I've been working on some issues with the site, and something very common happened to me again. Every time I don’t like the AI’s answer and ask again about the same problem, little by little the agent starts drifting further away from the main issue.
Of course, this usually happens with very specific problems, probably because of a lack of context—and also when I already understand the problem and the possible solution. I admit I’m too lazy to provide the full context (I know it’s necessary). So I’ve come to this axiom: you need to set very clear boundaries for the AI to avoid crazy answers/solutions, and you must stay focused on the problem. Don’t fall into the temptation of completely changing the approach (at least at the beginning). Ask it about the reasons for suggesting changes—you must be absolutely sure.
My case:
A GitHub workflow failed with a 'No space left on device' error. I investigated a bit and quickly noticed the huge size of the recently added PyTorch library. I realized we could install a minimal variant without GPU support. That was the problem, and that was the solution.
In the meantime, the AI insisted on changing my Dockerfile, removing the .venv folder, and doing some weird things with the models. To be honest, it’s not entirely wrong—if you need more space, you might have to delete files. But if the AI suggests that, it MUST tell you exactly which files to remove. Otherwise, delegating that decision to the AI could make you lose control of your development process.
So I asked the AI to remove packages and install the minimal one, all with uv. But I couldn’t get it to work. Again and again, the approach failed. I feel that neither ChatGPT nor Gemini fully understand uv yet, so the suggestions kept leading to the same bad results.
The solution: read the documentation and make the adjustments myself. Then, use the AI to help with documentation, style, and improving the current code.
Message from Charly to AI: Stay focused—that’s the key!