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 --prune               # act on it, reversibly
skillprune --json                # + skillprune.json
skillprune --selfcheck           # the guards below, as assertions
skillprune --help                # all of the above

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:

  171 loaded    35 fired    123 dead    13 too new to judge
  125 more sit on disk, not loaded by any plugin — those cost nothing

  ████████████████████░░░░░░░░  72% of your skill surface has never fired

  ~15,673 tokens in every system prompt, for skills that have never fired
  71% of the ~22,169 tokens your skills cost you per turn

Removing what it finds

skillprune --prune prints a plan, asks once, and applies it. Two levers, because they are two different things:

• Plugins where not one skill has ever fired → claude plugin disable. Reversible with enable. A plugin with even one live skill is never offered.
• Personal skills in ~/.claude/skills/ → moved, never deleted, to ~/.claude/.skillprune-trash/<timestamp>/, next to a generated restore.sh that puts every one of them back.

There is no per-skill disable for plugin skills — the plugin is the unit the tool manages, so that is the unit offered.

It follows symlinks, so the agent-neutral layout — one shared ~/.agents/skills tree linked into each tool's config — is counted rather than silently skipped, and skills a plugin ships under .claude/skills/ or .openclaw/skills/ are found too. A project's own ./.claude/skills is included when you run it there. A shared tree that is not linked into a config dir is deliberately ignored: nothing loads it, so it costs nothing.

Set CLAUDE_CONFIG_DIR if your Claude Code config lives somewhere other than ~/.claude. With nothing to audit it says so, rather than reporting 0%.

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. Five 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.
• On disk is not loaded. A marketplace you added and a plugin you later uninstalled both leave complete SKILL.md trees behind, and neither costs a single prompt token. Only what installed_plugins.json actually points at is counted. Skipping this inflated the count on the author's own machine from a true 171 to 291.
• 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.