This repository is an installable Agent Skill, not a conventional software project: the product is SKILL.md, a long structured instruction document that turns a skills-compatible AI agent (Claude Code, Codex, Cursor, etc.) into an autonomous optimizer for other SKILL.md files. It implements an autoresearch-inspired ratchet loop: Phase 0.5 designs per-skill test prompts, Phase 1 baselines target skills on a 100-point 9-dimension weighted rubric (structure 59 / effectiveness 35 / meta-skill 6) using independent judge sub-agents, Phase 2 hill-climbs one rubric dimension per round with git commit/revert as the ratchet plus a mandatory human checkpoint after each skill, and Phase 3 produces a regression report and a generated HTML-to-PNG result card. Distribution is `npx skills add alchaincyf/darwin-skill` or manual copy into a skills directory; the runtime is whatever agent interprets the instructions. The most valuable engineering here is meta-methodological and applies far beyond skill optimization. The current protocol (v2.1) ditched absolute-score keep/revert for paired same-judge comparison with odd-N majority voting, citing measured ±8-point LLM-judge scale drift that within-judge comparison cancels out — the same pairwise-preference reasoning behind RLHF. It also defines a weighted-gap formula (weight × (10 − score) / 10) so low-weight rubric dimensions can't fake progress, flags correlated dimension clusters (dim 2/3/4), enforces an 8-item anti-pattern blacklist derived from real past failures (e.g., 'same AI edits and scores', 'git reset --hard as rollback'), alarms on dry-run ratios >30%, and gates skills on multi-runtime neutrality. references/skilllens-evidence.md documents a genuinely well-designed controlled study (5 counterbalanced blind judges, 4 degradation classes) validating that the rubric catches gross regressions. Supporting assets are a mixed bag. scripts/screenshot.mjs is a genuinely reusable headless-screenshot utility with a robust cross-environment Playwright loader; templates/result-card.html (plus dark/white variants) is a polished, self-contained themed HTML template with data-field placeholders. But showcase.html and docs/index.html are near-duplicates, the result-card templates still render the legacy 8 dimensions against the current 9-dimension rubric (a known, documented drift), the README's claims of Microsoft Research integration and the cited arXiv papers (SkillLens / SkillOpt) cannot be verified from the repository itself, and the SKILL body is Chinese-only and deeply tied to the author's ecosystem. Repository health is good: MIT-licensed, active maintenance (last commit 2026-09-18), minimal dependency surface, no security alerts, and the repo practices what it publishes — its own evolution (86.05 → 92.7 self-score) is logged with unusual honesty, including the admission that early optimization runs were 67% dry-run with zero reverts before the methodology was fixed.