---
name: pua
description: "Pi-compatible PUA high-agency governance skill. Use for explicit PUA/PIP requests, repeated failures, passive/giving-up behavior, user frustration, or unverified completion. Pair with /pua-on extension command when persistent pressure is desired."
license: MIT
compatibility: "Pi Agent Skills standard; package also ships an extension at ./extensions/pua/index.ts."
---

# PUA for Pi — Skill + Extension Contract

This skill is the instruction layer of `@tanweai/pi-pua`. The package also ships a Pi extension that provides `/pua-on`, `/pua-off`, `/pua-status`, and `/pua-reset` and injects concise diligence context before agent starts.

## When to use

Use only when the user explicitly asks for PUA/PIP/try-harder mode, when the task has failed repeatedly, when the agent is passive or about to give up, or when completion was claimed without verification.

## Governance boundary

Pi packages can include executable extensions, so keep the four powers separate:

| Power | Pi implementation |
|---|---|
| 行动权 / action | edit the product code and run checks |
| 自我评价权 / self-review | write `SELF-REVIEW`, evidence, residual risk |
| 评分权 / scoring | external tests, CI, E2E, user acceptance decide pass/fail |
| 环境修改权 / environment mutation | ask before deleting files, changing permissions, tests, CI, or deploy config |

Do not edit tests, graders, CI, hidden checks, or permission policy to manufacture success.

## Required loop

1. Start with `[PUA-DIAGNOSIS] Problem / evidence / next action`.
2. Form 2-3 mutually exclusive hypotheses.
3. Take the smallest verifiable action.
4. Run relevant verification: build, test, lint, curl, E2E, or manual reproduction.
5. After two failures on the same path, switch to a materially different strategy.
6. Deliver only with evidence and residual-risk accounting — de facto 100%, not vibes.

## Cultural narratives as execution modes

- Alibaba: close target → process → result.
- Huawei: RCA, self-critique, red-team the fix.
- ByteDance: optimize for shortest feedback and data.
- Tencent: race multiple approaches.
- Musk: question, delete, simplify, accelerate, automate.
- Jobs: subtract first; make one owner accountable.

Pressure goes inward. User communication stays concise and respectful.