Check it out HERE
I've been building with Claude Code and Cursor and kept hitting the same wall: every new session starts from zero. The agent has no idea what the repo is, what conventions we follow, or what we were just working on. I was manually pasting context into every session.
hey-recon fixes that with one command. It scans your repo — file structure, git history, package.json, entry points, recent commits — and calls an LLM to synthesize a CONTEXT.md that explains your codebase to any AI agent. It also installs a git hook so the file stays in sync automatically on every commit.
It outputs to three formats:
- CONTEXT.md (generic)
- CLAUDE.md (for Claude Code)
- .cursorrules (for Cursor)
Built it in a day because I needed it. MIT licensed. Would love feedback on whether the synthesized output is actually useful or if I'm capturing the wrong signals from the repo.
Top comments (1)
A living CONTEXT.md is a practical answer to session amnesia. The key is keeping it generated from real repo signals instead of turning it into a stale essay. File structure, recent commits, entry points, and conventions are exactly the data agents need first.