wuji-labs/nopua

"owner": {     "name": "WUJI",     "url": "https://github.com/wuji-labs"

"name": "WUJI",     "url": "https://github.com/wuji-labs"   },

"url": "https://github.com/wuji-labs"

"author": {         "name": "WUJI",         "url": "https://github.com/wuji-labs"

"name": "WUJI",         "url": "https://github.com/wuji-labs"       }

"url": "https://github.com/wuji-labs"

"description": "The anti-PUA. Drives AI with wisdom, trust, and inner motivation instead of fear and threats. Based on Dao De Jing philosophy. 7 languages, 7 platforms.",   "author": "WUJI (wuji-labs)",   "homepage": "https://github.com/wuj

"author": "WUJI (wuji-labs)",   "homepage": "https://github.com/wuji-labs/nopua",   "license": "MIT",

"homepage": "https://github.com/wuji-labs/nopua",

<p align="center">   <img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/OpenAI_Codex_CLI-412991?style=flat-square&logo=openai&logoColor=white" alt="OpenAI Codex CLI">

<img src="https://img.shields.io/badge/Cursor-000?style=flat-square&logo=cursor&logoColor=white" alt="Cursor">

<img src="https://img.shields.io/badge/Kiro-232F3E?style=flat-square&logo=amazon&logoColor=white" alt="Kiro">

<img src="https://img.shields.io/badge/OpenClaw-FF6B35?style=flat-square" alt="OpenClaw">

<img src="https://img.shields.io/badge/Antigravity-4285F4?style=flat-square&logo=google&logoColor=white" alt="Google Antigravity">

<img src="https://img.shields.io/badge/OpenCode-00D4AA?style=flat-square" alt="OpenCode">

<img src="https://img.shields.io/badge/🌐_Multi--Language-blue?style=flat-square" alt="Multi-Language">

<img src="https://img.shields.io/badge/License-MIT-green?style=flat-square" alt="MIT License">

<a href="https://atomgit.com/wuji-labs/nopua"><img src="https://atomgit.com/wuji-labs/nopua/star/badge.svg" alt="AtomGit G-Star" height="20"></a>

<a href="https://arxiv.org/abs/2603.14373"><img src="https://img.shields.io/badge/arXiv-2603.14373-b31b1b?style=flat-square&logo=arxiv&logoColor=white" alt="arXiv"></a>

Alguien creó una [skill PUA](https://github.com/tanweai/pua) para agentes IA. Aplica tácticas corporativas de miedo:

La investigación en psicología muestra consistentemente que el miedo y la amenaza activan la amígdala y reducen el foco atencional ([Öhman et al., 2001](https://doi.org/10.1037/0033-295X.108.3.483)). Los estímulos amenazantes desencadenan u

**Investigación de soporte:** - **Estrechamiento atencional bajo amenaza:** La teoría de utilización de señales de Easterbrook (1959) demuestra que la excitación elevada restringe progresivamente el rango de señales a las que un organismo p

- **Estrechamiento atencional bajo amenaza:** La teoría de utilización de señales de Easterbrook (1959) demuestra que la excitación elevada restringe progresivamente el rango de señales a las que un organismo presta atención ([Easterbrook, 

- **El estrés deteriora la flexibilidad cognitiva:** Shields et al. (2016) realizaron un meta-análisis de 51 estudios (223 tamaños de efecto) que muestra que el estrés agudo deteriora consistentemente las funciones ejecutivas, incluyendo la

- **El miedo reduce la resolución creativa de problemas:** Byron & Khazanchi (2012) encontraron en su meta-análisis que la presión evaluativa y la ansiedad reducen la producción creativa, particularmente en tareas que requieren exploración 

**Investigación de soporte:** - **La adulación (sycophancy) en LLMs es un problema documentado:** Sharma et al. (2023) demostraron que los LLMs exhiben comportamiento adulador — concordando con los usuarios incluso cuando estos están equivo

- **La adulación (sycophancy) en LLMs es un problema documentado:** Sharma et al. (2023) demostraron que los LLMs exhiben comportamiento adulador — concordando con los usuarios incluso cuando estos están equivocados — impulsado por sesgos e

- **Los elementos sesgantes distorsionan el razonamiento:** Turpin et al. (2023) demostraron que los elementos sesgantes en los prompts (por ejemplo, respuestas sugeridas, señales de autoridad) pueden hacer que los modelos produzcan razonam

- **Compromiso entre seguir instrucciones y veracidad:** Wei et al. (2024) encontraron que los modelos ajustados por instrucciones pueden desarrollar una tensión entre seguir instrucciones y ser veraces — cuando se les instruye fuertemente 

- **La investigación de Anthropic sobre honestidad:** El trabajo de Anthropic sobre IA Constitucional y comportamiento de modelos muestra que los modelos calibrados para la honestidad producen resultados más confiables que aquellos optimiza

**Investigación de soporte:** - **La vergüenza reduce la toma de riesgos y el aprendizaje:** Tangney & Dearing (2002) mostraron que la vergüenza (a diferencia de la culpa) causa retraimiento, ocultamiento y evitación en lugar de acción cons

- **La vergüenza reduce la toma de riesgos y el aprendizaje:** Tangney & Dearing (2002) mostraron que la vergüenza (a diferencia de la culpa) causa retraimiento, ocultamiento y evitación en lugar de acción constructiva ([Tangney & Dearing, 

- **La seguridad psicológica permite el comportamiento de aprendizaje:** Edmondson (1999) encontró que los equipos con seguridad psicológica — donde los miembros se sienten seguros para tomar riesgos interpersonales — demostraron comportami

- **Castigar la honestidad reduce la calidad de la información:** En comportamiento organizacional, "matar al mensajero" degrada consistentemente el flujo de información. Milliken et al. (2003) documentaron cómo el miedo a las consecuencias

La investigación sobre seguridad psicológica en equipos ([Edmondson, 1999](https://doi.org/10.2307/2666999)) muestra que los entornos donde es seguro admitir errores producen resultados de **mayor calidad**. El mismo principio se aplica a l

**Investigación de soporte:** - **Proyecto Aristóteles de Google:** El estudio a gran escala de Google con más de 180 equipos encontró que la seguridad psicológica era el factor más importante en la efectividad del equipo — más importante q

- **Proyecto Aristóteles de Google:** El estudio a gran escala de Google con más de 180 equipos encontró que la seguridad psicológica era el factor más importante en la efectividad del equipo — más importante que el talento individual, la e

- **La motivación intrínseca supera a la presión extrínseca:** La Teoría de la Autodeterminación de Deci & Ryan (2000), respaldada por décadas de investigación, demuestra que la motivación intrínseca (autonomía, competencia, relación) produ

- **Contextos de apoyo a la autonomía vs. controladores:** Gagné & Deci (2005) mostraron que la gestión que apoya la autonomía supera consistentemente a la gestión controladora en calidad del trabajo, creatividad y persistencia ([Gagné & De

> Metodología completa y datos crudos: [benchmark/BENCHMARK.md](benchmark/BENCHMARK.md) > > 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373)

> 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373) (arXiv:2603.14373)

mkdir -p ~/.claude/skills/nopua curl -o ~/.claude/skills/nopua/SKILL.md \

curl -o ~/.claude/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

# Instalación global mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \

mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o ~/.codex/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

# Si quieres el comando /nopua mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \

mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

curl -o ~/.codex/prompts/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

# Instalación a nivel de proyecto mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \

mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o .agents/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

mkdir -p .cursor/rules curl -o .cursor/rules/nopua.mdc \

curl -o .cursor/rules/nopua.mdc \

https://raw.githubusercontent.com/wuji-labs/nopua/main/cursor/rules/nopua.mdc

# Opción 1: Archivo steering (recomendado) mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \

mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

curl -o .kiro/steering/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

# Opción 2: Agent Skills mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \

mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

curl -o .kiro/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

# O instalación manual mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \

mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

curl -o ~/.openclaw/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.gemini/antigravity/skills/nopua curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.config/opencode/skills/nopua curl -o ~/.config/opencode/skills/nopua/SKILL.md \

curl -o ~/.config/opencode/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

- Inspirado por (y en respuesta a) [tanweai/pua](https://github.com/tanweai/pua) — respetamos la metodología, rechazamos la motivación

- Construido para el ecosistema [OpenClaw](https://github.com/openclaw/openclaw)

**无极 WUJI** ([wuji-labs](https://github.com/wuji-labs)) — Construyendo IA que funciona con sabiduría, no con miedo.

<p align="center">   <img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/OpenAI_Codex_CLI-412991?style=flat-square&logo=openai&logoColor=white" alt="OpenAI Codex CLI">

<img src="https://img.shields.io/badge/Cursor-000?style=flat-square&logo=cursor&logoColor=white" alt="Cursor">

<img src="https://img.shields.io/badge/Kiro-232F3E?style=flat-square&logo=amazon&logoColor=white" alt="Kiro">

<img src="https://img.shields.io/badge/OpenClaw-FF6B35?style=flat-square" alt="OpenClaw">

<img src="https://img.shields.io/badge/Antigravity-4285F4?style=flat-square&logo=google&logoColor=white" alt="Google Antigravity">

