tofu-harness (formerly tofable)
🇰🇷 한국어 버전: README.ko.md
Keep an AI-agent harness honest. Transfer a working style into rules and mechanical gates, then audit the harness documents themselves so they keep telling the truth about your tools, paths, and models.
Two pillars
| Skill | What it does |
|---|
| Transfer & enforce | /tofable | A copyable rule layer plus verification gates that bounce a turn when evidence is missing. The original half of this repo. |
| Audit & repair | /harness-audit | A deadline-bounded census of your harness documents — finds where CLAUDE.md / skills / rules have drifted from reality, re-verifies each finding independently, repairs only what's confirmed. Added 2026-07. |
Plugin packaging is skill-only. The gates under hooks/ install separately — see hooks/README.md.
Start here
Install
claude plugin marketplace add treylom/tofu-harness
/plugin install tofu-harness@tofu-harness
Updating later:
claude plugin marketplace update tofu-harness
claude plugin update tofu-harness@tofu-harness
/harness-audit — does your harness still tell the truth?
Claude Opus 5 shipped on 2026-07-24 (TechCrunch · Axios) — and with it, Anthropic's own guidance on how a harness should be written. Two first-party documents are what the audit behind this skill was actually run against: the official Prompting Claude Opus 5 guide (verification instructions cause over-verification; minimizers like "only report high-severity" shrink output literally — report everything, filter downstream) and The new rules of context engineering for Claude 5 generation models (Anthropic blog — the post that also ships /doctor). The table below follows a community digest that condensed this guidance into 7 checkable principles (@nextcocoai on Threads) — used here as a cross-check, with the author's own caveat "I'm not an expert — verify everything" applied and our verification notes included:
| # | Principle (digest) | Check it against your harness |
|---|
| 1 | Don't write what the model already does (verify/re-check instructions duplicate built-in behavior) | ⚠️ Partially — delete generic "double-check" lines, but keep gates born from real regressions: our audit found the model repeatedly not doing the thing the rule enforces. Rule of thumb: no incident history behind it → delete candidate; incident history → keep |
| 2 | Only specify the four things models don't self-regulate: response length, reporting cadence, document length, delegation ceiling | ✅ good entry question for every new rule (retrofit existing docs next-touch only) |
| 3 | Positive examples over prohibitions | ✅ for style rules; prohibitions still earn their keep when they block one exact failure path |
| 4 | No minimizers ("only the serious ones") — collect everything, filter in a second stage | ✅ this is literally how /harness-audit's two-stage verify works (collect everything, then independent re-verification) |
| 5 | Stop scope creep — "exactly the requested scope", halt before clearly out-of-scope actions | ✅ the skill's guards ①–③ exist because of this failure mode |
| 6 | Give the full spec up front, then leave it alone; approval gates only for hard-to-reverse actions | ✅ matches the autonomy pricing this repo's rules use |
| 7 | Long files: repeat a 2–3 line reminder at the end | 🧪 worth an experiment — long-prompt dilution is real |
'%20stop-opacity%3D'0.16'%2F%3E%3Cstop%20offset%3D'1'%20stop-color%3D'rgb(200%2C90%2C60)'%20stop-opacity%3D'0.03'%2F%3E%3C%2FlinearGradient%3E%3C%2Fdefs%3E%3Crect%20width%3D'320'%20height%3D'200'%20fill%3D'url(%23g)'%2F%3E%3Ccircle%20cx%3D'250'%20cy%3D'56'%20r%3D'92'%20fill%3D'rgb(200%2C90%2C60)'%20fill-opacity%3D'0.06'%2F%3E%3Ccircle%20cx%3D'64'%20cy%3D'172'%20r%3D'58'%20fill%3D'rgb(200%2C90%2C60)'%20fill-opacity%3D'0.05'%2F%3E%3C%2Fsvg%3E)
