alchaincyf/darwin-skill

Darwin-Skill (达尔文.skill) is a v2.0 Agent Skill optimizer that applies machine-learning training loop concepts — specifically Andrej Karpathy's autoresearch ratchet mechanism — to the iterative improvement of SKILL.md files used by AI coding agents like Claude Code, Codex, and Trae. The system evaluates skills across a 9-dimension rubric (expanded from 8 in v1.0), incorporating three new dimensions directly from Microsoft Research's SkillLens paper: Failure Mechanism Encoding, Actionable Specificity, and High-Risk Action Blacklist. A key differentiator is its Human-in-the-Loop design with mandatory checkpoints, contrasting with SkillOpt's fully autonomous approach, since skill quality is considered more nuanced than a simple validation loss. The ratchet mechanism ensures scores only increase — any edit that doesn't improve the weighted score is automatically reverted via git. Notably, Microsoft's SkillOpt officially listed darwin-skill as an integration partner. The project is actively maintained (updated May 2026), installable via `npx skills add`, and has demonstrated measurable gains (e.g., huashu-gpt-image skill improved from 80.8 to 91.65). It is trending likely due to its timely alignment with emerging Microsoft Research papers and the rapidly growing Agent Skill ecosystem.