darwin-skill

Why we included this project

Teams that maintain a growing library of agent skills for Claude Code or any SKILL.md-compatible assistant eventually hit the point where hand-editing instruction files and hoping they work stops scaling. darwin-skill turns that maintenance into a measured loop: each skill is scored against a nine-dimension rubric, improved along one dimension per round, and either kept or rolled back with git revert instead of a destructive reset. The human-in-the-loop layer is what makes it usable, since it stops after every optimization round and waits for your confirmation, so nothing changes without your go-ahead. Two details worth knowing: the agent that edits a skill is never the one that scores it, and the 2.0 design builds on Microsoft Research's SkillLens and SkillOpt papers, with Microsoft listing darwin-skill among SkillOpt's official integrations. If you want verifiable, repeatable skill improvements rather than vibes-based rewriting, this is a concrete workflow to run.