skill-monitor

Track which Claude Code skills (slash commands) you actually use — how often, with what, and in which project.

Claude Code can accumulate dozens — even hundreds — of skills, but there's no built-in way to answer "which of these do I actually use?" (see anthropics/claude-code#35319). skill-monitor fills that gap with two lightweight hooks and a small query CLI.

100% local. Nothing is ever sent anywhere.

Example

$ skill-stats
Total: 128 calls  |  log: ~/.claude/skill-monitor/skill-usage.jsonl

COUNT  SKILL          LAST_USED
42     commit         2026-06-16T17:26:03+0900
31     pull-request   2026-06-16T16:10:55+0900
18     worktree-list  2026-06-15T09:41:12+0900
12     new-feature    2026-06-14T11:02:48+0900

$ skill-stats recent 4
TIME                      SKILL          SRC    ARGS         PROJECT
2026-06-16T17:26:03+0900  commit         typed  fix login    my-app
2026-06-16T17:10:22+0900  worktree-list  tool   -            my-app
2026-06-16T16:10:55+0900  pull-request   typed  -            my-app
2026-06-15T09:41:12+0900  new-feature    typed  DEVWEB-1234  api

$ skill-stats source
By source (tool vs typed):

COUNT  SOURCE
80     typed
48     tool

Features

• 🪶 Two hooks, full coverage — captures both skills Claude invokes (the Skill tool) and slash commands you type yourself.
• 📊 Instant stats — per-skill totals, recents, today, per-project, and a tool-vs-typed breakdown.
• 🔒 Local only — one JSONL file on your disk; no network, no telemetry.
• 🧩 Zero config — install the plugin and the hooks activate automatically; no settings.json edits.
• 🪝 Non-blocking — hooks always exit 0 and never touch stdout, so they can't interfere with anything.

Requirements

• Claude Code with plugin support
• jq

Install

/plugin marketplace add yokihara/skill-monitor
/plugin install skill-monitor@skill-monitor

Start a new session — the hooks are active automatically.

Local test before publishing

/plugin marketplace add /absolute/path/to/skill-monitor
/plugin install skill-monitor@skill-monitor
claude plugin validate /absolute/path/to/skill-monitor

Usage

Inside Claude Code (slash command)

After installing, type the slash command right in Claude Code:

/skill-monitor:skill-stats            # per-skill totals
/skill-monitor:skill-stats recent 20  # last 20 invocations
/skill-monitor:skill-stats today      # today's usage
/skill-monitor:skill-stats source     # tool vs typed

From your terminal (CLI)

The CLI also ships at bin/skill-stats inside the plugin. Locate it and alias it in your shell rc:

find ~/.claude/plugins -name skill-stats -path '*skill-monitor*'   # locate it
alias skill-stats="/path/from/above"

skill-stats            # per-skill totals + last used (default)
skill-stats recent 20  # last 20 invocations (with source)
skill-stats today      # today's usage
skill-stats project    # per-project totals (or: project <name>)
skill-stats source     # tool vs typed breakdown
skill-stats raw        # print the resolved log path

How it works

Every invocation is appended as one JSON line. Two paths are covered, and they don't overlap (so there's no double counting):

You do this	Hook that fires	Skill name from	source
Claude invokes a skill via the Skill tool	PreToolUse (matcher Skill)	tool_input.skill	tool
You type a /command yourself	UserPromptExpansion	the command field	typed

A record looks like:

{"ts":"2026-06-16T17:26:03+0900","epoch":1781598363,"skill":"commit","args":"fix login","source":"typed","session":"…","cwd":"…","project":"my-app"}

Data & privacy

• Log file: ${CLAUDE_PLUGIN_DATA}/skill-usage.jsonl, falling back to ~/.claude/skill-monitor/skill-usage.jsonl.
• A pointer file at ~/.claude/skill-monitor/.logpath lets the CLI find the log without environment variables.
• Local only — no network calls. Records contain the skill name, args, timestamp, session id, and cwd/project. Delete the file any time to reset.

Uninstall

/plugin uninstall skill-monitor@skill-monitor
/plugin marketplace remove skill-monitor

Then optionally remove leftover data:

rm -rf ~/.claude/skill-monitor

Caveats

• The UserPromptExpansion event and its command field follow the Claude Code hooks documentation. If a future version changes them, the typed path falls back to parsing the first /token of the prompt.
• Args for typed commands are best-effort: the structured payload exposes the command name reliably, while arguments are parsed from the prompt.

Contributing

Issues and PRs welcome. Please keep the hooks non-blocking (always exit 0, no stdout) and the CLI dependency-light (POSIX shell + jq).

License

MIT