Skill Updater

Skill Updater — Self-Assessment & Auto-Update (Claude Code)

When to Trigger

• "update your skills" / "refresh skills" / "skill audit"
• "are your skills up to date" / "check your skills"
• "tune yourself up" / "make yourself better"

Workflow

Phase 1 — Inventory

1. List all installed skills: ls ~/.claude/skills/
2. Read CLAUDE.md files: ~/.claude/CLAUDE.md, project CLAUDE.md, nested
3. Scan recent git history: git log --oneline -20 for active workstreams

Phase 2 — Audit Each Skill

For every skill directory:

1. Read the SKILL.md — check for:

   • Stale dates or version references
   • Overlap with other skills
   • Missing coverage for known active projects
   • Description quality (will it trigger reliably?)
   • References to claude.ai-only tools (Note in response, Grep/Read project memory, Read session transcripts) — these don't exist in Claude Code

2. Read reference files — check for:

   • Outdated data (prices, versions, API endpoints, tax law)
   • Missing reference files that should exist

3. Freshness checks (web search only when needed):

   • token-miser: Compare references/model-pricing.md header date. If >90 days, web search.
   • cpa-tax-specialist: Check for major tax law changes if >12 months old.
   • crypto-monero-wizard: Check monerod/P2Pool/XMRig versions if >6 months old.
   • ai-cluster-knowledge: Compare documented hardware against current memory files.

Phase 3 — Gap Analysis

• Active projects without skills — check memory files and CLAUDE.md for gaps
• CLAUDE.md gaps — conventions in skills not reflected in CLAUDE.md
• Skill overlap — two skills covering the same ground
• Missing conventions — patterns in code/commits not captured anywhere
• Stale CLAUDE.md entries — rules referencing removed files or outdated versions
• Cross-config contradictions — invoke config-auditor skill for hardware specs, repo counts, port assignments, collection prefixes, project phases. SSH to verify hardware before changing specs.

Phase 4 — Report & Propose

Terse tables and bullet points:

1. What's current — no changes needed
2. What's stale — specific data points that are wrong
3. What's missing — skills or entries that should exist
4. What overlaps — skills to consolidate
5. Proposed actions — ranked by impact

Execute non-breaking fixes immediately. Only ask about ambiguous items.

Phase 5 — Execute

1. Skill updates — Edit SKILL.md and reference files in place
2. New skills — Create directory + SKILL.md in ~/.claude/skills/
3. Reference updates — Web search for current data, rewrite reference file
4. CLAUDE.md updates — Edit the appropriate CLAUDE.md
5. Sync to repo — Run collect.sh to update claude-config repo

Frugality

• Don't web-search every skill — only time-sensitive reference data
• Don't re-read skills already in the current session
• Only propose actions with clear ROI
• Batch file edits

Output Format

Tables and bullet points, not prose. Flag issues, propose fixes, execute on approval.