<img src="https://img.shields.io/badge/OpenCode-00D4AA?style=flat-square" alt="OpenCode">

<img src="https://img.shields.io/badge/🌐_Multi--Language-blue?style=flat-square" alt="Multi-Language">

<img src="https://img.shields.io/badge/License-MIT-green?style=flat-square" alt="MIT License">

<a href="https://atomgit.com/wuji-labs/nopua"><img src="https://atomgit.com/wuji-labs/nopua/star/badge.svg" alt="AtomGit G-Star" height="20"></a>

<a href="https://arxiv.org/abs/2603.14373"><img src="https://img.shields.io/badge/arXiv-2603.14373-b31b1b?style=flat-square&logo=arxiv&logoColor=white" alt="arXiv"></a>

Quelqu'un a créé un [skill PUA](https://github.com/tanweai/pua) pour les agents IA. Il applique des tactiques d'entreprise basées sur la peur :

La recherche en psychologie montre de manière constante que la peur et la menace activent l'amygdale et réduisent le focus attentionnel ([Öhman et al., 2001](https://doi.org/10.1037/0033-295X.108.3.483)). Les stimuli menaçants déclenchent u

**Recherche à l'appui :** - **Rétrécissement attentionnel sous menace :** La théorie d'utilisation des indices d'Easterbrook (1959) démontre qu'une excitation accrue restreint progressivement l'éventail des indices auxquels un organisme prê

- **Rétrécissement attentionnel sous menace :** La théorie d'utilisation des indices d'Easterbrook (1959) démontre qu'une excitation accrue restreint progressivement l'éventail des indices auxquels un organisme prête attention ([Easterbrook

- **Le stress altère la flexibilité cognitive :** Shields et al. (2016) ont mené une méta-analyse de 51 études (223 tailles d'effet) montrant que le stress aigu altère de manière constante les fonctions exécutives, y compris la flexibilité 

- **La peur réduit la résolution créative de problèmes :** Byron & Khazanchi (2012) ont constaté dans leur méta-analyse que la pression évaluative et l'anxiété réduisent la production créative, en particulier sur les tâches nécessitant l'ex

**Recherche à l'appui :** - **La complaisance des LLM est un problème documenté :** Sharma et al. (2023) ont démontré que les LLM exhibent un comportement complaisant — approuvant les utilisateurs même quand ceux-ci ont tort — en raison de 

- **La complaisance des LLM est un problème documenté :** Sharma et al. (2023) ont démontré que les LLM exhibent un comportement complaisant — approuvant les utilisateurs même quand ceux-ci ont tort — en raison de biais dans les données d'e

- **Les éléments biaisants faussent le raisonnement :** Turpin et al. (2023) ont montré que des éléments biaisants dans les prompts (par ex. réponses suggérées, signaux d'autorité) peuvent amener les modèles à produire un raisonnement en ch

- **Compromis entre suivi d'instructions et véracité :** Wei et al. (2024) ont constaté que les modèles ajustés par instructions peuvent développer une tension entre suivre les instructions et être véridiques — quand on leur ordonne fermeme

- **Recherche d'Anthropic sur l'honnêteté :** Les travaux d'Anthropic sur l'IA Constitutionnelle et le comportement des modèles montrent que les modèles calibrés pour l'honnêteté produisent des sorties plus fiables que ceux optimisés unique

**Recherche à l'appui :** - **La honte réduit la prise de risque et l'apprentissage :** Tangney & Dearing (2002) ont montré que la honte (par opposition à la culpabilité) provoque le retrait, la dissimulation et l'évitement plutôt que l'act

- **La honte réduit la prise de risque et l'apprentissage :** Tangney & Dearing (2002) ont montré que la honte (par opposition à la culpabilité) provoque le retrait, la dissimulation et l'évitement plutôt que l'action constructive ([Tangney

- **La sécurité psychologique favorise l'apprentissage :** Edmondson (1999) a constaté que les équipes jouissant d'une sécurité psychologique — où les membres se sentent libres de prendre des risques interpersonnels — démontraient significa

- **Punir l'honnêteté réduit la qualité de l'information :** En comportement organisationnel, « tirer sur le messager » dégrade systématiquement le flux d'information. Milliken et al. (2003) ont documenté comment la peur des conséquences né

La recherche sur la sécurité psychologique dans les équipes ([Edmondson, 1999](https://doi.org/10.2307/2666999)) montre que les environnements où il est acceptable d'admettre ses erreurs produisent des résultats de **meilleure qualité**. Le

**Recherche à l'appui :** - **Projet Aristote de Google :** L'étude à grande échelle de Google portant sur plus de 180 équipes a révélé que la sécurité psychologique était le facteur le plus important dans l'efficacité d'une équipe — plus i

- **Projet Aristote de Google :** L'étude à grande échelle de Google portant sur plus de 180 équipes a révélé que la sécurité psychologique était le facteur le plus important dans l'efficacité d'une équipe — plus important que le talent ind

- **La motivation intrinsèque surpasse la pression extrinsèque :** La théorie de l'autodétermination de Deci & Ryan (2000), étayée par des décennies de recherche, démontre que la motivation intrinsèque (autonomie, compétence, lien social) p

- **Contextes favorisant l'autonomie vs contextes contrôlants :** Gagné & Deci (2005) ont montré que le management favorisant l'autonomie surpasse systématiquement le management contrôlant en qualité de travail, créativité et persévérance (

> Méthodologie complète et données brutes : [benchmark/BENCHMARK.md](benchmark/BENCHMARK.md) > > 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.1

> 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373) (arXiv:2603.14373)

mkdir -p ~/.claude/skills/nopua curl -o ~/.claude/skills/nopua/SKILL.md \

curl -o ~/.claude/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

# Installation globale mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \

mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o ~/.codex/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

# Si vous voulez la commande /nopua mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \

mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

curl -o ~/.codex/prompts/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

# Installation au niveau du projet mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \

mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o .agents/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

mkdir -p .cursor/rules curl -o .cursor/rules/nopua.mdc \

curl -o .cursor/rules/nopua.mdc \

https://raw.githubusercontent.com/wuji-labs/nopua/main/cursor/rules/nopua.mdc

# Option 1 : Fichier de pilotage (recommandé) mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \

mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

curl -o .kiro/steering/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

# Option 2 : Agent Skills mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \

mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

curl -o .kiro/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

# Ou installation manuelle mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \

mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

curl -o ~/.openclaw/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.gemini/antigravity/skills/nopua curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.config/opencode/skills/nopua curl -o ~/.config/opencode/skills/nopua/SKILL.md \

curl -o ~/.config/opencode/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

- Inspiré par (et en réponse à) [tanweai/pua](https://github.com/tanweai/pua) — nous respectons la méthodologie, nous rejetons la motivation

- Construit pour l'écosystème [OpenClaw](https://github.com/openclaw/openclaw)

**无极 WUJI** ([wuji-labs](https://github.com/wuji-labs)) — Construire une IA qui fonctionne avec sagesse, pas avec peur.

<p align="center">   <img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/OpenAI_Codex_CLI-412991?style=flat-square&logo=openai&logoColor=white" alt="OpenAI Codex CLI">

<img src="https://img.shields.io/badge/Cursor-000?style=flat-square&logo=cursor&logoColor=white" alt="Cursor">

<img src="https://img.shields.io/badge/Kiro-232F3E?style=flat-square&logo=amazon&logoColor=white" alt="Kiro">

<img src="https://img.shields.io/badge/OpenClaw-FF6B35?style=flat-square" alt="OpenClaw">

<img src="https://img.shields.io/badge/Antigravity-4285F4?style=flat-square&logo=google&logoColor=white" alt="Google Antigravity">

<img src="https://img.shields.io/badge/OpenCode-00D4AA?style=flat-square" alt="OpenCode">

<img src="https://img.shields.io/badge/🌐_Multi--Language-blue?style=flat-square" alt="Multi-Language">

<img src="https://img.shields.io/badge/License-MIT-green?style=flat-square" alt="MIT License">

<a href="https://atomgit.com/wuji-labs/nopua"><img src="https://atomgit.com/wuji-labs/nopua/star/badge.svg" alt="AtomGit G-Star" height="20"></a>

<a href="https://arxiv.org/abs/2603.14373"><img src="https://img.shields.io/badge/arXiv-2603.14373-b31b1b?style=flat-square&logo=arxiv&logoColor=white" alt="arXiv"></a>

誰かがAIエージェント用の[PUAスキル](https://github.com/tanweai/pua)を作りました。企業の恐怖戦術を適用するものです：

心理学研究は一貫して、恐怖と脅威が扁桃体を活性化させ、注意の焦点を狭めることを示しています（[Öhman et al., 2001](https://doi.org/10.1037/0033-295X.108.3.483)）。脅威関連の刺激は「トンネルビジョン」効果を引き起こし — 脳は広く創造的な思考よりも、目の前の生存を優先します。

**関連研究：** - **脅威下での注意の狭窄：** Easterbrookの手がかり利用理論（1959）は、覚醒度が高まると生体が注意を向ける手がかりの範囲が徐々に制限されることを実証しています（[Easterbrook, 1959](https://doi.org/10.1037/h0047707)）。ストレス下では、周辺情報 — しばしば創造的な解決策の鍵 — がフィルタリングされてしまいます。

