, without using the highway, in less than 40 minutes?" The second question always gives you a better answer.

If you do all this, you have a great chance of getting what we call a one-shot, the perfect first response, without having to iterate five or ten more times. That's the goal. Not to struggle with the AI to get it to understand you, but to give it everything it needs to answer you correctly on the first try.

My first complex prompt was one I created to categorize conversations in Klari and route them to agents who are experts in each topic. That leads you to other, more specific prompts. It was quite a challenge, and that piece of software is still up and running there. I read a lot to understand this, but once I did, it worked right from the start. It took me longer to write that prompt than to get the result. Working with AI isn't about waiting for the answer, but about phrasing the question well.

A "one-shot" is a very rewarding feeling. And once you experience it, you'll never want to write vague prompts again. If we get creative, we could ask it to write the result in HTML so we can view it in a browser, complete with colors and images, thereby putting together a professional presentation.

So far, we've seen how these tools work, how to use them, and how to get real value out of them. But everything we've done so far has been within our own little world, our company, our project.

Now it's time to look up and look outward. Because while we were learning to use AI, the whole world was changing around us. And what's happening out there is much bigger than any quote generator or any automation tool.

1. Stripe. A payment platform that allows businesses to accept payments online and manage invoices and subscriptions. https://stripe.com ↩
2. MCP (Model Context Protocol). An open protocol that connects artificial intelligence applications to external data sources and tools through a standardized interface. https://modelcontextprotocol.io ↩