skills-analyzer

Inventory installed Claude Code skills and cross-reference them against your actual usage history.

What it does

Claude Code skills live in ~/.claude/skills/ (user), .claude/skills/ (project), and inside installed plugins. Over time you accumulate skills you've forgotten about, skills that conflict with each other, and skills you've never actually invoked. skills-analyzer surfaces all of it.

Inventory — every SKILL.md on disk, tagged by where it came from (user / project / plugin), with description length, body word count, and any structural defects (missing-name, dangling-ref?, orphan-resource, etc.).

Usage — how many times each skill was actually invoked, across how many sessions and projects, mined from the Skill tool-use records in ~/.claude/projects/*.jsonl.

Overlap detection — pairs of skills whose descriptions compete for the same triggers, ranked by overlap score.

Ghost detection — skill names that appear in your transcripts but no longer have a SKILL.md on disk (renamed, uninstalled, or never installed on this machine).

Plugin audit — installed plugin entries whose installPath no longer exists on disk.

Built-in skills shipped with the Claude Code CLI are not files and are never reported here.

Installation

pip install skills-analyzer

Requires Python 3.9+. No third-party dependencies.

Usage

skills-analyzer [options]

By default, output is JSON (suitable for piping into jq or a dashboard). Pass --text for a human-readable summary.

Options

Examples

# Quick human-readable summary
skills-analyzer --text

# Only look at the last 30 days of usage
skills-analyzer --text --since 30

# Inventory only, no transcript reading
skills-analyzer --no-usage --text

# JSON output piped to jq
skills-analyzer | jq '.skills[] | select(.problems | length > 0)'

# Scan a specific project's skills
skills-analyzer --project ~/work/myproject --text

JSON output shape

{
  "counts": {
    "skills": 12,
    "unused": 3,
    "with_problems": 1,
    "transcripts": 847,
    "transcripts_with_skill_calls": 204
  },
  "skills": [
    {
      "name": "my-skill",
      "source": "user",
      "path": "/Users/you/.claude/skills/my-skill/SKILL.md",
      "description": "...",
      "desc_chars": 120,
      "desc_long": false,
      "body_words": 1340,
      "body_over_ideal": false,
      "body_over_max": false,
      "resources": ["references/rubric.md"],
      "problems": [],
      "usage": {
        "invocations": 47,
        "sessions": 12,
        "projects": 3,
        "first": "2026-09-01",
        "last": "2026-10-28"
      }
    }
  ],
  "ghosts": [...],
  "overlaps": [...],
  "plugins": [...]
}

Text output example

12 skills on disk / 1 with structural problems / 3 never invoked
transcripts: 847 scanned, 204 contain Skill calls

--- inventory ---
skill                    source              desc   body  used        last
my-skill                 user                 120   1340    47  2026-10-28
code-review              user                 310   2100   182  2026-10-29
deploy                   project              88    540     0           -

--- structural problems (1) ---
old-skill:
    dangling-ref?: body cites references/rubric.md, which does not exist — line: '...'

--- description overlap (1) ---
 0.40  code-review <-> simplify
        shared: changes, correctness, diff, review

--- invoked but not installed (1) ---
plugin:deprecated-tool     3x  last 2026-09-15

Structural checks

The tool flags unambiguous defects in SKILL.md files:

• empty — SKILL.md has no content
• no-frontmatter — missing or unparseable YAML block
• missing-name — no name: field in frontmatter
• name-mismatch — name: field doesn't match the directory name
• missing-description — no description: field
• empty-body — frontmatter only, no instruction body
• dangling-ref? — body references a path under references/, scripts/, assets/, or examples/ that doesn't exist (flagged with ? because a skill documenting other skills may quote paths it doesn't own)
• orphan-resource — a file in a resource subdirectory is never mentioned in the body

Description length flags

The inventory table reports desc_long: true for descriptions over 500 characters. Skill descriptions are loaded into every Claude Code prompt for the duration of the session; very long descriptions increase token usage on every request.

Body word count is reported against the 1,500–2,000 word ideal and 5,000 word hard guidance from Claude Code's skill development documentation.

License

MIT