Skill Overlap Audit

A library that grows by accretion accumulates near-duplicates: two skills covering the same ground, written at different times, never reconciled. This finds them.

Procedure

1. Extract name + one-line description for every skill in the library (from its README listing or its SKILL.md frontmatter directly).

2. Group candidates by textual similarity, not just exact match. Two skills can have slightly different wording but describe the identical thing (e.g. both saying "development conventions for the same project").

3. For each flagged pair, determine which is canonical, don't guess:

grep -rl "<skill-name>" . --include="*.md"   # who else references it back

The one other skills actually point to as their source/foundation is almost always the one to keep. An orphaned duplicate that nothing else references is the safer one to retire.

4. Diff the actual file content of flagged pairs before recommending removal. A near-identical description doesn't always mean near-identical content. If the bodies genuinely diverge in substance (not just wording), they may be intentionally distinct variants, not true duplicates. Flag those for a human judgment call rather than auto-recommending removal.

5. Check for explicit self-marked deprecation too. A skill whose own description says "deprecated, superseded by X" is an unambiguous removal candidate regardless of textual similarity to anything else.

6. Report before removing. This is an audit: surface the findings (candidate pairs, which one is canonical, confidence level) and let the removal happen as a separate, explicit action.

Real precedent

This exact method found two genuine removals in this project's own library: a skill explicitly self-labeled "deprecated, superseded by" a newer version, and a 95%+-identical duplicate of another skill that 16 other skills referenced back to. The duplicate itself was referenced by nothing.