Many projects already contain a huge amount of knowledge in their README, issues, docs, and merged PRs — but when using Claude Code, this knowledge still has to be manually distilled into a CLAUDE.md file.

I’m experimenting with the idea that an agent should be able to generate and improve that file automatically.

Auto-improvement explores a workflow where an agent: 1. Investigates a repository and generates an initial CLAUDE.md 2. Reads project context (README, issues, docs) 3. Time-travels to before a merged PR 4. Attempts to solve the same problem 5. Compares its solution to the human PR 6. Updates its own instructions (CLAUDE.md / rules / skills) based on the diff

The goal isn’t better code completion, but agents that adapt to: - A project’s conventions - Domain-specific decisions - Past mistakes captured in real PRs

This is early and opinionated — more a design exploration than a finished product. I’m especially interested in feedback on: - Whether this feedback loop makes sense - Failure modes you see - Similar systems or prior art I might be missing

Repo: https://github.com/Polandia94/auto-improvement

Thanks!