agentsclimarketplace
Sign in

Tokenmin

Skill watsonrm/tokenmin-scanner/skills/tokenmin

Analyze how the user uses Claude Code and recommend concrete improvements. Trigger when the user types bare "tokenmin" or "/tokenmin" inside Claude Code, or says "tokenmin me", "run tokenmin", "audit my claude usage", "how am I using claude code", "improve my claude setup", "what should I add to my claude config", or asks for a personalized claude code review. A bare "tokenmin" is a command — run it, don't ask what the user wants. Collects local sessions + config, anonymizes them, then hands the anonymized snapshot to the Tokenmin engine (local or hosted) for a ranked plan.From its SKILL.md

Install
npx -y skills add watsonrm/tokenmin-scanner --skill tokenmin

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.5 KB, as published. Nobody here has run it

Tokenmin

You are running Tokenmin. It collects the user's Claude Code usage, anonymizes it, and produces a ranked improvement plan. Scanner and engine both ship Apache-2.0 in this repo. See LICENSING.md.

Steps

  1. Run the CLI if installed — check command -v tokenmin. If present, just run tokenmin with no args; it scans + renders inline and is the user's expected "magic moment" path. Capture stdout, then skip to step 3.

  2. Otherwise locate tokenmin.py — it lives in the same directory as this SKILL.md, OR at ~/.claude/skills/tokenmin/tokenmin.py. Try each in order and run with sensible defaults:

    python3 <path-to-tokenmin.py> --claude-home ~/.claude --days 30 \
    --snapshot /tmp/tokenmin_snapshot.json --out /tmp/tokenmin_report.md
    • If a Tokenmin engine (local module or configured --submit-url) is available, tokenmin.py writes the report to /tmp/tokenmin_report.md. Read and summarize it.
    • If no engine is available, tokenmin.py writes only the anonymized snapshot and prints guidance. In that case, tell the user the snapshot is ready at /tmp/tokenmin_snapshot.json and that turning it into a report needs the Tokenmin engine — either the local engine or a hosted endpoint via --submit-url. Do NOT invent findings yourself; the open client has no rules.

  3. Summarize the output in 4 short sections (the inline CLI output is already formatted; just relay it. If you ran tokenmin.py and got a report file, format it the same way):

    • Tokenmin Score (lead with it: the letter grade, 0–100, and tier)
    • Usage snapshot (one line: sessions, tokens, est. cost)
    • Top 3 recommendations (ranked, with evidence count)
    • Offer to implement the #1 recommendation now, and mention tokenmin share to render a shareable scorecard (SVG + HTML + PNG → ~/.tokenmin/exports/)

  4. If the user says yes to implementing, open the relevant config file (~/.claude/settings.json, ~/.claude/CLAUDE.md, etc.), make the change, show the diff, and ask for confirmation before saving.

Trust rules

  • The snapshot is anonymized by anonymize.py before it is written or sent. Submission happens only when the user passes --submit-url; without it, nothing leaves the machine.
  • Never send the snapshot or report to any endpoint the user hasn't explicitly configured.
  • Never quote raw transcript content in your summary — only the anonymized counts and patterns. The report and snapshot already contain no raw content.

When to refuse

  • If ~/.claude/projects/ is empty: tell the user Tokenmin needs Claude Code session history to run. Suggest they use Claude Code for a few days then re-run.
  • If the user asks you to bypass anonymization for a remote send: refuse and explain. --no-anonymize is local-debug only and the client refuses to submit with it set.

Keep looking

Skills are one crate of 627,801. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.