Haircut measures, prices, and safely trims the configuration overhead loaded by every coding agent on a machine.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install haircut
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install haircut using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is haircut?
haircut is an Openclaw Skills configuration auditing and cleanup tool for coding agents such as Claude Code, Codex, OpenCode, Cursor, Gemini, Copilot, Windsurf, Continue, Claude Desktop, and VS Code. It scans instruction files, memory, MCP servers, hooks, skills, commands, subagents, plugins, duplicates, and leftovers, then prices their token and latency impact using real transcript data where available.
The skill is designed for safe, evidence-based cleanup rather than blind deletion. It reports measured startup costs, proposes verdicts, asks for explicit decisions, moves approved items into a dated recovery bag, creates a receipt, and supports one-command restoration. As an Openclaw Skills workflow, haircut complements built-in checks such as /context, /doctor, and /skill-doctor by auditing every supported agent on the machine.
haircut Use Cases
- Investigate why a coding-agent session is slow or why context is full before the first prompt.
- Measure average and worst-case tokens consumed before the first word of a cold Claude Code session.
- Audit Codex rollout overhead, including MCP schema costs paid on every model call.
- Find unused MCP servers, skills, commands, hooks, plugins, subagents, or instruction files.
- Identify duplicate installations and project settings that shadow global configuration.
- Review the cost and latency of hooks that fire on every prompt.
- Audit CLAUDE.md, AGENTS.md, and MEMORY.md without modifying their contents.
- Decide which credential-bearing configuration items need rotation without exposing secrets.
- Safely park approved configuration with a manifest and restore it if the change causes problems.
- Use Openclaw Skills to create a measured, reversible configuration-maintenance workflow across multiple agents.
How haircut Works
- Measure: Run the read-only scanner to inspect supported agent configuration, transcripts, usage history, and optional MCP schema costs.
- Read the report: Review the concise
REPORT.mdsummary rather than loading the fullreport.json; optionally open the self-containedreport.htmlinterface. - Explain findings: Present the before-first-word token figure, the three largest savings opportunities, and counts grouped by verdict.
- Resolve decisions: Ask the user about
ASKandROTATEitems in small, grouped rounds. Unanswered items remainlaterand are not changed. - Write decisions: Save approved choices to
decisions.jsonin the run folder, using item IDs fromdecisions.template.json. - Confirm the plan: Show the exact list of items to be parked, including what will stop working, and wait for explicit approval.
- Dry-run trimming: Execute
trim --dry-runand reconcile any difference between the proposed plan and the user-approved list. - Apply changes: Run the real trim command only after confirmation. Approved items are moved, not deleted.
- Provide recovery: Print the receipt, point to
MANIFEST.jsonl, and provide the exactrestorecommand for the generated bag. - Start fresh: Explain that the current session already paid its startup cost; the user should begin a new session and rerun haircut to measure the improvement.
haircut Setup
haircut is invoked through the installed skill or plugin and uses ${CLAUDE_PLUGIN_ROOT} as its runtime root. The provided skill documentation does not specify a separate package-install command; ensure the skill is available to the coding agent before running these commands.
Run a default, read-only audit:
node "${CLAUDE_PLUGIN_ROOT}/bin/haircut.mjs" --json
Optional measurement controls:
node "${CLAUDE_PLUGIN_ROOT}/bin/haircut.mjs" --json --days 30
node "${CLAUDE_PLUGIN_ROOT}/bin/haircut.mjs" --json --agents claude-code,codex
node "${CLAUDE_PLUGIN_ROOT}/bin/haircut.mjs" --json --probe
Use --probe only when the user explicitly wants MCP tool-schema costs measured, because it starts configured MCP servers. Codex transcript scans may take a minute on machines with extensive history.
