mcpField
S
skill-curator

skill-curator

github.com/wilbeibiAgent workflows278 installsUpdated 5d ago
Repository

About skill-curator

A skill for auditing collections of agent skills and identifying overlap, conflicting guidance, bloat, and internal defects. It provides a classification workflow and recommends fixes while limiting changes to verified issues.

Key features

  1. 01Inventory and cluster skills
  2. 02Classify overlap and defects
  3. 03Verify suspected issues in source files
  4. 04Recommend a fix before broad refactors
  5. 05Check references and content budgets

Use cases

  1. 01Reviewing a repository of skills
  2. 02Finding conflicting instructions
  3. 03Identifying redundant or oversized files
  4. 04Preparing scoped skill maintenance

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/wilbeibi/wilbeibi-skills --skill skill-curator

After installing, type/skill-curatorin Claude Code to confirm it loaded.

Bundled filesagents/openai.yaml

Details

From the project README

skill-curator

Workflow

  1. Inventory — for d in <root>/*/*/; do grep -E '^(name|description):' "$d/SKILL.md"; done
  2. Cluster — group by topic; flag pairs whose descriptions overlap or triggers could both fire.
  3. Read suspects — never classify from description or regex alone.
  4. Classify against taxonomy below.
  5. Surface findings + recommended fix — get direction before multi-file refactor; small isolated fixes can proceed.
  6. Fix and verify. Don't commit unless asked.

Taxonomy

  • Forked script — same script copied across skills. Diff for drift. One owner; the other points to it.
  • Contradiction — conflicting instructions for the same artifact. Pick owner; other defers.
  • Restatement — one skill re-derives another's content. Replace with a one-line pointer.
  • Bloat — frequently-loaded file carrying rare detail. Externalize to references/; keep quick-fixes inline.
  • Unused or over budget — a skill no session log has read, or descriptions that together exceed the harness's list budget (Codex: 2% of context, then equal truncation). steipete/agent-scripts skill-cleaner measures both for Codex; elsewhere, sum description bytes by hand.
  • Internal defect — duplicate headers/numbering, dead path refs, dead code.
  • Smell — a single skill that fails on its own. From a 238-skill survey (arXiv 2607.01456), the ones worth flagging: stepless workflow; option buffet (choices with no decision rule); rationalization loophole ("usually", "if needed"); buried gotchas (traps mid-paragraph instead of in Notes); confusing description (what, when, and keywords blurred into one sentence); time-sensitive claims; non-third-person description; XML or backslash paths in frontmatter.

Rules

  • Read every claimed defect before asserting it — heuristics misfire on code blocks and intentional parallel structures.
  • One owner per capability; the other references it by explicit path.
  • Progressive disclosure: description → SKILL.md (terse) → references/REFERENCE.md (deep). Terseness scales with load frequency.
  • Fix the defect only — no adjacent refactoring, no extra cross-links.

See also

  • write-skill — author one new skill.

Comments

Markdown supported