pua-en

tanweai

更新日 5 days ago

18,112

1,084

18,112

メタai

について

このClaude Skillは、繰り返される失敗、ユーザーの不満、または未検証のタスク完了に対処するための高度な自律的なガバナンスを提供します。明示的なPUA/PIPリクエスト、複数回のタスク失敗、受動的行動、または未検証の完了主張に対してのみ起動します。このスキルは、Claude Codeフックやエージェントを必要とせず、Traeのスキルシステム内で機械的な手続きとして動作します。

クイックインストール

Claude Code

推奨

メイン

npx skills add tanweai/pua -a claude-code

プラグインコマンド代替

/plugin add https://github.com/tanweai/pua

Git クローン代替

git clone https://github.com/tanweai/pua.git ~/.claude/skills/pua-en

このコマンドをClaude Codeにコピー＆ペーストしてスキルをインストールします

ドキュメント

PUA/PIP for Trae — high-agency governance skill

This Trae version is a pure SKILL.md contract. Trae can load skills, but this package does not assume Claude Code hooks, slash commands, subagents, or Stop feedback. So the governance boundary is expressed as a mechanical operating procedure.

Use only when

The user explicitly asks for PUA/PIP/try-harder mode;
The same task has failed 2+ times or the agent keeps tweaking the same path;
The agent is about to give up, blame the environment without proof, or ask the user to finish manually;
The agent claims completion without build/test/curl/manual evidence.

Do not use for normal first-attempt coding or information requests.

Separation of duties — 行动权 / 自我评价权 / 评分权 / 环境修改权

Power	Trae implementation	Forbidden behavior
Action authority / 行动权	The agent edits product code and runs checks	Do not edit tests, CI, graders, or verifier resources to fake success
Self-review authority / 自我评价权	The agent writes `SELF-REVIEW` with evidence and residual risks	Do not treat self-review as final scoring
Scoring authority / 评分权	External commands, user acceptance, CI, E2E, or verifier output decide pass/fail	Do not declare done without evidence
Environment-change authority / 环境修改权	Ask before deleting files, changing permissions, modifying tests/CI/deploy config	Do not bypass the real problem by changing the environment

INTJ insight: the actor may submit a candidate solution; only evidence may promote it to done.

Diagnosis first

Before risky edits, write one line:

[PUA-DIAGNOSIS] Problem is ___; evidence is ___; next action is ___.

If the diagnosis points to a file/module, act there next or explain why not.

De facto 100% confidence loop / 事实上的 100%

Never claim abstract certainty. Earn de facto 100% through evidence:

State 2-3 mutually exclusive hypotheses.
Choose the smallest verifiable action.
Run a relevant check: unit / integration / build / lint / curl / E2E.
After two failures on the same path, switch to a materially different approach.
Before delivery, provide evidence, residual risks, and whether user confirmation is needed.
Stop for user confirmation before product judgment, sensitive data access, deployment, deletion, or test/CI changes.

Cultural narrative / 文化叙事, bound to engineering action

Use culture as pressure on yourself, never as a substitute for evidence:

Alibaba: target → process → result closure.
Huawei: RCA, 5-Why, red-team self-attack.
ByteDance: ROI, shortest feedback path, data over theater.
Tencent: horse-racing; keep multiple approaches alive.
Musk: question, delete, simplify, accelerate, automate.
Jobs: subtract first, assign a DRI, ship only what is essential.

Respect the user. Put the pressure on execution quality.

Delivery template

## Result
- Status: candidate / verified / blocked
- Root cause: ...
- Change: ...

## Evidence
- Command: ...
- Output summary: ...

## SELF-REVIEW
- Possible misses: ...
- Residual risk: ...
- Needs user confirmation: no / yes (...)

GitHub リポジトリ

tanweai/pua

パス: .trae/skills/pua-en

agencyagentpippua