The first important lesson in vibe coding

As I mentioned once or twice, I'm currently enrolled in a course that teaches developers how to use AI. Among many pros and cons are tasks which encourage the approach of “vibe coding the shit out of it”.

I decided to ask people at work how they approach the tasks. I thought that everyone is doing it the way I do: create a code and then adjust, adjust, adjust. To my surprise, some people thought that those tasks could be done with just one prompt. So if the prompt doesn't nail it, they adjust the prompt and start from scratch.

And so, for better or worse, I decided to do that for the next task.

I wrote a prompt, executed it, and realized it does not solve the task. So I deleted all the files the AI generated, adjusted the prompt, executed it, realized it does not solve the task, deleted the files, adjusted the prompt, executed it...

I did not count how many approaches it took.

Frankly, I don't remember if it even solved the task 😂

But I did learn something from it. And the funny part is, you can get to this conclusion without adjusting the prompt, but just by executing it multiple times and studying the result.

Why?

Well, the results will not be the same. They will differ. And sometimes the implementation method will be 80% different from the previous approach chosen.

And don't get me started on the fact that they might not work at first, as well.

AI will not necessarily lead you to a working solution.

That’s not how it works.

Not because there are multiple ways to solve a problem (if it generates code that doesn’t work, then no problem was solved, right?), but because the AI is not the competent assistant we hope it is.

Instead of doing it right, the AI keeps “starting over” with a new approach. This lack of stability, control, and the developer’s over-reliance on luck creates unpredictability. You might get a great result once and then spend 10 more runs trying to replicate something remotely similar. Without any chances.

Conclusions?

The idea of game templates comes back to me, as if I can’t achieve the same result by prompting, I’d better save my work for later.

And I’ll stick to tweaking the code, rather than making a perfect prompt.

I guess good developer practices still work, especially in vibe coding.