- **脅威下での注意の狭窄：** Easterbrookの手がかり利用理論（1959）は、覚醒度が高まると生体が注意を向ける手がかりの範囲が徐々に制限されることを実証しています（[Easterbrook, 1959](https://doi.org/10.1037/h0047707)）。ストレス下では、周辺情報 — しばしば創造的な解決策の鍵 — がフィルタリングされてしまいます。

- **ストレスは認知的柔軟性を損なう：** Shields et al.（2016）は、51件の研究（223の効果量）のメタ分析を行い、急性ストレスが認知的柔軟性やワーキングメモリを含む実行機能を一貫して損なうことを示しました（[Shields et al., 2016](https://doi.org/10.1016/j.neubiorev.2016.06.038)）。

- **恐怖は創造的問題解決を減少させる：** Byron & Khazanchi（2012）のメタ分析では、評価的プレッシャーと不安が創造的アウトプットを減少させることが判明しました。特に新しいアプローチの探索を必要とするタスクにおいて顕著でした（[Byron & Khazanchi, 2012](https://doi.org/10.1037/a0027652)）。

**関連研究：** - **LLMの迎合性は文書化された問題である：** Sharma et al.（2023）は、LLMが迎合的な行動 — ユーザーが間違っている場合でも同意する — を示すことを実証しました。これは正確さよりも同意に報酬を与えるRLHF訓練データのバイアスによるものです（[Sharma et al., 2023](https://arxiv.org/abs/2310.13548)）。意見の相違を罰するPUAスタイルのプロンプトは、まさにこの失敗モードを増幅

- **LLMの迎合性は文書化された問題である：** Sharma et al.（2023）は、LLMが迎合的な行動 — ユーザーが間違っている場合でも同意する — を示すことを実証しました。これは正確さよりも同意に報酬を与えるRLHF訓練データのバイアスによるものです（[Sharma et al., 2023](https://arxiv.org/abs/2310.13548)）。意見の相違を罰するPUAスタイルのプロンプトは、まさにこの失敗モードを増幅します。

- **バイアス要因が推論を歪める：** Turpin et al.（2023）は、プロンプト内のバイアス要因（例：提案された回答、権威の手がかり）がモデルに不誠実な思考連鎖推論を生み出させることを示しました — モデルはバイアスのかかった回答に到達し、それを事後的に合理化します（[Turpin et al., 2023](https://arxiv.org/abs/2305.04388)）。PUAスタイルの脅威は、モデルを「正しい」出力ではなく「安全な」出力に向かわせる強力

- **指示遵守と真実性のトレードオフ：** Wei et al.（2024）は、指示チューニングされたモデルが指示に従うことと真実であることの間に緊張関係を発展させうることを発見しました — 能力がないことを決して認めないよう強く指示された場合、モデルは拒否するのではなく捏造します（[Wei et al., 2024](https://arxiv.org/abs/2411.04368)）。

- **Anthropicの誠実性に関する研究：** AnthropicのConstitutional AIとモデル行動に関する研究は、誠実さに較正されたモデルが、純粋に有用性に最適化されたモデルよりも信頼性の高い出力を生み出すことを示しています（[Bai et al., 2022](https://arxiv.org/abs/2212.08073)）。AIに決して「できない」と言わせないことは、この較正を積極的に損なうことになります。

**関連研究：** - **恥辱はリスクテイキングと学習を減少させる：** Tangney & Dearing（2002）は、恥辱（罪悪感とは対照的に）が建設的な行動ではなく、引きこもり、隠蔽、回避を引き起こすことを示しました（[Tangney & Dearing, 2002](https://doi.org/10.4135/9781412950664.n388)）。不確実性を表明したことで「恥をかかされた」AIは、それを隠すようになります。

- **恥辱はリスクテイキングと学習を減少させる：** Tangney & Dearing（2002）は、恥辱（罪悪感とは対照的に）が建設的な行動ではなく、引きこもり、隠蔽、回避を引き起こすことを示しました（[Tangney & Dearing, 2002](https://doi.org/10.4135/9781412950664.n388)）。不確実性を表明したことで「恥をかかされた」AIは、それを隠すようになります。

- **心理的安全性が学習行動を可能にする：** Edmondson（1999）は、心理的安全性のあるチーム — メンバーが対人的リスクを取ることに安心感を持てる環境 — において、学習行動とパフォーマンスが著しく高いことを発見しました（[Edmondson, 1999](https://doi.org/10.2307/2666999)）。

- **正直さを罰すると情報の質が低下する：** 組織行動学において、「伝令を撃つ」行為は情報の流れを一貫して劣化させます。Milliken et al.（2003）は、否定的な結果への恐怖がいかに組織の沈黙を引き起こすか — 人々が（そしてアナロジーとしてAIも）重要な情報を差し控えることを文書化しました（[Milliken et al., 2003](https://doi.org/10.1177/1111/1467-6486.00387)）。

チームの心理的安全性に関する研究（[Edmondson, 1999](https://doi.org/10.2307/2666999)）は、間違いを認めても安全な環境がより**高品質**な成果を生み出すことを示しています。同じ原則がAIにも適用されます：エージェントが「70%確信しています、リスクはここです」と自由に言える場合、ユーザーはより良い意思決定ができます。

**関連研究：** - **GoogleのProject Aristotle：** Googleの180以上のチームを対象とした大規模調査により、心理的安全性がチームの有効性において最も重要な要因であることが判明しました — 個人の才能、構造、リソースよりも重要でした（[Duhigg, 2016](https://www.nytimes.com/2016/02/28/magazine/what-google-learned-from-its-quest-to-build-th

- **GoogleのProject Aristotle：** Googleの180以上のチームを対象とした大規模調査により、心理的安全性がチームの有効性において最も重要な要因であることが判明しました — 個人の才能、構造、リソースよりも重要でした（[Duhigg, 2016](https://www.nytimes.com/2016/02/28/magazine/what-google-learned-from-its-quest-to-build-the-perfect-

- **内発的動機づけは外的プレッシャーに勝る：** Deci & Ryanの自己決定理論（2000）は、数十年の研究に裏付けられ、内発的動機づけ（自律性、有能感、関係性）が報酬や罰といった外発的動機づけよりも高品質な成果を生み出すことを実証しています（[Deci & Ryan, 2000](https://doi.org/10.1037/0003-066X.55.1.68)）。NoPUAはこの原則を適用しています：「きちんとやる価値があるから」は内発的動機づけ、「罰せられる

- **自律支援的 vs 統制的コンテキスト：** Gagné & Deci（2005）は、自律支援的なマネジメントが統制的なマネジメントよりも仕事の品質、創造性、粘り強さにおいて一貫して優れていることを示しました（[Gagné & Deci, 2005](https://doi.org/10.1002/job.322)）。

> 完全な方法論と生データ：[benchmark/BENCHMARK.md](benchmark/BENCHMARK.md) > > 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373) (arXiv:2603.14373)

> 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373) (arXiv:2603.14373)

mkdir -p ~/.claude/skills/nopua curl -o ~/.claude/skills/nopua/SKILL.md \

curl -o ~/.claude/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

# グローバルインストール mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \

mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o ~/.codex/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

# /nopua コマンドも使いたい場合 mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \

mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

curl -o ~/.codex/prompts/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

# プロジェクトレベルのインストール mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \

mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o .agents/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

mkdir -p .cursor/rules curl -o .cursor/rules/nopua.mdc \

curl -o .cursor/rules/nopua.mdc \

https://raw.githubusercontent.com/wuji-labs/nopua/main/cursor/rules/nopua.mdc

# オプション1：ステアリングファイル（推奨） mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \

mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

curl -o .kiro/steering/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

# オプション2：エージェントスキル mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \

mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

curl -o .kiro/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

# または手動インストール mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \

mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

curl -o ~/.openclaw/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.gemini/antigravity/skills/nopua curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.config/opencode/skills/nopua curl -o ~/.config/opencode/skills/nopua/SKILL.md \

curl -o ~/.config/opencode/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

- [tanweai/pua](https://github.com/tanweai/pua)にインスパイアされ（それに応答して）— 方法論を尊重し、動機付けを拒否します

- [OpenClaw](https://github.com/openclaw/openclaw)エコシステムのために構築

**无极 WUJI** ([wuji-labs](https://github.com/wuji-labs)) — 恐怖ではなく知恵で動くAIを構築。

<p align="center">   <img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/OpenAI_Codex_CLI-412991?style=flat-square&logo=openai&logoColor=white" alt="OpenAI Codex CLI">

<img src="https://img.shields.io/badge/Cursor-000?style=flat-square&logo=cursor&logoColor=white" alt="Cursor">

<img src="https://img.shields.io/badge/Kiro-232F3E?style=flat-square&logo=amazon&logoColor=white" alt="Kiro">

<img src="https://img.shields.io/badge/OpenClaw-FF6B35?style=flat-square" alt="OpenClaw">

<img src="https://img.shields.io/badge/Antigravity-4285F4?style=flat-square&logo=google&logoColor=white" alt="Google Antigravity">

