skill-doctor - SKILL.md Agent Skill

name: skill-doctor description: Use when a repo skill fails, produces errors, gives wrong instructions, or references stale information. Also use when a command from a skill returns "not found" or "unknown flag", when a skill references a file path that doesn't exist, when a skill's behavior doesn't match the codebase, or to list available skills, manage local overrides, check staleness, or optimize token usage and quality.

Skill Doctor

Diagnose issues with repository skills, manage local overrides, and optimize skill quality.

Scope: Skills in the current repository's .claude/skills/ directory only.

Routing

Parse $ARGUMENTS:

Pattern	Mode
`list`	list (all repo skills)
`info <skill>`	list (detail for one skill)
`feedback [skill] "<message>"`	feedback (quick note)
(empty) or `<skill-name>`	diagnose
`--optimize <skill>`	optimize
`--manage [skill]`	manage (status)
`clone/diff/sync/revert <skill>`	manage (lifecycle)
`report/gist/pr <skill>`	deliver
`resolve <skill>`	resolve

Unrecognized input: show this table and ask to rephrase. Then read ./reference.md for detailed procedures.

List

Dynamically discovers repo skills by reading .claude/skills/*/SKILL.md. Shows name, description, invocation, and override status. info <skill> reads the full SKILL.md to provide detailed usage guidance. All information comes from the skill files themselves — nothing is hardcoded. See reference.md for details.

Feedback

Lightweight path for quick notes. No investigation or quality gate — just scrub, preview, confirm, post. If skill name is omitted, ask which skill (or use the most recently invoked skill if known from conversation context). See reference.md for details.

Manage

Lifecycle: clone -> edit -> diff -> sync/revert -> pr/gist. Local overrides live at ~/.claude/skills/<name>/ with .skill-origin.json tracking.

Diagnose

Identify skill, validate it exists
Classify: invalid_command | wrong_assumption | stale_info | other
Investigate -- reproduce, root cause, determine fix, check related issues. Quality gate: confirmed problem + root cause + suggested fix
Duplicate check in Discussion thread. Match found: upvote, add context, or report anyway
Route: local override -> fix inline, stop. Repo -> offer clone-and-fix or report-only
Scrub PII, review with developer, deliver to GitHub Discussion + Slack

Optimize

Checks frontmatter, token efficiency (<500 words), structure quality, and staleness (file paths, commands, flags). Scored report with fixes. Local overrides fixed directly; repo versions cloned first.

Deliver & Resolve

Deliver: GitHub Discussion (primary, one thread per skill) + Slack #log-skill-feedback (optional notification). Check prerequisites first (gh auth, repo access). See reference.md.

Resolve: After fix PR merges, mark Discussion reports resolved, post Slack summary, offer to lock thread.

Common Mistakes

See reference.md for the full table. Key rules:

Never send reports without root cause + suggested fix
Always check Discussion thread for duplicates before reporting
Always run gh auth status before any delivery action
revert always shows diff and requires name confirmation

Override Warning

When ANY skill is invoked and ~/.claude/skills/<name>/.skill-origin.json exists, output before skill content:

SKILL OVERRIDE ACTIVE: <name>
  Using: ~/.claude/skills/<name>/ (local)
  Instead of: .claude/skills/<name>/ (repo)
  Cloned from: <sha> | Repo is N commits ahead
  Run /skill-doctor --manage <name> for options