ai coding agents were driving me crazy. same mistakes over and over, switching models and rewriting prompts didn't help much.
so I flipped the question if I can't fix the model, maybe the problem is my repo? to figure this out I dug into Claude Code's internals, went through 265 versions of its system prompt to see how it evolved, read a bunch of papers on instruction-following, and tore apart a lot of open source repos. turns out most of the problems are in repo-level details. like: cramming too many "IMPORTANT" keywords actually makes the model ignore them (Anthropic themselves cut it from 12 down to 4). files over 256KB silently fail to load. and if your .env isn't gitignored, the agent can just read all your secrets.
I turned all of this into AgentLint 33 checks covering findability, safety, instruction quality and more. every check is backed by a paper or real data, not vibes. it scores your repo, shows whats wrong, and auto-fixes some of it. currently a Claude Code plugin, but its all bash + jq under the hood. MIT licensed. feedback welcome Github https://github.com/0xmariowu/Age...
