tune
Find and fix real defects in a skill, AGENTS.md, or CLAUDE.md, then prove the description still routes. Runs the bundled token-spending runner, so it starts only on the human's explicit instruction.
- License 许可
- MIT
- 最近更新
- 2026-11-15
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- tune
- description
- Find and fix real defects in a skill, AGENTS.md, or CLAUDE.md, then prove the description still routes. Runs the bundled token-spending runner, so it starts only on the human's explicit instruction.
- argument-hint
- <path to SKILL.md / AGENTS.md / CLAUDE.md>
- disable-model-invocation
- true
02 资源文件 1
- 参考文档agents/openai.yaml205 B
03 SKILL.md 正文
Tune the document at $ARGUMENTS.
You are running the loop the human just asked for, so the runner may spend tokens. Report the estimate before the first call and the actual after.
${CLAUDE_PLUGIN_ROOT} below is Claude Code's expansion for the plugin's
install root. In a client that does not expand it, resolve it yourself: the
plugin root is the directory three levels up from this SKILL.md — the one
that contains skills/. The command blocks below use <target-path> as a
slot, not a variable: replace it with the document path the human gave,
keeping the surrounding double quotes so a path with spaces stays one
argument. Never pass an unexpanded placeholder through to a shell.
1. Read it
Read $ARGUMENTS in full, and the bundled doctrine at
${CLAUDE_PLUGIN_ROOT}/skills/skill-tuner/SKILL.md. You need both: the probe
tells you what is wrong, the doctrine tells you what good looks like when
you fix it.
2. Probe it
python3 ${CLAUDE_PLUGIN_ROOT}/skills/skill-tuner/scripts/tune.py probe \
--target "<target-path>" --run-id tune-$(date +%s) --yes --budget-usd 3 --verify-trials 3
Every finding it reports has already survived three independent skeptics in fresh contexts, plus a code-level check that the quoted text genuinely appears in the document. Treat confirmed findings as real and act on them. They are not suggestions to evaluate; the evaluating already happened.
Read the run's report.md for each finding's quote and proposed fix.
Refuted findings are listed too, with the mode that killed them
(verifier = the rule did not apply; quote_not_found = the probe invented
its evidence). Treat refuted findings as closed: their one remaining job is
the report's refuted line, where the human sees what the probe got wrong.
3. Fix what it found
Apply each confirmed finding. The doctrine's rules are the target shape — positive phrasing over prohibition, one trigger per branch, checkable completion bounds, one meaning in one place.
Fix the defect, not the sentence containing it. A finding about a vague completion bound is asking for a bound that can be checked, not for the word "clearly" to be deleted.
Before each edit, one question: would a competent agent with the current instructions still be expected to fail this way? If yes, there is a gap. If no, defer — the text already required the behavior. When nothing clears that bar, opening no change is the correct outcome, and the report says so per finding; a speculative change is worse than none.
One check before acting: the skeptic panel verifies text against text — it cannot see the filesystem. A finding that asserts an environment fact (a path resolves, a file exists, a config value holds) needs that fact checked on disk first. When the claim turns out false, the ambiguity that made the text read as wrong is usually still real: fix that, not the false claim.
4. Re-probe
Run step 2 again against the edited file.
Done when either the re-probe confirms nothing, or it confirms only findings disjoint from every earlier pass. A finding that overlaps an earlier fix means that fix has not landed — address it and re-probe again; a fix that trades one confirmed defect for another has not landed either.
Disjoint fresh findings at that point are the instrument's floor, not your failure: the probe surfaces ~2 genuine defects per pass on almost any document, including already-audited ones (its measured marginal-value property), so a rich document never re-probes to zero — one loop here ran eight passes before that was accepted. Apply the fresh ones that are cheap and clearly right, and report the rest as residuals.
5. If you changed the description, prove it still routes
A reworded description is a routing change, and routing is the one thing here
that can be measured properly on a single document. Build a battery config
following ${CLAUDE_PLUGIN_ROOT}/configs/receipts-routing-001-neutral.json, with the original
description as original and yours as pruned, and keep its "neutral_ids": true: the
battery lists each skill's id beside its description, and an id routes on its own, so with
the id visible a parity verdict can be carried by the name rather than by the description
you are testing (the doctrine's routing-parity rule records the measurement). Then:
python3 ${CLAUDE_PLUGIN_ROOT}/skills/skill-tuner/scripts/tune.py routing-parity \
--config <your-config> --run-id route-$(date +%s) --yes --budget-usd 3
Keep the new description only on a land verdict. On refuse, restore the
original and say which prompts it lost.
What you may not claim
You probed one document. A before/after count on one document is not evidence of improvement — this project has the receipts for that, having recorded LOST, WON and WORSE for one question at n=1, 6 and 16 respectively.
So: report the defects you fixed and the routing verdict, both of which are real. Do not report that the document is "better" as a measured fact, and do not compare finding counts across the two probe runs as if the difference meant something.
Measuring an actual improvement takes a paired comparison across many
documents — tune.py compare, which needs at least 3 and realistically many
more. Reach for it when changing something that affects a whole suite, such as
a shared convention or a doctrine, and never to grade a single edit.
Report
- Confirmed defects found, and what you changed for each
- Anything refuted, so the human sees what the probe got wrong
- The routing verdict, if the description moved
- Actual spend, from the run output
04 同仓库更多 Skills
review-adversarial-first
Operator-invoked adversarial review: 3 refute-by-default lenses plus a deterministic judge against a PR, branch, or path, then a decision on whether /ce-code-review's topic fan-out runs. Off the default merge path since the 2026-10-17 merge-and-end ruling; run only when asked for by name.
memory-mine
Run one deterministic, budgeted directed-capture batch for an explicit Claude Code project store. Use manual workspace mode only when cross-agent cass sessions are intentionally in scope.
papers
Search preprints from arXiv, bioRxiv, and medRxiv without API keys. Supports search, lookup, extraction, citations, enrichment, and paper downloads that bundle PDFs plus linked GitHub code.
papers
Search preprints from arXiv, bioRxiv, and medRxiv without API keys. Supports search, lookup, extraction, citations, enrichment, and paper downloads that bundle PDFs plus linked GitHub code.
wsl-cdp
Drive the user's real, logged-in Windows browser (Chrome, Edge, or Brave) from a Claude Code session running inside WSL2, over the Chrome DevTools Protocol - read authenticated dashboards and consoles, take real screenshots, open and read pages with the user's sessions. Use when a task needs the user's existing web logins and the session runs in WSL2, where the Claude in Chrome extension cannot work. Not for native-Windows or macOS Claude Code (use the official Claude in Chrome extension), not for anonymous scraping or CI (use headless Playwright inside WSL), not for getting past login walls (those are handed to the human).
token-eater
[BETA] Spend a subscription's idle, about-to-expire credits on safe, machine-verified cleanup. You pick the service (Grok / Codex / Claude) and a token-heavy maintenance skill (de-monolithize, de-slop, simplify, dead-code); that service runs the whole skill -> gate -> self-review -> draft-PR loop on its own credits, in an isolated worktree. token-eater verifies the gate and never merges. Manual invocation only during beta.