Audit, track usage, and manage your AI coding skills - 7 focused skills, zero dependencies.

Works with Claude Code and OpenAI Codex.

A plugin that keeps your skills ecosystem clean, organized, and healthy. Automatically detects and scans skills from both Claude Code (~/.claude/skills/) and OpenAI Codex (~/.agents/skills/).

Correctness release — bundled five fixes that resolve real data-loss and noise problems found in the wild.

• fix.sh --apply is safe again — recognizes nested metadata.version (the canonical layout). Previously would have injected duplicate top-level version: lines into every modern skill.
• Name-collision detection — janitor-report now flags skills with the same name living at different real paths (the situation that confuses skill triggering), not just description-similarity overlaps.
• Symlink shadows no longer reported as duplicates — the same physical SKILL.md reachable from ~/.claude/skills/ and ~/.agents/skills/ is counted once.
• tokencost reflects actual cost — no more double-counted symlinks inflating the wasted-budget number.
• Plugin info populated — installed_plugins.json parser updated for Claude Code v2 schema (was always reporting 0 plugins).

Plus prior unreleased lint fixes (Windows compatibility, multiline descriptions, support-folder false-positives).

• Cross-platform — works with both Claude Code and OpenAI Codex
• Pre-install overlap check — /janitor-precheck checks if a new skill duplicates existing ones before installing
• Context window token cost — /janitor-tokens shows per-skill token consumption
• Consolidated from 9 to 7 commands — fewer commands, same coverage

Plugin install (recommended):

/plugin marketplace add khendzel/skills-janitor
/plugin install skills-janitor

Or clone directly:

git clone https://github.com/khendzel/skills-janitor ~/.claude/skills/skills-janitor

Each skill has its own slash command with autocomplete:

/janitor-audit          -> full skill inventory
/janitor-report         -> health check (lint + duplicates + broken)
/janitor-usage          -> which skills you actually invoke
/janitor-tokens         -> context window cost per skill
/janitor-search         -> find skills on GitHub
/janitor-search --compare my-skill  -> market analysis vs alternatives
/janitor-precheck https://github.com/user/skill  -> check before installing
/janitor-fix            -> auto-fix (dry-run by default)
/janitor-fix --prune    -> find and remove broken skills

Or use natural language:

"audit my skills"
"which skills do I use?"
"how many tokens do my skills cost?"
"check this skill before installing"
"search for n8n skills"

Parses your conversation history to show which skills you invoke and which you never use:

=== Skills Janitor - Usage Report ===
Period: 2026-02-24 to 2026-03-23 (4 weeks)

Active skills: 4 / 36 (11%)
Unused skills: 32 (89%)
Most used: n8n-workflows (17 total)
Recommendation: Remove 32 unused skills

Shows how many context window tokens each skill consumes:

=== Skills Janitor - Context Window Cost ===
Budget: 200,000 tokens

  Skill                               Tokens  Budget  Used?
  marketing-copywriting                2,340   1.2%    yes
  marketing-seo-audit                  1,560   0.8%     NO
  ...

  Total token cost: 18,720 (9.4% of budget)
  Unused skill cost: 14,300 (7.2% of budget wasted)

Check if a new skill overlaps with existing ones before installing:

=== Skills Janitor - Pre-Install Check ===

  Checking: marketing-seo-v2
  Scanned 35 installed skills

  --- HIGH OVERLAP (likely duplicates) ---
    [72%] marketing-seo-audit (user)
         Shared: seo, audit, ranking, meta

  VERDICT: High overlap detected - likely duplicate

/janitor-report now flags two distinct kinds of duplicates separately:

=== Skills Janitor - Duplicate Detection ===

Total skill records: 73
Unique skill files (after symlink dedup): 60

--- Name Collisions ---
Found 2 skill name(s) at multiple distinct paths:

  baseline-ui
    [user] ~/.claude-account-personal/skills/baseline-ui
    [user] ~/.agents/skills/baseline-ui

  fixing-accessibility
    [user] ~/.claude-account-personal/skills/fixing-accessibility
    [user] ~/.agents/skills/fixing-accessibility

--- Description Overlap (Jaccard > 30%) ---
Found 3 potential overlap(s):

  [50%] janitor-audit <-> janitor-usage
       Scopes: user / user
       Shared keywords: show, skills

  [33%] n8n-code-javascript <-> n8n-code-python
       Scopes: user / user
       Shared keywords: code, input, json, node, nodes, syntax

Name collisions are exact-name conflicts at different real paths — the situation that makes Claude pick the wrong skill ambiguously. Description overlaps are semantic-similarity warnings for skills with overlapping triggers but different names.

Symlink shadows (the same physical SKILL.md reachable from both ~/.claude/skills/ and ~/.agents/skills/) are deduped via realpath and no longer counted as duplicates.

• Never deletes anything without explicit confirmation
• Never modifies plugin/marketplace skills
• Respects that some overlap is intentional
• Dry-run by default for all destructive operations

Skills Janitor auto-detects which platforms are installed and scans all of them.

• Bash, Python 3, curl
• No pip installs, no node modules

skills-janitor/
├── .claude-plugin/
│   └── marketplace.json      # Plugin manifest (7 skills)
├── skills/
│   ├── janitor-audit/SKILL.md
│   ├── janitor-report/SKILL.md
│   ├── janitor-fix/SKILL.md
│   ├── janitor-usage/SKILL.md
│   ├── janitor-tokens/SKILL.md
│   ├── janitor-search/SKILL.md
│   └── janitor-precheck/SKILL.md
├── scripts/                  # Shared bash+python scripts
├── demo.gif
├── LICENSE                   # MIT
└── README.md

If you used v1.0 (9 skills), here's what changed:

PRs welcome. Each skill is self-contained in skills/janitor-*/SKILL.md.

MIT