Skill Cleaner

Skill Cleaner

Audit and trim your AI agent's skill prompt budget. The analyzer mirrors the same rendering rules agents use: 2% budget from context window, ceil(utf8 / 4) token cost, then proportional description truncation or skill omission.

Quick start

node --experimental-strip-types scripts/skill-cleaner.ts --months 3

Scans all skill roots (~/.codex/skills, ~/.config/opencode/skills, ~/.claude/skills, project .agents/skills/, and plugins) for the last 3 months, then prints:

• Skill Budget — context window, 2% budget, token costs, pre-budget full-list pressure, remaining budget
• Description candidates — long descriptions where trimming saves budget
• Duplicates — same name or near-identical body across roots
• Unused candidates — no $skill mention or SKILL.md read in recent logs
• Root summary — origin and enabled/disabled status per root

Workflows

1. Full audit with deep log scanning

node --experimental-strip-types scripts/skill-cleaner.ts --months 6 --max-log-mb 800 --deep-logs

Scans archived Codex sessions, OpenClaw, Clawd, and Claude Code task JSONs. Use quarterly or when budget pressure is high.

2. Quick budget-only check

node --experimental-strip-types scripts/skill-cleaner.ts --no-logs

Skips session log scanning — only analyzes SKILL.md files, descriptions, and config. Use for a fast budget check without heavy I/O.

3. Audit a custom skill root

node --experimental-strip-types scripts/skill-cleaner.ts --root ~/Dropbox/boxd/skills --no-logs

Scans additional skill directories beyond the standard roots.

4. Budget-constrained audit

node --experimental-strip-types scripts/skill-cleaner.ts --context-tokens 272000 --budget-percent 2 --no-logs

Override context window (default: GPT-5.5 at 272K) and budget percentage (default: 2%).

Reading the report

Read sections in this order:

Before deleting or editing

• Verify the kept copy loads correctly before deleting the duplicate.
• Prefer deleting repo-local or agent-scripts duplicates when agent built-ins cover them.
• Keep repo-local maintainer skills that encode project policy, domain language, or live operations.
• Preserve trigger nouns in descriptions: product, tool, action, object — removing these breaks auto-triggering.

Analyzer behavior

The script mirrors the agent's model-visible line shape (- name: description (file: path)) and applies the same rendering rules from render.rs:

• YAML frontmatter only; default name = parent directory
• Budget: 2% of context_window; token cost = ceil(utf8_bytes / 4)
• Description rendering: full descriptions → equal truncation → omitted to minimum lines
• Reads ~/.codex/models_cache.json for context_window (fallback: 272K)
• Scans default roots plus --root extras
• Dedupes by realpath — symlinked roots don't create false duplicates
• For duplicate names: reports description/body similarity; suggests deletion only when bodies are near copies
• Log scanning: ~/.codex/history.jsonl, ~/.codex/sessions/, ~/.config/opencode/sessions/, ~/.claude/projects/ by default

All flags

--months <n>         Look back N months for usage (default: 3)
--no-logs            Skip log scanning (budget-only check)
--deep-logs          Scan archived sessions, OpenClaw, Clawd logs
--max-log-mb <n>     Max log bytes to scan (default: 300)
--model <name>       Model name for context window (default: gpt-5.5)
--budget-percent <n> Skills budget percent (default: 2)
--context-tokens <n> Override context window size
--chars-per-token <n> UTF-8 bytes per token (default: 4)
--all                Include disabled skills
--root <path>        Add extra skill root (repeatable)
--json               Output JSON instead of report
--help               This help

Output policy

• Suggest first — print findings and recommendations. Do NOT delete or edit without confirmation.
• When asked to apply: make small grouped commits per action (descriptions, deletes, config disables).
• Do not delete ignored/untracked skill directories without naming the destination and confirming they are disposable.