playbook/brooks-lint/.claude/skills/new-skill/SKILL.md

2.8 KiB
Raw Permalink Blame History

name description disable-model-invocation
new-skill Scaffold a new brooks-lint analysis skill so it passes `npm run validate` and `npm run evals` on the first try — generates skills/{name}/SKILL.md (with the mandatory "Do NOT trigger for:" clause and a Process section citing guide step ranges) plus skills/{name}/{name}-guide.md (sequentially numbered steps), then appends paired eval scenarios. Triggers when the maintainer asks to "add a new skill", "scaffold a skill", or "create a brooks-lint mode". Do NOT trigger for: editing an existing skill's content, adding a single eval to an existing skill, or authoring skills for some other plugin. true

brooks-lint — New Skill Scaffold

Skill name comes from $ARGUMENTS (kebab-case, e.g. brooks-security). If empty, ask for the name and a one-line purpose first.

Follow the repo's "Adding a New Skill" checklist exactly:

  1. Create skills/{name}/SKILL.md with frontmatter — name, a description that ends with a Do NOT trigger for: clause (validate-enforced for shipped skills; omitting it causes false triggering), and a Process section of 36 bullets that cite the guide's step ranges inline (e.g. Scan risks (Steps 16 of the guide)). Mirror the structure of an existing skill such as skills/brooks-review/SKILL.md (Setup → Process → Mode line).
  2. Create skills/{name}/{name}-guide.md with sequentially numbered steps — no gaps, no duplicates. Sub-steps like Step 2a are allowed. The guide owns the detailed steps; the SKILL.md Process is just an orientation skeleton.
  3. Wire the framework. The new SKILL.md Setup section must Read the relevant _shared/ files (common.md for Iron Law + Report Template, plus decay-risks.md / test-decay-risks.md as applicable) — _shared/ is NOT auto-loaded.
  4. Add eval coverage to evals/evals.json: append with the next sequential id at least one happy-path scenario (with the relevant risk code in expected_output) and at least one false-positive scenario flagged no_risk_codes: true. Each scenario needs id, name, prompt, expected_output, mode, files.
  5. Validate. npm run validate (structure + step continuity + Process-section presence) and npm run evals (eval schema). Fix until both pass.
  6. Local-test. cp -r skills/* ~/.claude/skills/brooks-lint/ (if you've symlinked ~/.claude/skills/brooks-lint to the repo, skip this — edits are already live), trigger the new skill in a Claude session, verify the Iron Law output, then restore the marketplace copy: /plugin marketplace update/plugin install brooks-lint@brooks-lint-marketplace.

Do NOT register a slash command by hand — short forms are auto-installed by the session-start hook. Report the files created and the validate/evals results.