Demo

Sample output — shows the full flow: Phase 1 (user skills) with confirmation prompt, then Phase 2 (plugins) with per-plugin confirmation. No real ~/.claude data shown.

Why this exists

Every Claude Code skill adds its description to the model context on every turn. You pay that token cost whether or not you use the skill. With a handful of skills the overhead is negligible — but as your library grows, unused skills quietly spend hundreds or thousands of tokens on every single prompt, pushing you toward your usage limit faster.

optimize helps you answer:

• Which skills have I actually used recently?
• Which can be hidden without uninstalling them?
• How many tokens am I wasting per turn?
• Which plugin skills appear completely unused?

Savings depend on your skill count and description sizes. Even a modest trim can meaningfully extend how long your daily limit lasts.

How it works

/skill-optimize runs a local audit over your Claude Code data:

1. Reads installed user skills from ~/.claude/skills/.
2. Reads enabled plugin skills from Claude Code plugin cache.
3. Scans recent session history in ~/.claude/projects/**/*.jsonl.
4. Counts direct slash-command usage and Skill tool invocations.
5. Groups user skills as:
   • dead: 0 uses in the lookback window, proposed as off
   • situational: 1-2 uses, proposed as name-only
   • kept: 3+ uses, left unchanged
6. Shows estimated token savings before changing anything.
7. Asks for confirmation before writing to ~/.claude/settings.json.

Plugin skills are reported separately because Claude Code does not expose them through skillOverrides in the same way as user skills. If a plugin appears completely unused, the workflow can ask whether to disable that plugin.

Install

curl -fsSL https://raw.githubusercontent.com/codeprakhar25/optimize/main/install.sh | bash

Restart Claude Code, then run:

/skill-optimize

Requirements

• Claude Code with skills enabled
• Python 3
• curl for the installer
• A Unix-like shell environment

No Python packages are required.

Usage

/skill-optimize

The workflow will:

• scan the last 60 days of Claude Code session history
• estimate token savings from hiding unused skill context
• show separate summaries for user skills and plugin skills
• ask whether to apply all recommendations, review them, skip, or cancel
• create a backup before writing any changes

Nothing is changed until you explicitly confirm.

What Gets Changed

For user skills, optimize updates skillOverrides in:

~/.claude/settings.json

Example:

{
  "skillOverrides": {
    "rarely-used-skill": "name-only",
    "unused-skill": "off"
  }
}

For fully unused plugins, optimize may update enabledPlugins only after a separate confirmation.

Modes

Undo

Every apply creates a timestamped backup of ~/.claude/settings.json.

Restore the latest backup:

python3 ~/.claude/skills/skill-optimize/scripts/restore.py

List available backups:

python3 ~/.claude/skills/skill-optimize/scripts/restore.py --list

Restore a specific backup:

python3 ~/.claude/skills/skill-optimize/scripts/restore.py --backup settings.json.bak-YYYY-MM-DDTHH:MM:SS

Restart Claude Code after restoring.

Manual Alternative

For one-off changes, you can use Claude Code's built-in skill picker:

1. Run /skills.
2. Highlight a skill.
3. Press Space to cycle between on, name-only, and off.
4. Press Enter to save.

Use the manual flow when you already know which skill to change. Use /skill-optimize when you want an audit based on usage history.

Safety

• Fully local: session history is read from disk and is not sent anywhere.
• Reversible: settings are backed up before every write.
• Reviewable: recommendations are shown before they are applied.
• Conservative defaults: rarely used skills are moved to name-only, not immediately disabled.
• Protected skills: skills referenced in ~/.claude/CLAUDE.md are kept.
• Write lock: concurrent runs are guarded by a lock file.

Risks and Limits

The audit is based on detectable usage — slash commands and Skill tool calls found in your session history. Review recommendations before applying.

Implicit usage may be missed. If Claude invoked a skill automatically (its description matched your prompt) without an explicit slash command, the audit counts it as zero. Keep borderline skills at name-only rather than off.

Fresh installs have thin history. If you've barely used Claude Code yet, everything looks unused. Wait a few days, then re-run.

Plugin disables are broader. Turning off a plugin removes all its skills and agents — not just one. Plugin changes are always shown separately and need their own confirmation.

Token savings are estimates. Calculated from description lengths, not Claude's internal tokenizer. Treat the number as a directional guide, not an exact figure.

Your prompts are read locally. The audit reads .jsonl session files on disk. Nothing is sent anywhere — all processing stays on your machine.

If a skill you need becomes hidden, restore from backup or re-enable it with /skills.

Uninstall

rm -rf ~/.claude/skills/skill-optimize

Uninstalling removes the skill-optimize command. It does not remove any skillOverrides already written to ~/.claude/settings.json; use the restore command first if you want to undo applied changes.

Development

Run the audit script directly:

python3 scripts/audit.py --days 60

Emit JSON:

python3 scripts/audit.py --days 60 --json

The project is intentionally small:

• skills/skill-optimize/SKILL.md: Claude Code workflow
• scripts/audit.py: usage audit and token estimate
• scripts/apply.py: settings patcher with backup and lockfile
• scripts/restore.py: backup restore helper
• scripts/record-demo.sh: regenerate the README demo GIF
• install.sh: curl installer

License

MIT