skills-audit

# /skills-audit — keep the toolkit from rotting Skills degrade in a specific way: they accumulate. Every installed skill's `description` is resident in context at all times, so the more there are, the less of each one survives — and the model picks by matching a request against exactly those descriptions. A toolkit that grows without pruning stops selecting correctly, and nothing announces it. **This skill never modifies anything.** It reports, and proposes. Removals are yours to approve. **This is not `/doctor`, and it is not an eval.** Three different things: | | Reads | Can see | |---|---|---| | `/doctor` | usage logs and settings | which skills were *never invoked*, slow hooks, CLAUDE.md duplication. **Run it too** — this skill cannot see usage. | | this skill | the files | over-constraint, budget, overlapping triggers, Cursor incompatibility | | an eval | with-skill versus without-skill runs | whether a skill actually *helps* | Nothing here measures whether a skill works. Say so when reporting, rather than letting a clean audit read as a clean bill of health. ## Preconditions | Condition | If unmet | |---|---| | `~/.agents/skills` exists, or a path was given | Stop and report it | | The dotagents checkout is locatable (`~/.claude/.dotagents-managed.json` → `repo`) | Continue, but skip the checks that need `verify-skills.sh` and say so | ## Position in the workflow | Upstream | This skill | Downstream | |---|---|---| | about to add a skill, or a quarterly clean-u...