See which Claude skills you actually use — and what to cut.
Most people accumulate dozens of skills and plugins. Every enabled one puts its name and description into context every session, whether you ever use it or not. skill-usage reads your local Claude transcripts, shows you the real numbers, and recommends what to keep, modernize, turn off, or prune.
It answers two questions you otherwise can't see:
- How often is each skill used, and what does it cost? — usage counts, last-used date, and token cost across your global, project, and plugin skills.
- What should I do about each one? — a Keep / Modernize / Turn off / Prune verdict per skill, with a one-line reason.
It runs entirely on your machine. It reads ~/.claude transcripts locally and writes a local HTML file. Nothing is uploaded. There is no server.
- Always-on tax — a skill's name + description sit in context every session, paid even if never used. This scales with the number of enabled skills; it's what turning a skill off reduces.
- On-invoke cost — the full skill body, loaded once only when the skill actually fires.
The biggest savings come from the skills you never invoke.
The dashboard: stat cards, an always-on token-tax donut, top token consumers, a Keep / Modernize / Turn off / Prune action strip, a bucket board, and a sortable table.
As a Claude Code plugin (recommended):
claude plugin marketplace add paultaki/claude-skill-usageThen, inside Claude Code:
/plugin install skill-usage@skill-usage
It's now available as /skill-usage. Update later with claude plugin marketplace update skill-usage.
As a standalone script (no Claude required for the usage view):
git clone https://github.com/paultaki/claude-skill-usage
python3 claude-skill-usage/skills/skill-usage/scripts/skill-usage.py --openManually (drop-in skill): copy skills/skill-usage/ into ~/.claude/skills/.
Fast, deterministic view (usage + token cost + a heuristic baseline verdict):
python3 skills/skill-usage/scripts/skill-usage.py --openFlags: --open open the dashboard · --prune print the never-used prune list · --rescan ignore the cache · --demo build a sample-data dashboard (no transcripts needed — handy for a hosted demo).
The first run parses all transcripts; later runs only re-parse new or changed files (cached on mtime+size), so they're near-instant. Token counts use tiktoken if installed, otherwise a chars/4 estimate (the dashboard labels which). pip install tiktoken for a closer number.
The usage view ships with quick heuristic verdicts. For real recommendations — graded against your own CLAUDE.md and Anthropic's current skill-authoring docs — run the skill in Claude Code and ask for the recommendation pass. It writes recommendations.json, which the dashboard renders. See skills/skill-usage/SKILL.md for the steps.
Everything is written to ~/.claude/skill-usage/:
dashboard.html— the dashboarddata/skill-usage.json— usage + tokens + lint + baselinedata/recommendations.json— verdicts from the recommendation pass
The dashboard is read-only. In Claude Code you can then say "modernize these / prune those / turn off the rest" and it will edit your own skills, archive-move prunes to ~/.claude/skills-archive/ (reversible), or print the /plugin steps for plugin skills. Nothing runs without you; archive, not delete, is the default.
MIT — see LICENSE.