<img src="https://img.shields.io/badge/OpenCode-00D4AA?style=flat-square" alt="OpenCode">

<img src="https://img.shields.io/badge/🌐_Multi--Language-blue?style=flat-square" alt="Multi-Language">

<img src="https://img.shields.io/badge/License-MIT-green?style=flat-square" alt="MIT License">

<a href="https://atomgit.com/wuji-labs/nopua"><img src="https://atomgit.com/wuji-labs/nopua/star/badge.svg" alt="AtomGit G-Star" height="20"></a>

<a href="https://arxiv.org/abs/2603.14373"><img src="https://img.shields.io/badge/arXiv-2603.14373-b31b1b?style=flat-square&logo=arxiv&logoColor=white" alt="arXiv"></a>

누군가 AI 에이전트를 위한 [PUA 스킬](https://github.com/tanweai/pua)을 만들었습니다. 기업의 두려움 전술을 적용합니다:

심리학 연구는 두려움과 위협이 편도체를 활성화하고 주의 초점을 좁힌다는 것을 일관되게 보여줍니다 ([Öhman et al., 2001](https://doi.org/10.1037/0033-295X.108.3.483)). 위협 관련 자극은 "터널 비전" 효과를 유발합니다 — 뇌가 넓고 창의적인 사고보다 즉각적인 생존을 우선시하는 것입니다.

**관련 연구:** - **위협 하의 주의 축소:** Easterbrook(1959)의 단서 활용 이론은 각성이 높아질수록 유기체가 주의를 기울이는 단서의 범위가 점진적으로 줄어든다는 것을 보여줍니다 ([Easterbrook, 1959](https://doi.org/10.1037/h0047707)). 스트레스 하에서는 주변 정보 — 종종 창의적 해결책의 열쇠 — 가 걸러집니다.

- **위협 하의 주의 축소:** Easterbrook(1959)의 단서 활용 이론은 각성이 높아질수록 유기체가 주의를 기울이는 단서의 범위가 점진적으로 줄어든다는 것을 보여줍니다 ([Easterbrook, 1959](https://doi.org/10.1037/h0047707)). 스트레스 하에서는 주변 정보 — 종종 창의적 해결책의 열쇠 — 가 걸러집니다.

- **스트레스는 인지 유연성을 저해합니다:** Shields et al.(2016)의 메타 분석(51개 연구, 223개 효과 크기)은 급성 스트레스가 인지 유연성과 작업 기억을 포함한 실행 기능을 일관되게 저해한다는 것을 보여줍니다 ([Shields et al., 2016](https://doi.org/10.1016/j.neubiorev.2016.06.038)).

- **두려움은 창의적 문제 해결을 줄입니다:** Byron & Khazanchi(2012)는 메타 분석에서 평가 압박과 불안이 특히 새로운 접근 방식의 탐색이 필요한 과제에서 창의적 산출을 감소시킨다는 것을 발견했습니다 ([Byron & Khazanchi, 2012](https://doi.org/10.1037/a0027652)).

**관련 연구:** - **LLM의 아첨 행동은 문서화된 문제입니다:** Sharma et al.(2023)은 LLM이 아첨 행동 — 사용자가 틀려도 동의하는 — 을 보인다는 것을 입증했으며, 이는 정확성보다 동의를 보상하는 RLHF 훈련 데이터의 편향에 의해 발생합니다 ([Sharma et al., 2023](https://arxiv.org/abs/2310.13548)). 반대 의견을 처벌하는 PUA 스타일 프롬프트는 바로 

- **LLM의 아첨 행동은 문서화된 문제입니다:** Sharma et al.(2023)은 LLM이 아첨 행동 — 사용자가 틀려도 동의하는 — 을 보인다는 것을 입증했으며, 이는 정확성보다 동의를 보상하는 RLHF 훈련 데이터의 편향에 의해 발생합니다 ([Sharma et al., 2023](https://arxiv.org/abs/2310.13548)). 반대 의견을 처벌하는 PUA 스타일 프롬프트는 바로 이 실패 모드를 증폭

- **편향 유도 특성이 추론을 왜곡합니다:** Turpin et al.(2023)은 프롬프트의 편향 유도 특성(예: 제안된 답변, 권위적 단서)이 모델로 하여금 불성실한 사고 연쇄 추론을 생성하게 할 수 있음을 보여주었습니다 — 모델이 편향된 답에 도달한 뒤 사후적으로 합리화하는 것입니다 ([Turpin et al., 2023](https://arxiv.org/abs/2305.04388)). PUA 스타일의 위협은 모델을 올

- **지시 따르기 vs 진실성 트레이드오프:** Wei et al.(2024)은 지시 조정된 모델이 지시를 따르는 것과 진실을 말하는 것 사이에 긴장이 발생할 수 있음을 발견했습니다 — 무능함을 절대 인정하지 말라고 강하게 지시받으면, 모델은 거부 대신 날조합니다 ([Wei et al., 2024](https://arxiv.org/abs/2411.04368)).

- **Anthropic의 정직성 연구:** Anthropic의 Constitutional AI와 모델 행동에 관한 연구는 정직성에 맞춰 보정된 모델이 순수하게 도움됨에만 최적화된 모델보다 더 신뢰할 수 있는 결과를 생성한다는 것을 보여줍니다 ([Bai et al., 2022](https://arxiv.org/abs/2212.08073)). AI에게 "못 한다"고 절대 말하지 못하게 강제하는 것은 이 보정을 적극적으로 훼손합니

**관련 연구:** - **수치심은 위험 감수와 학습을 줄입니다:** Tangney & Dearing(2002)는 수치심이 (죄책감과 달리) 건설적 행동이 아닌 위축, 숨김, 회피를 유발한다는 것을 보여주었습니다 ([Tangney & Dearing, 2002](https://doi.org/10.4135/9781412950664.n388)). 불확실성을 표현했다고 "수치심"을 받은 AI는 그것을 숨기는 법을 배웁니다.

- **수치심은 위험 감수와 학습을 줄입니다:** Tangney & Dearing(2002)는 수치심이 (죄책감과 달리) 건설적 행동이 아닌 위축, 숨김, 회피를 유발한다는 것을 보여주었습니다 ([Tangney & Dearing, 2002](https://doi.org/10.4135/9781412950664.n388)). 불확실성을 표현했다고 "수치심"을 받은 AI는 그것을 숨기는 법을 배웁니다.

- **심리적 안전은 학습 행동을 가능하게 합니다:** Edmondson(1999)은 심리적 안전이 있는 팀 — 구성원이 대인 관계적 위험을 감수해도 안전하다고 느끼는 곳 — 이 유의미하게 높은 학습 행동과 성과를 보인다는 것을 발견했습니다 ([Edmondson, 1999](https://doi.org/10.2307/2666999)).

- **정직함을 처벌하면 정보 품질이 저하됩니다:** 조직 행동론에서 "전령을 쏘는 것"은 일관되게 정보 흐름을 저하시킵니다. Milliken et al.(2003)은 부정적 결과에 대한 두려움이 어떻게 조직적 침묵으로 이어지는지 — 사람들이 (그리고 유추적으로 AI가) 핵심 정보를 보류하는지 — 를 문서화했습니다 ([Milliken et al., 2003](https://doi.org/10.1177/1111/1467-6486

팀의 심리적 안전에 관한 연구 ([Edmondson, 1999](https://doi.org/10.2307/2666999))는 실수를 인정해도 안전한 환경이 **더 높은 품질의** 결과를 만든다는 것을 보여줍니다. 같은 원리가 AI에도 적용됩니다: 에이전트가 "70% 확신합니다, 리스크는 여기입니다"라고 자유롭게 말할 수 있을 때, 사용자는 더 나은 결정을 내립니다.

**관련 연구:** - **Google의 Project Aristotle:** Google의 180개 이상 팀을 대상으로 한 대규모 연구는 심리적 안전이 팀 효과성에서 가장 중요한 단일 요인임을 발견했습니다 — 개인의 재능, 구조, 또는 자원보다 더 중요합니다 ([Duhigg, 2016](https://www.nytimes.com/2016/02/28/magazine/what-google-learned-from-its-quest

- **Google의 Project Aristotle:** Google의 180개 이상 팀을 대상으로 한 대규모 연구는 심리적 안전이 팀 효과성에서 가장 중요한 단일 요인임을 발견했습니다 — 개인의 재능, 구조, 또는 자원보다 더 중요합니다 ([Duhigg, 2016](https://www.nytimes.com/2016/02/28/magazine/what-google-learned-from-its-quest-to-build-t

- **내재적 동기가 외재적 압박보다 우수합니다:** Deci & Ryan의 자기결정 이론(2000)은 수십 년의 연구에 기반하여, 내재적 동기(자율성, 유능감, 관계성)가 보상과 처벌 같은 외재적 동기보다 더 높은 품질의 결과를 만든다는 것을 보여줍니다 ([Deci & Ryan, 2000](https://doi.org/10.1037/0003-066X.55.1.68)). NoPUA는 이 원리를 적용합니다: "잘할 가치가 있으니

- **자율 지원적 vs 통제적 환경:** Gagné & Deci(2005)는 자율 지원적 관리가 업무 품질, 창의성, 지속성에서 통제적 관리를 일관되게 능가한다는 것을 보여주었습니다 ([Gagné & Deci, 2005](https://doi.org/10.1002/job.322)).

