Purpose

Detect duplicates, enforce category consistency, flag deprecation candidates, verify discoverability, and maintain catalog consistency across an entire skill library. Produces structured curation reports with prioritized action items.

When to use

• Library has grown past ~20 skills and overlaps or inconsistencies have appeared
• Before a major release or after a bulk import
• Periodic maintenance pass (monthly for active libraries)
• User asks to "audit the library", "clean up skills", or "find duplicate skills"

When NOT to use

• Improving or refining a single skill →

  skill-improver

• Creating a new skill from scratch →

  skill-creator

• Splitting a single skill into focused variants →

  skill-variant-splitting

• Promoting, deprecating, or archiving individual skills through lifecycle gates →

  skill-lifecycle-management

Procedure

1. Build inventory

• List every skill directory at the repo root that contains a

  SKILL.md

• For each skill record: name (directory name), category (inferred from pipeline position in

  # Next steps

  cross-references, e.g., creation-pipeline vs. improvement-pipeline), last-modified date (from

  git log -1 --format=%cI -- <skill-dir>

  )
• Infer status from location: root directories → active,

  archive/

  → archived,

  corpus/

  → test fixture
• Count skills per inferred category; flag uncategorized or miscategorized entries

2. Detect duplicates and overlaps

• Extract action signatures: For each skill, extract the first verb+object phrase from the description (e.g., "Audit a skill library" →

  audit library

  , "Compare skill variants" →

  compare variants

  ). This is the skill's action signature.
• Group by action signature: Skills with the same or synonymous action signature (e.g.,

  audit library

  ≈

  review catalog

  ) are potential duplicates. Compare trigger phrases within each group — if >50% of one skill's trigger phrases also appear in or paraphrase the other, flag as duplicate.
• Check cross-references: For skills not grouped by action signature, inspect "Do not use for" sections. Mutual cross-references (A says "not for X, use B" and B says "not for Y, use A") suggest related scopes that may overlap or have ambiguous boundaries.
• Flag identical names at different paths.
• For each flagged pair, recommend one of: merge, differentiate (rewrite boundaries), or keep (with rationale).

3. Audit categories

• Verify each skill's category matches its actual function (read the procedure, not just the name)
• Singleton categories (1 skill) → review whether the skill fits naturally in an existing category. Do not merge categories that represent distinct capability areas solely based on count
• Categories with > 15 skills → propose a split axis

4. Check discoverability

• Does each description start with an action verb?
• Are negative boundaries present and naming the correct neighbor skills?
• Would a user with a realistic task phrase find this skill via keyword match?

Flag concrete defects using these thresholds:

• Too terse: Description under 20 words → insufficient for reliable routing. Recommend expanding to at least one sentence with verb, scope, and context.
• No trigger phrases: Description lacks quoted example phrases (e.g.,

  "audit the library"

  ) → routing relies entirely on keyword overlap, which is fragile. Recommend adding 2–3 realistic trigger phrases.
• Weak boundaries: "Do not use" section names fewer than 2 alternative skills → the skill's scope edges are undefined. Check the catalog for the most likely confused neighbors and add them.

5. Flag deprecation candidates

• Superseded by a newer skill with the same coverage
• Targets a tool or framework no longer in the stack
• If usage metrics exist, flag skills with zero invocations over the review window

6. Compile report

Output the curation report using the structure in Output contract below.

Output contract

The report MUST contain all eight sections. Omit rows only when a section has zero findings; keep the heading with "None found."

## Catalog Curation Report

### Inventory
- Total active skills: <N> (root directories with SKILL.md)
- Archived: <N> (in archive/)
- Test fixtures: <N> (in corpus/)
- Inferred categories: <list each category with skill count>

### Description Quality
| Skill | Word count | Starts with verb? | Has trigger phrases? | Issues |
|-------|-----------|-------------------|---------------------|--------|
| ...   | <N>       | yes/no            | yes/no              | <details> |

### Cross-Reference Graph
| Skill | References out (Next steps) | Referenced by | Boundary clarity |
|-------|-----------------------------|---------------|-----------------|
| ...   | <list>                      | <list>        | clear / ambiguous / missing |

### Duplicates / Overlaps
| Skill A | Skill B | Overlap evidence | Recommendation |
|---------|---------|-----------------|----------------|
| ...     | ...     | <shared trigger phrases or purpose text> | merge / differentiate / keep |

### Naming Convention Audit
| Skill | Convention | Issue |
|-------|-----------|-------|
| ...   | kebab-case / other | <details if non-conforming> |

### Gap Analysis
| Pipeline role | Expected skill | Status |
|--------------|---------------|--------|
| ...          | ...           | present / missing / weak coverage |

### Deprecation Candidates
| Skill | Reason | Replacement |
|-------|--------|-------------|
| ...   | ...    | ...         |

### Prioritized Actions
1. **[High]** <action> — <reason>
2. **[Medium]** <action> — <reason>
3. **[Low]** <action> — <reason>

Failure handling

• Cannot compare descriptions meaningfully (e.g., descriptions are one-word stubs): report the skill as "unassessable — description too short to evaluate" and recommend a description rewrite before the next curation pass.
• Category scheme is incoherent (no consistent axis): propose a replacement taxonomy with explicit grouping criteria and flag it as a blocking action before other category fixes.
• No usage metrics available: fall back to last-modified date and whether the skill's target tool/framework still exists in the stack.
• Findings exceed 30 action items: split into phases — Phase 1: duplicates and broken boundaries (high), Phase 2: category restructuring (medium), Phase 3: discoverability polish (low). Do not emit an unprioritized list.

Merge procedure

When the audit identifies true duplicates (recommendation: "merge"), execute:

1. Choose the survivor — the skill with broader coverage, better routing, or more support files
2. Inventory the absorbed skill — list all unique procedure steps, failure cases, and support files not present in the survivor
3. Merge content — add unique elements from the absorbed skill into the survivor. Do not duplicate content that already exists.
4. Update routing — rewrite the survivor's description to cover trigger phrases from both skills
5. Update cross-references — find all skills that reference the absorbed skill name and update to point to the survivor
6. Remove the absorbed skill — confirm with user before deleting:

   About to delete [absorbed-skill]/ directory. Proceed? [y/N]
   

   Do NOT delete without explicit user confirmation.
7. Update the inventory — update the README skill inventory to reflect the merge

Next steps

After curation:

• Execute merge recommendations → use the merge procedure above
• Fix discoverability issues →

  skill-trigger-optimization

• Deprecate identified candidates →

  skill-lifecycle-management