Skill Self Refiner

Use when an existing SKILL.md failed or drifted in a real run and must be patched from evidence. Not for blank-slate authoring or one-off trajectory packaging.

cassidyrice Updated

File contents

Skill Self-Refiner

Patch existing skills from runtime failure, not vibes.

When to Use

  • A loaded skill gave wrong steps, missing gates, or stale commands
  • Fresh-context dry run failed a checklist item
  • User corrected the agent mid-skill and the fix should stick

Not for: creating a skill with no prior file → writing-skills or trajectory-to-packaged-skill.
Not for: prose-only polish → punchy-insight-refiner.

Process

  1. Load current SKILL.md + failure evidence (log, user correction, test output).
  2. Name the broken contract line (when_to_use / step N / NEVER / produces).
  3. Patch minimum text that would have prevented the failure.
  4. Add or tighten a Verification bullet that catches the same miss.
  5. Write a 3–5 item regression checklist; run it once in fresh wording.
  6. Record diff summary (before → after) in the response.

NEVER

  • Rewrite the whole skill when one step failed
  • Invent new capabilities not proven in the failure run
  • Drop produces / requires during polish

Collision parents

trajectory-to-packaged-skill, writing-skills, punchy-insight-refiner

cassidyrice/ai-skills-drop/tree/main/skills/skill-self-refiner commit 63a2dbd27e

Frequently asked questions

npx skillmds@latest add cassidyrice/skill-self-refiner