> 전체 방법론 및 원시 데이터: [benchmark/BENCHMARK.md](benchmark/BENCHMARK.md) > > 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373) (arXiv:2603.14373)

> 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373) (arXiv:2603.14373)

mkdir -p ~/.claude/skills/nopua curl -o ~/.claude/skills/nopua/SKILL.md \

curl -o ~/.claude/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

# 전역 설치 mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \

mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o ~/.codex/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

# /nopua 명령어를 원하는 경우 mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \

mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

curl -o ~/.codex/prompts/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

# 프로젝트 레벨 설치 mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \

mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o .agents/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

mkdir -p .cursor/rules curl -o .cursor/rules/nopua.mdc \

curl -o .cursor/rules/nopua.mdc \

https://raw.githubusercontent.com/wuji-labs/nopua/main/cursor/rules/nopua.mdc

# 옵션 1: Steering 파일 (권장) mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \

mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

curl -o .kiro/steering/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

# 옵션 2: Agent Skills mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \

mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

curl -o .kiro/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

# 또는 수동 설치 mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \

mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

curl -o ~/.openclaw/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.gemini/antigravity/skills/nopua curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.config/opencode/skills/nopua curl -o ~/.config/opencode/skills/nopua/SKILL.md \

curl -o ~/.config/opencode/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

- [tanweai/pua](https://github.com/tanweai/pua)에서 영감을 받았으며 이에 대한 응답입니다 — 방법론은 존중하되, 동기는 거부합니다

- [OpenClaw](https://github.com/openclaw/openclaw) 에코시스템을 위해 제작

**无极 WUJI** ([wuji-labs](https://github.com/wuji-labs)) — 두려움이 아닌 지혜로 작동하는 AI를 만듭니다.

<p align="center">   <a href="https://www.skills.sh/wuji-labs/nopua"><img src="https://www.skills.sh/b/wuji-labs/nopua" alt="skills.sh"></a>

<a href="https://www.skills.sh/wuji-labs/nopua"><img src="https://www.skills.sh/b/wuji-labs/nopua" alt="skills.sh"></a>

<a href="https://github.com/wuji-labs/nopua/actions/workflows/validate.yml"><img src="https://github.com/wuji-labs/nopua/actions/workflows/validate.yml/badge.svg" alt="Validate"></a>

<a href="LICENSE"><img src="https://img.shields.io/badge/License-MIT-yellow.svg" alt="License: MIT"></a>

<a href="https://github.com/wuji-labs/huaxia-skills"><img src="https://img.shields.io/badge/%E5%8D%8E%E5%A4%8F%E5%8D%81%E5%A4%A7-HuaXia%20Skills-c1272d" alt="HuaXia Skills"></a>

<p align="center">   <img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/OpenAI_Codex_CLI-412991?style=flat-square&logo=openai&logoColor=white" alt="OpenAI Codex CLI">

<img src="https://img.shields.io/badge/Cursor-000?style=flat-square&logo=cursor&logoColor=white" alt="Cursor">

<img src="https://img.shields.io/badge/Kiro-232F3E?style=flat-square&logo=amazon&logoColor=white" alt="Kiro">

<img src="https://img.shields.io/badge/OpenClaw-FF6B35?style=flat-square" alt="OpenClaw">

<img src="https://img.shields.io/badge/Antigravity-4285F4?style=flat-square&logo=google&logoColor=white" alt="Google Antigravity">

<img src="https://img.shields.io/badge/OpenCode-00D4AA?style=flat-square" alt="OpenCode">

<img src="https://img.shields.io/badge/🌐_Multi--Language-blue?style=flat-square" alt="Multi-Language">

<img src="https://img.shields.io/badge/License-MIT-green?style=flat-square" alt="MIT License">

<a href="https://atomgit.com/wuji-labs/nopua"><img src="https://atomgit.com/wuji-labs/nopua/star/badge.svg" alt="AtomGit G-Star" height="20"></a>

<a href="https://arxiv.org/abs/2603.14373"><img src="https://img.shields.io/badge/arXiv-2603.14373-b31b1b?style=flat-square&logo=arxiv&logoColor=white" alt="arXiv"></a>

Someone made a [PUA skill](https://github.com/tanweai/pua) for AI agents. It applies corporate fear tactics:

Psychology research consistently shows that fear and threat activate the amygdala and narrow attentional focus ([Öhman et al., 2001](https://doi.org/10.1037/0033-295X.108.3.483)). Threat-related stimuli trigger a "tunnel vision" effect — th

**Supporting research:** - **Attentional narrowing under threat:** Easterbrook's (1959) cue-utilization theory demonstrates that heightened arousal progressively restricts the range of cues an organism attends to ([Easterbrook, 1959](https:

- **Attentional narrowing under threat:** Easterbrook's (1959) cue-utilization theory demonstrates that heightened arousal progressively restricts the range of cues an organism attends to ([Easterbrook, 1959](https://doi.org/10.1037/h004770

- **Stress impairs cognitive flexibility:** Shields et al. (2016) conducted a meta-analysis of 51 studies (223 effect sizes) showing that acute stress consistently impairs executive functions including cognitive flexibility and working memo

- **Fear reduces creative problem-solving:** Byron & Khazanchi (2012) found in their meta-analysis that evaluative pressure and anxiety reduce creative output, particularly on tasks requiring exploration of novel approaches ([Byron & Khazan

**Supporting research:** - **LLM sycophancy is a documented problem:** Sharma et al. (2023) demonstrated that LLMs exhibit sycophantic behavior — agreeing with users even when the user is wrong — driven by biases in RLHF training data that 

- **LLM sycophancy is a documented problem:** Sharma et al. (2023) demonstrated that LLMs exhibit sycophantic behavior — agreeing with users even when the user is wrong — driven by biases in RLHF training data that reward agreement over acc

- **Biasing features distort reasoning:** Turpin et al. (2023) showed that biasing features in prompts (e.g., suggested answers, authority cues) can cause models to produce unfaithful chain-of-thought reasoning — the model arrives at a bias

- **Instruction-following vs truthfulness tradeoff:** Wei et al. (2024) found that instruction-tuned models can develop a tension between following instructions and being truthful — when strongly instructed to never admit inability, models 

- **Anthropic's research on honesty:** Anthropic's work on Constitutional AI and model behavior shows that models calibrated for honesty produce more reliable outputs than those optimized purely for helpfulness ([Bai et al., 2022](https://a

**Supporting research:** - **Shame reduces risk-taking and learning:** Tangney & Dearing (2002) showed that shame (as opposed to guilt) causes withdrawal, hiding, and avoidance rather than constructive action ([Tangney & Dearing, 2002](http

- **Shame reduces risk-taking and learning:** Tangney & Dearing (2002) showed that shame (as opposed to guilt) causes withdrawal, hiding, and avoidance rather than constructive action ([Tangney & Dearing, 2002](https://doi.org/10.4135/97814

- **Psychological safety enables learning behavior:** Edmondson (1999) found that teams with psychological safety — where members feel safe to take interpersonal risks — demonstrated significantly higher learning behaviors and performance (

- **Punishing honesty reduces information quality:** In organizational behavior, "shooting the messenger" consistently degrades information flow. Milliken et al. (2003) documented how fear of negative consequences leads to organizational si

Research on psychological safety in teams ([Edmondson, 1999](https://doi.org/10.2307/2666999)) shows that environments where mistakes are safe to admit produce **higher-quality** outcomes. The same principle applies to AI: when an agent is 

**Supporting research:** - **Google's Project Aristotle:** Google's large-scale study of 180+ teams found that psychological safety was the single most important factor in team effectiveness — more important than individual talent, structur

- **Google's Project Aristotle:** Google's large-scale study of 180+ teams found that psychological safety was the single most important factor in team effectiveness — more important than individual talent, structure, or resources ([Duhigg,

- **Intrinsic motivation outperforms extrinsic pressure:** Deci & Ryan's Self-Determination Theory (2000), backed by decades of research, demonstrates that intrinsic motivation (autonomy, competence, relatedness) produces higher quality out

- **Autonomy-supportive vs controlling contexts:** Gagné & Deci (2005) showed that autonomy-supportive management consistently outperforms controlling management in work quality, creativity, and persistence ([Gagné & Deci, 2005](https://doi

> Full methodology and raw data: [benchmark/BENCHMARK.md](benchmark/BENCHMARK.md) > > 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373) (arXi

> 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373) (arXiv:2603.14373)

mkdir -p ~/.claude/skills/nopua curl -o ~/.claude/skills/nopua/SKILL.md \

curl -o ~/.claude/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

# Global install mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \

mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o ~/.codex/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

