skillprune

Every agent skill you install puts its name and description into the system prompt on every turn, forever. Installing is one click; nothing tells you what to remove.

The OSS ecosystem has solved discovery and install several times over — SkillDock (519★), skillfish (315★), skilld (308★). The other half of the loop is empty: every audit/eval/conflict tool on GitHub sits at 1–14 stars.

skillprune is that other half. It reads your real transcript history and tells you what to turn off.

uvx skillprune                   # run it, install nothing
pipx install skillprune          # or keep it around

skillprune                       # report
skillprune --json                # + skillprune.json
skillprune --selfcheck           # the four guards below, as assertions

No install needed either way — it is one stdlib-only file, so python3 skillprune.py works straight from a clone.

Example, on a real 291-skill install:

  291 skills installed · 39 have ever fired · 225 dead (77%) · 27 too new to judge
  ~26,043 tokens paid on every turn for skills you never use

How it decides

Nothing leaves the machine. No dependencies beyond the standard library.

Where it deliberately errs

A tool that says delete this has one unacceptable failure: naming something you actually use. Four guards, each pinned by an assertion in --selfcheck:

• Slash commands count. A skill invoked only as /foo never produces a Skill tool call. Counting just the tool call marks it dead.
• Hooks count. A plugin shipping a SessionStart hook runs every session with an invocation count of zero. Those are held back, never auto-disabled.
• Bare-name matching. Transcripts log <plugin>:<skill>, and that prefix often disagrees with the marketplace directory the skill was found in. Two plugins sharing a skill name will cross-credit — that errs toward in use, the only safe direction here.
• 14-day grace. Something installed yesterday hasn't had its chance. Tune GRACE_DAYS; skills under it are reported separately, never as dead.

Collision detection is Jaccard overlap on description word-sets (COLLIDE, default 0.35) — crude next to embeddings, but it needs no model and no network.

MIT.