I've been using a `/feedback ...` command with claude code where I give it either positive or negative feedback about some action it just did, and it'll look through the session to make some educated guesses about why it did some thing - notably, checking for "there was guidance for this, but I didn't follow it", or "there was no guidance for this".

the outcome is usually a new or tweaked skill file.

it doesn't always fix the problem, but it's definitely been making some great improvements.

CodeRabbit in particular is gold here. I don't know what they do but it is far better at reviewing than any AI model I've seen. From the deep kinds of things it finds, I highly suspect they have a lot of agents routing code to extremely specialized subagents that can find subtle concurrency bugs, misuse of some deep APIs etc. I often have to do the architecture l/bug picture/how this fits into project vision review myself, but for finding actual bugs in code, or things that would be self evident from reading one file, it is extremely good.

That is actually a gold tip. Codex CLI is way less pleasant to use than Opus, but way better at finding bugs, so I combine them.

Codex is a sufficiently good reviewer I now let it review my hand-coded work too. It's a really, really good reviewer. I think I make this point often enough now that I suspect OpenAI should be paying me. Claude and Gemini will happily sign off work that just doesn't work, OpenAI is a beast at code-review.