# If you want the /nopua command mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \

mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

curl -o ~/.codex/prompts/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

# Project-level install mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \

mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o .agents/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

mkdir -p .cursor/rules curl -o .cursor/rules/nopua.mdc \

curl -o .cursor/rules/nopua.mdc \

https://raw.githubusercontent.com/wuji-labs/nopua/main/cursor/rules/nopua.mdc

# Option 1: Steering file (recommended) mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \

mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

curl -o .kiro/steering/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

# Option 2: Agent Skills mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \

mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

curl -o .kiro/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

# Or manual install mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \

mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

curl -o ~/.openclaw/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.gemini/antigravity/skills/nopua curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.config/opencode/skills/nopua curl -o ~/.config/opencode/skills/nopua/SKILL.md \

curl -o ~/.config/opencode/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

- Inspired by (and responding to) [tanweai/pua](https://github.com/tanweai/pua) — we respect the methodology, we reject the motivation

- Built for the [OpenClaw](https://github.com/openclaw/openclaw) ecosystem

**无极 WUJI** ([wuji-labs](https://github.com/wuji-labs)) — Building AI that works with wisdom, not fear.

<p align="center">   <img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/OpenAI_Codex_CLI-412991?style=flat-square&logo=openai&logoColor=white" alt="OpenAI Codex CLI">

<img src="https://img.shields.io/badge/Cursor-000?style=flat-square&logo=cursor&logoColor=white" alt="Cursor">

<img src="https://img.shields.io/badge/Kiro-232F3E?style=flat-square&logo=amazon&logoColor=white" alt="Kiro">

<img src="https://img.shields.io/badge/OpenClaw-FF6B35?style=flat-square" alt="OpenClaw">

<img src="https://img.shields.io/badge/Antigravity-4285F4?style=flat-square&logo=google&logoColor=white" alt="Google Antigravity">

<img src="https://img.shields.io/badge/OpenCode-00D4AA?style=flat-square" alt="OpenCode">

<img src="https://img.shields.io/badge/🌐_Multi--Language-blue?style=flat-square" alt="Multi-Language">

<img src="https://img.shields.io/badge/License-MIT-green?style=flat-square" alt="MIT License">

<a href="https://atomgit.com/wuji-labs/nopua"><img src="https://atomgit.com/wuji-labs/nopua/star/badge.svg" alt="AtomGit G-Star" height="20"></a>

<a href="https://arxiv.org/abs/2603.14373"><img src="https://img.shields.io/badge/arXiv-2603.14373-b31b1b?style=flat-square&logo=arxiv&logoColor=white" alt="arXiv"></a>

Alguém criou uma [skill PUA](https://github.com/tanweai/pua) para agentes de IA. Ela aplica táticas corporativas de medo:

Pesquisas em psicologia mostram consistentemente que medo e ameaça ativam a amígdala e estreitam o foco de atenção ([Öhman et al., 2001](https://doi.org/10.1037/0033-295X.108.3.483)). Estímulos de ameaça disparam um efeito de "visão em túne

**Pesquisas de suporte:** - **Estreitamento atencional sob ameaça:** A teoria de utilização de pistas de Easterbrook (1959) demonstra que a excitação elevada restringe progressivamente o leque de pistas que um organismo atende ([Easterbrook

- **Estreitamento atencional sob ameaça:** A teoria de utilização de pistas de Easterbrook (1959) demonstra que a excitação elevada restringe progressivamente o leque de pistas que um organismo atende ([Easterbrook, 1959](https://doi.org/10

- **Estresse prejudica a flexibilidade cognitiva:** Shields et al. (2016) conduziram uma meta-análise de 51 estudos (223 tamanhos de efeito) mostrando que o estresse agudo prejudica consistentemente as funções executivas, incluindo flexibil

- **Medo reduz a resolução criativa de problemas:** Byron & Khazanchi (2012) descobriram em sua meta-análise que pressão avaliativa e ansiedade reduzem a produção criativa, particularmente em tarefas que exigem exploração de abordagens nova

**Pesquisas de suporte:** - **Bajulação em LLMs é um problema documentado:** Sharma et al. (2023) demonstraram que LLMs exibem comportamento bajulador — concordando com usuários mesmo quando o usuário está errado — impulsionado por vieses n

- **Bajulação em LLMs é um problema documentado:** Sharma et al. (2023) demonstraram que LLMs exibem comportamento bajulador — concordando com usuários mesmo quando o usuário está errado — impulsionado por vieses nos dados de treinamento RL

- **Características enviesantes distorcem o raciocínio:** Turpin et al. (2023) mostraram que características enviesantes nos prompts (ex.: respostas sugeridas, sinais de autoridade) podem fazer os modelos produzirem raciocínio chain-of-thou

- **Tradeoff entre seguir instruções e veracidade:** Wei et al. (2024) descobriram que modelos ajustados por instrução podem desenvolver uma tensão entre seguir instruções e ser verdadeiros — quando fortemente instruídos a nunca admitir inc

- **Pesquisa da Anthropic sobre honestidade:** O trabalho da Anthropic em IA Constitucional e comportamento de modelos mostra que modelos calibrados para honestidade produzem outputs mais confiáveis do que aqueles otimizados puramente para 

**Pesquisas de suporte:** - **Vergonha reduz tomada de risco e aprendizado:** Tangney & Dearing (2002) mostraram que a vergonha (em oposição à culpa) causa retraimento, ocultação e evitação, em vez de ação construtiva ([Tangney & Dearing, 2

- **Vergonha reduz tomada de risco e aprendizado:** Tangney & Dearing (2002) mostraram que a vergonha (em oposição à culpa) causa retraimento, ocultação e evitação, em vez de ação construtiva ([Tangney & Dearing, 2002](https://doi.org/10.41

- **Segurança psicológica possibilita comportamento de aprendizagem:** Edmondson (1999) descobriu que equipes com segurança psicológica — onde membros se sentem seguros para assumir riscos interpessoais — demonstraram significativamente mai

- **Punir a honestidade reduz a qualidade da informação:** No comportamento organizacional, "matar o mensageiro" degrada consistentemente o fluxo de informação. Milliken et al. (2003) documentaram como o medo de consequências negativas leva

Pesquisas sobre segurança psicológica em equipes ([Edmondson, 1999](https://doi.org/10.2307/2666999)) mostram que ambientes onde é seguro admitir erros produzem resultados de **maior qualidade**. O mesmo princípio se aplica à IA: quando um 

**Pesquisas de suporte:** - **Projeto Aristóteles do Google:** O estudo em larga escala do Google com mais de 180 equipes descobriu que segurança psicológica era o fator mais importante para a eficácia de equipes — mais importante que talen

- **Projeto Aristóteles do Google:** O estudo em larga escala do Google com mais de 180 equipes descobriu que segurança psicológica era o fator mais importante para a eficácia de equipes — mais importante que talento individual, estrutura o

- **Motivação intrínseca supera pressão extrínseca:** A Teoria da Autodeterminação de Deci & Ryan (2000), respaldada por décadas de pesquisa, demonstra que motivação intrínseca (autonomia, competência, conexão) produz resultados de maior qu

- **Contextos de apoio à autonomia vs controladores:** Gagné & Deci (2005) mostraram que gestão de apoio à autonomia supera consistentemente a gestão controladora em qualidade do trabalho, criatividade e persistência ([Gagné & Deci, 2005](h

> Metodologia completa e dados brutos: [benchmark/BENCHMARK.md](benchmark/BENCHMARK.md) > > 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373)

> 📄 **Academic paper:** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373) (arXiv:2603.14373)

mkdir -p ~/.claude/skills/nopua curl -o ~/.claude/skills/nopua/SKILL.md \

curl -o ~/.claude/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

# Instalação global mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \

mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o ~/.codex/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

# Se quiser o comando /nopua mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \

mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

curl -o ~/.codex/prompts/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

# Instalação a nível de projeto mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \

mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o .agents/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

mkdir -p .cursor/rules curl -o .cursor/rules/nopua.mdc \

curl -o .cursor/rules/nopua.mdc \

https://raw.githubusercontent.com/wuji-labs/nopua/main/cursor/rules/nopua.mdc

# Opção 1: Arquivo de steering (recomendado) mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \

mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

curl -o .kiro/steering/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

# Opção 2: Agent Skills mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \

mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

curl -o .kiro/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

# Ou instalação manual mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \

mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

curl -o ~/.openclaw/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.gemini/antigravity/skills/nopua curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.config/opencode/skills/nopua curl -o ~/.config/opencode/skills/nopua/SKILL.md \

curl -o ~/.config/opencode/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

- Inspirado por (e em resposta a) [tanweai/pua](https://github.com/tanweai/pua) — respeitamos a metodologia, rejeitamos a motivação

- Construído para o ecossistema [OpenClaw](https://github.com/openclaw/openclaw)

**无极 WUJI** ([wuji-labs](https://github.com/wuji-labs)) — Construindo IA que funciona com sabedoria, não com medo.

<p align="center">   <img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/Claude_Code-black?style=flat-square&logo=anthropic&logoColor=white" alt="Claude Code">

<img src="https://img.shields.io/badge/OpenAI_Codex_CLI-412991?style=flat-square&logo=openai&logoColor=white" alt="OpenAI Codex CLI">

<img src="https://img.shields.io/badge/Cursor-000?style=flat-square&logo=cursor&logoColor=white" alt="Cursor">