Open the browser-friendly report when the result is large:
node "${CLAUDE_PLUGIN_ROOT}/bin/haircut.mjs" report --open
After the user approves a decisions file, preview and apply the changes:
node "${CLAUDE_PLUGIN_ROOT}/bin/haircut.mjs" trim <path to decisions.json> --dry-run
node "${CLAUDE_PLUGIN_ROOT}/bin/haircut.mjs" trim <path to decisions.json>
Restore a bag if needed:
node "${CLAUDE_PLUGIN_ROOT}/bin/haircut.mjs" restore --bag <bag id> --dry-run
node "${CLAUDE_PLUGIN_ROOT}/bin/haircut.mjs" restore --bag <bag id>
All scans, measurements, probes, and reports are read-only. Only the approved trim operation changes files. Never use haircut to edit the contents of CLAUDE.md, AGENTS.md, or MEMORY.md, and never print credential values.
haircut Data Schema & Taxonomy
Run outputs
Each audit creates a run folder whose path is returned in the JSON summary as run_dir.
| File | Purpose |
|---|---|
REPORT.md |
Human-readable findings, measurements, savings, and verdicts; preferred for review. |
report.html |
Self-contained browser interface for reviewing items and exporting decisions. |
report.json |
Complete machine-readable report containing every item; inspect selectively because it can be large. |
decisions.template.json |
Candidate item IDs and proposal structure used to build decisions. |
decisions.json |
User-approved choices and notes for the current run. |
MANIFEST.jsonl |
Operation-by-operation receipt stored inside the recovery bag. |
Decision format
{
"run": "<run id from the summary>",
"items": {
"<item id>": { "decision": "park", "note": "unused in 90 days, user confirmed" },
"<item id>": { "decision": "keep", "note": "" }
}
}
Supported decisions are park, keep, and later. Only park moves files. later is the safe default for unanswered items.
Verdict taxonomy
PARK: Candidate for removal from active configuration and placement in the recovery bag.TRIM: Useful but oversized or excessive and should be reduced rather than removed.FIX: Wanted configuration that is currently broken.ROTATE: Credential appears in plain text and should be rotated by the user.ASK: Requires an explicit user decision, especially for shared team files.KEEP: Remains active.INERT: Has no measurable cost in its current location.
The report can include headline.claude_code.before_first_word.avg_cold, cold_sessions, and max_cold, plus Codex fixed-per-rollout and per-call MCP schema costs. Savings are summarized in top_savings, while verdict counts appear in totals.by_verdict.
haircut Advanced Features
- Multi-agent scanning across Claude Code, Codex, OpenCode, Cursor, Gemini, Copilot, Windsurf, Continue, Claude Desktop, and VS Code.
- Transcript-based token pricing using API usage fields instead of hand-built estimates.
- Optional live MCP probing that measures tool-schema overhead by starting servers as the agent would.
- Hook analysis covering both injected token cost and latency per execution.
- Detection of duplicate skills, commands, and configuration plus scope-shadowing behavior.
- Plugin-aware handling: skills, hooks, and servers shipped inside a plugin are treated as one unit and are not parked independently.
- Interactive
ASKandROTATEdecisions with grouped questions and explicit consequences. - Browser-based review through
report.html, including export of the samedecisions.jsonused by the CLI. - Dry-run enforcement before any real trim operation.
- Reversible dated recovery bags with
MANIFEST.jsonlreceipts and restore support. - Safety controls that preserve team files, avoid credential disclosure, and keep all changes read-only until approval.
- A focused Openclaw Skills workflow that scales the response to the user's question, from a quick startup-cost answer to a complete configuration audit.
SKILL.md
Loading
Related Openclaw Skills
Movie Night Decider
Movie Night Decider finds and ranks sourced movie recommendations that match your preferences while respecting licensing, approval, and verification rules.
XinJianXue Anxiety and Rumination Advisor
An AI advisor that analyzes anxiety and rumination loops through XinJianXue data and converts them into practical, manageable next steps.
XinJianXue Parent–Child Conflict Advisor
A consent-based XinJianXue advisor that analyzes parent–child conflict patterns and provides neutral, practical ways to de-escalate and communicate.
XinJianXue Marriage & Partner Selection Advisor
A consent-first AI advisor that uses XinJianXue reports to explore relationship compatibility, partner selection, and long-term marriage fit without making the decision for the user.
Glasser CLI
Glasser lets Openclaw Skills discover, price, and safely run more than 1,000 third-party data API endpoints through one authenticated Workspace.
1688 One-Stop Procurement Assistant
A complete 1688 procurement workflow that takes users from product discovery and supplier comparison to cart management, checkout, and user-approved payment.
Featured*
