Hone Skill

hone-skill

Improve one existing skill: analyze it, apply targeted edits for the requested improvement, then lint and validate. Preserve what already works — refine, don't rewrite.

The request (target skill + what to improve): $ARGUMENTS

1. Identify the target and the goal

• Parse the arguments for the skill name and the improvement wanted. If either is missing or ambiguous, ask before changing anything.
• Locate the skill's SKILL.md. Prefer the repo source: search the working directory and its ancestors for the folder named after the skill, sitting beside CONVENTIONS.md, _template/, and scripts/. That parent is REPO_ROOT.
• Fall back to ~/.claude/skills/NAME/SKILL.md. If that path is a symlink, follow it and edit the real source so the repo stays the source of truth.
• If the skill isn't found or several match, list candidates and ask which one.

2. Analyze the current skill

• Read the full SKILL.md. Note the frontmatter (name, description, which keys), the single job it does, the body's structure and length, and any scripts//references/.
• If REPO_ROOT/CONVENTIONS.md exists, read it so edits stay compliant.
• Pin down exactly what the improvement needs and where it lands: triggering (description), a step, a new edge case, tighter wording, or scope.

3. Plan the change

• For anything beyond a trivial wording tweak, state a short plan — what you'll change and why, in a few lines — before editing.
• Keep the skill to one job. If the request really adds a second job, say so and propose a separate skill (hand off to /new-skill) instead of bloating this one.
• Make targeted edits; don't rewrite a working skill wholesale.

4. Apply the edits

Edit SKILL.md in place, respecting the conventions:

• description ≤ 1024 chars, leads with the use case + real trigger phrases, no XML/angle-bracket tags, no unquoted : (colon-space) that can break YAML.
• frontmatter uses only Claude Code-recognized keys — never add license, metadata, or compatibility (they silently hide the skill from the / menu).
• body under 500 lines, imperative and concise; move long material into references/ and link it.
• If the improvement is about triggering, sharpen the description's trigger phrases (skills under-trigger — lean slightly pushy).

5. Lint and validate

• Run REPO_ROOT/scripts/validate.sh NAME. Fix anything it flags and re-run until clean.
• If the validator isn't reachable, hand-check against the conventions checklist (name/folder match, allowed keys, description cap, body length, no secrets, one job).
• If the skill is symlinked into ~/.claude/skills, the edit is already live (body changes hot-reload). Otherwise remind the user to reinstall (scripts/install.sh NAME) and, for description/frontmatter changes, fully restart Claude Code to pick them up.

6. Report

• Summarize what changed (frontmatter vs. body), the lint result, and what's worth dogfooding next.
• Note anything you deliberately left out of scope.
• For trigger-accuracy tuning or benchmark-grade evals, hand off to the skill-creator skill — this skill stops at a solid, linted, improved draft.