My "development environment" - Vivaldi, WezTerm, nvim, AeroSpace and SimpleBar for tiling window manager.

This morning I was asked, if I vibe-coded all or parts of Hule. The asker wasn’t accusing me, they wanted to know if that was possible, since they had a few ideas for the app but weren’t familiar with Python.

Short answer: No. I don’t trust AI generated code.

Longer answer: No, but it depends.

I use a local LLM with nvim to suggest code completions and keep track of things. For example, I keep all my :root variables in one central location, and more often than not, I do not have to type out long CSS blocks, since the LLM understands that .reply-card:hover wants to invert on hover and suggests an according block. Same for git commits, which I instruct an LLM to look over the code and create good commit messages.

I also used claude and codex for code reviews more than once. I, too, am learning to do this (I am a medical humanist and emergency medicine man, not a coder, after all), and especially Claude’s suggestions and warnings of n+1 mistakes saved me hours, if not days, of coding and bug hunting. Codex on the other hand understands ActivityPub better and was able to point out, that I was not sending tags and mentions correctly.

Finally, I let claude write the scripts to use Claude to create meaningful commits. After doing so, I had to manually fix so many things, I was cured of my vibe coding curiosity.