darwin-skill for Claude Code: score and improve your SKILL.md files

• Picks the skills to cover, makes a git branch and a results.tsv log

  Creates the branch auto-optimize/YYYYMMDD-HHMM and results.tsv if it is missing, and reads earlier results. If you are not in a git repo it asks whether to run git init or use file backups.

  Try: “Optimize all my skills.”

  Runs code · Phase 0: Set up

• Writes 2 to 3 typical prompts per skill to test-prompts.json

  You see the prompts and confirm them before scoring starts.

  Write · Phase 0.5: Test prompts

• Scores every skill, with sub-agents running each test prompt with and without the skill

  The absolute scores only rank which skill is weakest. It then stops and waits for you.

  Try: “Evaluate the quality of all my skills.”

  Runs code · Phase 1: Baseline scores

• One edit per round, committed, then judged; up to 3 rounds per skill

  3 judges read the old and new version in one call and vote better, worse or tie. A majority of worse triggers git revert, a new commit that undoes the edit. You approve each skill before the next.

  Try: “Optimize the skill called huashu-slides.”

  Runs code · Phase 2: Optimize loop

• Offers a full rewrite when two skills stop improving in round 1

  It uses git stash to keep the best version and runs only after you agree.

  Runs code · Phase 2.5: Exploratory rewrite

• Summarises experiments kept and reverted, and score changes

  Read · Phase 3: Report

• Nine scored areas; the live test is worth the most, 23 of 100 points

  In order: frontmatter 7, workflow 12, failure modes 12, checkpoints 6, specificity 18, resources 4, architecture 12, live test 23, do-not list 6.

  Rubric (9 parts, 100 points)

• Runs a grep for wording that ties a skill to one agent

  It greps SKILL.md and README.md for phrases such as "Claude Code skill" and fixes real hits first.

  Runs code · Runtime-neutrality scan

• Fills an HTML template and screenshots it to a PNG

  It runs node scripts/screenshot.mjs, with npx playwright screenshot as the fallback.

  Runs code · Result card

• Every 30 days, compares your clone with its origin and may tell you to git pull

  It reads .last-update-check, runs git ls-remote origin HEAD, writes the date, and leaves the update to you.

  Runs code · Update check

• Evaluate only, optimize one skill, or show results.tsv history

  Read · Other uses

Watch out for

• It edits and commits your other skills

  Each round edits a SKILL.md and commits it on an auto-optimize branch. The README says to run it from a git repo and to commit or stash your own edits first. github.com

• Test prompts run the skills you are testing

  Sub-agents execute each test prompt with the skill loaded. A skill that runs shell commands or calls services will do so during the test, so read the prompts it shows you before you confirm. github.com

• Result cards need Node and playwright

  The screenshot script launches Chromium through playwright and opens the PNG with your system viewer. Its error text tells you to install playwright-core and Chromium if they are missing. github.com

• It checks for updates against the git remote

  After 30 days it runs git ls-remote origin HEAD and writes the date to .last-update-check. It only tells you about a newer version, and the SKILL.md leaves updating to you. github.com

• It stops for your approval

  SKILL.md marks CHECKPOINT stops after the baseline scores, after each skill, and before an exploratory rewrite. Use them to review the diff. github.com