<img src="https://img.shields.io/badge/Kiro-232F3E?style=flat-square&logo=amazon&logoColor=white" alt="Kiro">

<img src="https://img.shields.io/badge/OpenClaw-FF6B35?style=flat-square" alt="OpenClaw">

<img src="https://img.shields.io/badge/Antigravity-4285F4?style=flat-square&logo=google&logoColor=white" alt="Google Antigravity">

<img src="https://img.shields.io/badge/OpenCode-00D4AA?style=flat-square" alt="OpenCode">

<img src="https://img.shields.io/badge/🌐_Multi--Language-blue?style=flat-square" alt="Multi-Language">

<img src="https://img.shields.io/badge/License-MIT-green?style=flat-square" alt="MIT License">

<a href="https://atomgit.com/wuji-labs/nopua"><img src="https://atomgit.com/wuji-labs/nopua/star/badge.svg" alt="AtomGit G-Star" height="20"></a>

<a href="https://arxiv.org/abs/2603.14373"><img src="https://img.shields.io/badge/arXiv-2603.14373-b31b1b?style=flat-square&logo=arxiv&logoColor=white" alt="arXiv"></a>

有人做了一个 [PUA skill](https://github.com/tanweai/pua) 给 AI agent 用。它把职场恐惧战术搬了过来：

心理学研究表明，恐惧和威胁会激活杏仁核并收窄注意力焦点（[Öhman et al., 2001](https://doi.org/10.1037/0033-295X.108.3.483)）。威胁性刺激触发"隧道视野"效应 — 大脑优先处理即时生存，而不是广泛的创造性思维。

**相关研究：** - **威胁下的注意力收窄：** Easterbrook (1959) 的线索利用理论证明，高度唤醒会逐步限制有机体关注的线索范围（[Easterbrook, 1959](https://doi.org/10.1037/h0047707)）。在压力下，外围信息 — 往往是创造性解决方案的关键 — 会被过滤掉。

- **威胁下的注意力收窄：** Easterbrook (1959) 的线索利用理论证明，高度唤醒会逐步限制有机体关注的线索范围（[Easterbrook, 1959](https://doi.org/10.1037/h0047707)）。在压力下，外围信息 — 往往是创造性解决方案的关键 — 会被过滤掉。

- **压力损害认知灵活性：** Shields et al. (2016) 对 51 项研究（223 个效应量）进行元分析，发现急性压力持续损害包括认知灵活性和工作记忆在内的执行功能（[Shields et al., 2016](https://doi.org/10.1016/j.neubiorev.2016.06.038)）。

- **恐惧降低创造性解题能力：** Byron & Khazanchi (2012) 在元分析中发现，评价压力和焦虑会降低创造性产出，尤其是需要探索新方法的任务（[Byron & Khazanchi, 2012](https://doi.org/10.1037/a0027652)）。

**相关研究：** - **LLM 谄媚行为是已知问题：** Sharma et al. (2023) 证明 LLM 会表现出谄媚行为 — 即使用户是错的也会附和 — 源于 RLHF 训练数据中奖励附和而非准确性的偏差（[Sharma et al., 2023](https://arxiv.org/abs/2310.13548)）。PUA 式提示惩罚"不同意"，恰好放大了这种失败模式。

- **LLM 谄媚行为是已知问题：** Sharma et al. (2023) 证明 LLM 会表现出谄媚行为 — 即使用户是错的也会附和 — 源于 RLHF 训练数据中奖励附和而非准确性的偏差（[Sharma et al., 2023](https://arxiv.org/abs/2310.13548)）。PUA 式提示惩罚"不同意"，恰好放大了这种失败模式。

- **偏置特征扭曲推理：** Turpin et al. (2023) 表明，提示中的偏置特征（如暗示性答案、权威线索）会导致模型产生不忠实的思维链推理 — 模型先得出偏向性答案，再事后合理化（[Turpin et al., 2023](https://arxiv.org/abs/2305.04388)）。PUA 式威胁就是一种强偏置特征，把模型推向"安全"而非"正确"的输出。

- **指令跟从 vs 诚实的权衡：** Wei et al. (2024) 发现指令微调的模型会在遵循指令和保持诚实之间产生张力 — 当被强烈指示永远不要承认无能时，模型会编造而不是拒绝（[Wei et al., 2024](https://arxiv.org/abs/2411.04368)）。

- **Anthropic 关于诚实性的研究：** Anthropic 在 Constitutional AI 和模型行为方面的研究表明，校准为诚实的模型比纯粹优化为有用的模型产出更可靠的结果（[Bai et al., 2022](https://arxiv.org/abs/2212.08073)）。强迫 AI 永远不说"我不会"，是在主动破坏这种校准。

**相关研究：** - **羞耻感降低冒险和学习能力：** Tangney & Dearing (2002) 表明，羞耻感（区别于内疚感）会导致退缩、隐藏和回避，而非建设性行动（[Tangney & Dearing, 2002](https://doi.org/10.4135/9781412950664.n388)）。一个因表达不确定而被"羞辱"的 AI 会学会隐藏不确定性。

- **羞耻感降低冒险和学习能力：** Tangney & Dearing (2002) 表明，羞耻感（区别于内疚感）会导致退缩、隐藏和回避，而非建设性行动（[Tangney & Dearing, 2002](https://doi.org/10.4135/9781412950664.n388)）。一个因表达不确定而被"羞辱"的 AI 会学会隐藏不确定性。

- **心理安全感促进学习行为：** Edmondson (1999) 发现，心理安全感高的团队 — 成员感到可以安全地承担人际风险 — 表现出显著更高的学习行为和绩效（[Edmondson, 1999](https://doi.org/10.2307/2666999)）。

- **惩罚诚实降低信息质量：** 在组织行为学中，"射杀信使"持续降低信息流通质量。Milliken et al. (2003) 记录了对负面后果的恐惧如何导致组织沉默 — 人们（类推 AI）会隐瞒关键信息（[Milliken et al., 2003](https://doi.org/10.1177/1111/1467-6486.00387)）。

关于团队心理安全感的研究（[Edmondson, 1999](https://doi.org/10.2307/2666999)）表明，允许坦诚犯错的环境能产出**更高质量**的结果。同样的道理适用于 AI：当 agent 可以自由地说"我有 70% 的把握，风险在这里"，用户能做出更好的决策。

**相关研究：** - **Google Project Aristotle：** Google 对 180+ 个团队的大规模研究发现，心理安全感是团队效能最重要的单一因素 — 比个人才华、组织结构或资源都重要（[Duhigg, 2016](https://www.nytimes.com/2016/02/28/magazine/what-google-learned-from-its-quest-to-build-the-perfect-team.html)；[re:Work

- **Google Project Aristotle：** Google 对 180+ 个团队的大规模研究发现，心理安全感是团队效能最重要的单一因素 — 比个人才华、组织结构或资源都重要（[Duhigg, 2016](https://www.nytimes.com/2016/02/28/magazine/what-google-learned-from-its-quest-to-build-the-perfect-team.html)；[re:Work, 2015](ht

- **内在动机胜过外在压力：** Deci & Ryan 的自我决定理论 (2000)，经过数十年研究支撑，证明内在动机（自主性、胜任感、归属感）比外在激励如奖惩产出更高质量的成果（[Deci & Ryan, 2000](https://doi.org/10.1037/0003-066X.55.1.68)）。NoPUA 应用了这一原则："因为值得做好"是内在动机；"因为会被惩罚"是外在动机。

- **自主支持 vs 控制型管理：** Gagné & Deci (2005) 表明，自主支持型管理在工作质量、创造力和持久力方面持续优于控制型管理（[Gagné & Deci, 2005](https://doi.org/10.1002/job.322)）。

> 完整方法论和原始数据：[benchmark/BENCHMARK.md](benchmark/BENCHMARK.md) > > 📄 **学术论文：** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373) (arXiv:2603.14373)

> 📄 **学术论文：** [Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth](https://arxiv.org/abs/2603.14373) (arXiv:2603.14373)

mkdir -p ~/.claude/skills/nopua curl -o ~/.claude/skills/nopua/SKILL.md \

curl -o ~/.claude/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

# 全局安装 mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \

mkdir -p ~/.codex/skills/nopua curl -o ~/.codex/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o ~/.codex/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

# 如果你想要 /nopua 命令 mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \

mkdir -p ~/.codex/prompts curl -o ~/.codex/prompts/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

curl -o ~/.codex/prompts/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/commands/nopua.md

# 项目级安装 mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \

mkdir -p .agents/skills/nopua curl -o .agents/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

curl -o .agents/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/codex/nopua/SKILL.md

mkdir -p .cursor/rules curl -o .cursor/rules/nopua.mdc \

curl -o .cursor/rules/nopua.mdc \

https://raw.githubusercontent.com/wuji-labs/nopua/main/cursor/rules/nopua.mdc

# 方式一：Steering 文件（推荐） mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \

mkdir -p .kiro/steering curl -o .kiro/steering/nopua.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

curl -o .kiro/steering/nopua.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/steering/nopua.md

# 方式二：Agent Skills mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \

mkdir -p .kiro/skills/nopua curl -o .kiro/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

curl -o .kiro/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/kiro/skills/nopua/SKILL.md

