Hey, I got inspired by GPT Repo Loader and built aidev over the weekend, asking it to build a big chunk of itself as a meta-experiment.

I was curious to find the current limits of how far I can push this AI coding stuff. Here's an example of a prompt that I believe represents the very limit of what GPT-4 can handle:

> We're working on support for paths in only/include patterns. The match func has a bug: the dir argument is an absolute path, but we need a relPath, and we compute it incorrectly via filepath.Join. Instead, we need to compute a relative path against directory of .aidev file that this pattern came from. But that information is lost!

> Let's kill the idea of merging TreeConfigs completely. Instead, save a pointer to a parent TreeConfig loadConfig. Change ShouldIgnore to build and process a list of applicable TreeConfigs: builtin config, a list of configs for the directory, and override config.

...Overall, though, I did not enjoy using aidev for coding. Definitely did not save me time.

There is a part that I've enjoyed, though: having aidev write documentation. In a real project, you would definitely want to edit those yourself, but having a draft seems beneficial. The README of aidev is unedited AI output, so you can judge the quality yourself.