I have always enjoyed programming. The process of writing code line by line, letter by letter, and iterating until it works is quite an experience. Even when there are problems or the documentation isn't clear—which can be frustrating—it is a process of creation.
Now, with the arrival of AI agents, our work is different since it is no longer necessary to write every single line or search through all the documentation. You write a prompt and that's it... Is Vibe Coding really like that?
The Rise of Vibe Coding
Recently, a new concept has emerged: Vibe Coding. It consists of writing a prompt in natural language and letting the AI do the rest. The term was introduced by Andrej Karpathy in February 2025.
There are a multitude of tools like Cursor or Codex (the ones I use most today) that allow you to "vibe code," and it is certainly a revolution. In my case, I observe how I can develop much faster and dedicate my mental energy to what is truly important, such as designing the architecture, thinking about how to validate the idea, and the essential parts of the product. Then, using the AI as a programmer, I specify exactly what I need, and it takes care of the rest.
Personally, I move fast when I have deep knowledge of the tools I work with; I know what I need to review and what I can leave a bit more to the side, and it saves me time searching through documentation. However, I always look up important parameters in the documentation to validate that they are correct for the software version I am using and that they do what they are supposed to.
I love vibe coding, although I insist that I know what I want to do, I know how it should be done, and this is precisely what allows me to move fast.
My experience "vibe coding"
In my opinion, vibe coding requires knowing what you want to achieve. It requires understanding the foundations of computing, the tools you use, and having good automated tests that allow you to make changes with greater confidence.
If you don't know how the code should be structured, technical debt can become a problem, as you can end up with monstrous spaghetti code quite easily if you don't guide the LLM correctly. The key here is to be specific in the prompt—the more, the better. For this, Cursor has released "Plan mode," which first prepares a technical design document in markdown that you can refine before it executes.
For me, the important thing here is supervision, which requires a human to think about what is needed, define what is to be done, and guarantee that this is what actually happens.
There are also no-code tools like n8n, which I personally love for certain tasks and automations due to development speed, but I wouldn't use them for more advanced processes as I feel they become more complex to maintain or develop in my case.
Finally, an important note: vibe coding requires knowing. Understanding systems, protocols, patterns, and limits. That is why it is important to study, test, and make mistakes, because that is how you really learn.
We will continue talking about these topics in future posts, discussing the evolution of these tools as well as future proposals like TOON (an alternative to JSON that uses fewer tokens) or work methodologies.