Find low-use Agent Skills by combining local session references, evidence coverage, install age, duplicate locations, source trust, and workflow purpose. A skill with zero matched evidence belongs in a review queue. Confirm its role, archive it through a recoverable ledger, and keep the archive through an observation period.
Why a usage count needs context
Local logs provide valuable first-hand evidence: Codex reading a SKILL.md, an explicit skill load, or a Claude Skill tool invocation. That evidence only represents files available to the scanner. Old logs may have been deleted, large logs may exceed the scan limit, and some tools store history elsewhere.
The strongest candidate for review combines several signals: zero recent evidence, a clear replacement, old install or modification dates, duplicated instructions, and low personal value. Any single signal carries less confidence.
Build the review queue
$ npx github:Ryan-yang125/skill-manager audit- Check evidence coverage. Record which Codex and Claude session roots exist and the number of logs scanned.
- Filter for zero or stale evidence. Use the result as a candidate list.
- Compare duplicate locations. Keep the canonical source and review shadowed copies.
- Read the skill. Rare recovery, security, travel, and annual workflows can have long quiet periods.
- Confirm the archive set. Record the reason for each selection.
Use a four-signal decision
| Evidence | Stronger review signal | Keep signal |
|---|---|---|
| Usage | Zero matches across broad log coverage | Recent, traceable session references |
| Overlap | Same job and content in another trusted path | Unique instructions or agent-specific behavior |
| Source | Unknown or abandoned package | Maintained source with a clear update route |
| Purpose | Experiment ended and replacement confirmed | Defined rare or critical workflow |
Coverage rule: “No evidence found” accurately describes the scan result. Keep that wording in reports so future readers can see the boundary of the conclusion.
Archive, observe, restore
A safe cleanup keeps the skill contents available. Skill Manager records the skill ID, original path, archive path, timestamp, content hashes, and operation status. Restore checks the original destination and protects an existing folder from overwrite.
- Archive a small confirmed batch.
- Run a fresh inventory and open your common agent workflows.
- Keep the archive through a useful observation period.
- Restore quickly when a workflow needs the skill again.
Unused skills FAQ
How can I tell whether an Agent Skill is unused?
Check local session references, evidence coverage, install age, duplicate copies, source trust, and the role in rare workflows. Zero evidence starts a review and preserves room for human judgment.
Which local logs does Skill Manager inspect?
It inspects recent and archived Codex sessions plus Claude project logs when those folders exist and files fit the configured scan limits.
What is the safest removal workflow?
Archive with a durable ledger, scan again, observe your workflows, and restore to the recorded original path when needed.
Audit before you archive
Generate the inventory, review evidence coverage, and make the final selection from a traceable report.
Run Skill Manager