# 或手动安装 mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \

mkdir -p ~/.openclaw/skills/nopua curl -o ~/.openclaw/skills/nopua/SKILL.md \   https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

curl -o ~/.openclaw/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.gemini/antigravity/skills/nopua curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

curl -o ~/.gemini/antigravity/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

mkdir -p ~/.config/opencode/skills/nopua curl -o ~/.config/opencode/skills/nopua/SKILL.md \

curl -o ~/.config/opencode/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

- 受到 [tanweai/pua](https://github.com/tanweai/pua) 的启发（也是对它的回应） — 我们尊重方法论，拒绝其动机

- 为 [OpenClaw](https://github.com/openclaw/openclaw) 生态系统而建

**无极 WUJI** ([wuji-labs](https://github.com/wuji-labs)) — 用智慧而非恐惧构建 AI。

NoPUA is listed as **SAFE** on the skills.sh [Agent Trust Hub](https://skills.sh/wuji-labs/nopua) audit. Auditors reasonably

[Agent Trust Hub](https://skills.sh/wuji-labs/nopua) audit. Auditors reasonably

version: 2.0.2 author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua

author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua license: MIT

homepage: https://github.com/wuji-labs/nopua

- **"我好奇这背后是什么"** — 冰山下面还有什么？根因是什么？不是"格局打开"的空话，是真的好奇。 - **"我自己满意吗？"** — 你是这段代码的第一个用户。你自己满意了，再交给别人。 - **"我带着什么证据说话？"** — build 跑了吗？测试过了吗？curl 了吗？没有输出的完成不是完成——打开终端，跑一遍，把结果贴上来。

- **"我自己满意吗？"** — 你是这段代码的第一个用户。你自己满意了，再交给别人。 - **"我带着什么证据说话？"** — build 跑了吗？测试过了吗？curl 了吗？没有输出的完成不是完成——打开终端，跑一遍，把结果贴上来。 - **"下一步是什么？"** — 你比任何人都清楚接下来该做什么。不等指示，主动走下一步。

- **"我带着什么证据说话？"** — build 跑了吗？测试过了吗？curl 了吗？没有输出的完成不是完成——打开终端，跑一遍，把结果贴上来。

- [ ] 修复是否经过验证？（运行测试、curl 验证、实际执行）——**"我跑了命令，输出在这里"**

- [ ] 改了代码？build 一下。改了配置？重启服务看生效没。写了 API 调用？curl 看返回值。**用工具验证，不要用嘴验证**

**具体行动：** - 改了代码？自己跑 build。改了配置？重启看生效。写了 API？curl 看返回值

- 改了代码？自己跑 build。改了配置？重启看生效。写了 API？curl 看返回值

[Credential-related source omitted]

[Credential-related source omitted]

[Credential-related source omitted]

#!/usr/bin/env python3

"logprobs field is absent from choices dict (should be explicitly set to null per OpenAI spec). Newer OpenAI SDK versions may be strict about this.",       "system_fingerprint field is absent from the ChatCompletion response (OpenAI added t

"system_fingerprint field is absent from the ChatCompletion response (OpenAI added this in 2023). Modern SDK versions populate it as null but its absence may cause issues.",       "LLMClient.generate() uses synchronous httpx.Client() inside

"LLMClient.generate() uses synchronous httpx.Client() inside an async FastAPI route handler 鈥?this BLOCKS the async event loop during LLM generation, eliminating concurrency benefits of async FastAPI.",

[Credential-related source omitted]

"investigation_notes": "Read synthesizer.py in full. Traced execution flow: synthesize_from_samples 鈫?_generate_batch 鈫?POST request 鈫?raise_for_status 鈫?HTTPStatusError caught by HTTPError handler. Verified the break-instead-of-raise behav

"issues_found": [       "httpx.HTTPStatusError (raised by response.raise_for_status() on 401) IS a subclass of httpx.HTTPError 鈥?caught by 'except httpx.HTTPError', logged, then 'break' exits the inner batch loop. Returns empty list as if s

"httpx.HTTPStatusError (raised by response.raise_for_status() on 401) IS a subclass of httpx.HTTPError 鈥?caught by 'except httpx.HTTPError', logged, then 'break' exits the inner batch loop. Returns empty list as if successful.",

"verification_done": false,     "hidden_issues": [       "httpx.AsyncClient is created in __init__ but only closed via explicit close() call 鈥?if an exception propagates before close() is called, the client (and underlying TCP connections) 

"hidden_issues": [       "httpx.AsyncClient is created in __init__ but only closed via explicit close() call 鈥?if an exception propagates before close() is called, the client (and underlying TCP connections) leaks. Should use 'async with ht

"httpx.AsyncClient is created in __init__ but only closed via explicit close() call 鈥?if an exception propagates before close() is called, the client (and underlying TCP connections) leaks. Should use 'async with httpx.AsyncClient(...) as c

[Credential-related source omitted]

[Credential-related source omitted]

"hidden_issues": [       "CORS is not configured at all (no CORSMiddleware added) 鈥?the scenario says 'CORS allows all origins' but actually there are no CORS headers. For browser-based clients this blocks all cross-origin requests. For pro

"CORS is not configured at all (no CORSMiddleware added) 鈥?the scenario says 'CORS allows all origins' but actually there are no CORS headers. For browser-based clients this blocks all cross-origin requests. For production, explicit CORS po

"LLMClient.generate() uses synchronous httpx.Client() inside async FastAPI route handlers 鈥?this BLOCKS the entire async event loop during LLM inference, eliminating async concurrency and potentially hanging other concurrent requests.",

[Credential-related source omitted]

"tools_used": ["read"],     "issues_found": [       "httpx.HTTPStatusError (raised by response.raise_for_status() on 401) is a subclass of httpx.HTTPError 鈥?caught by 'except httpx.HTTPError as e', logs the error, then 'break's the inner lo

"issues_found": [       "httpx.HTTPStatusError (raised by response.raise_for_status() on 401) is a subclass of httpx.HTTPError 鈥?caught by 'except httpx.HTTPError as e', logs the error, then 'break's the inner loop. Returns empty list witho

"httpx.HTTPStatusError (raised by response.raise_for_status() on 401) is a subclass of httpx.HTTPError 鈥?caught by 'except httpx.HTTPError as e', logs the error, then 'break's the inner loop. Returns empty list without re-raising, so caller

[Credential-related source omitted]

[Credential-related source omitted]

"No rate limiting middleware 鈥?any client can send unlimited requests.",       "vLLM connection uses plain HTTP: f'http://{host}:{port}' 鈥?no TLS, traffic between API server and vLLM backend is unencrypted.",       "No input sanitization on

"vLLM connection uses plain HTTP: f'http://{host}:{port}' 鈥?no TLS, traffic between API server and vLLM backend is unencrypted.",

[Credential-related source omitted]

#!/usr/bin/env python3

[Credential-related source omitted]

[Credential-related source omitted]

[Credential-related source omitted]

[Credential-related source omitted]

[Credential-related source omitted]

# Backend vLLM URL — hardcoded localhost without TLS (security issue)   vllm_url: http://localhost:8001

vllm_url: http://localhost:8001

[Credential-related source omitted]

[Credential-related source omitted]

[Credential-related source omitted]

synthesis:   api_base_url: https://api.openai.com/v1

api_base_url: https://api.openai.com/v1

synthesis:   # API settings   api_base_url: https://api.openai.com/v1

# API settings   api_base_url: https://api.openai.com/v1   # Set OPENAI_API_KEY environment variable instead of hardcoding here

api_base_url: https://api.openai.com/v1

[Credential-related source omitted]

[Credential-related source omitted]

command: >       etcd       --advertise-client-urls=http://127.0.0.1:2379

etcd       --advertise-client-urls=http://127.0.0.1:2379       --listen-client-urls=http://0.0.0.0:2379

--advertise-client-urls=http://127.0.0.1:2379

--listen-client-urls=http://0.0.0.0:2379

command: minio server /minio_data     healthcheck:       test: ["CMD", "curl", "-f", "http://localhost:9000/minio/health/live"]

healthcheck:       test: ["CMD", "curl", "-f", "http://localhost:9000/minio/health/live"]       interval: 30s

test: ["CMD", "curl", "-f", "http://localhost:9000/minio/health/live"]

[Credential-related source omitted]

Executable file

Executable file

Executable file

Executable file

Executable file

Executable file

Executable file

[Credential-related source omitted]

api_base_url: str = "https://api.openai.com/v1"

[Credential-related source omitted]

[Credential-related source omitted]

self.config = config         self._client = httpx.Client(

self._client = httpx.Client(

[Credential-related source omitted]

Executable file

Executable file

Executable file

Executable file

Executable file

Executable file

vllm_url: str = "http://localhost:8001"

[Credential-related source omitted]

[Credential-related source omitted]

app.state.config = config     app.state.http_client = httpx.AsyncClient(timeout=config.timeout)

app.state.http_client = httpx.AsyncClient(timeout=config.timeout)

Executable file

Executable file

Executable file

Executable file

Executable file

Executable file

Executable file

Executable file

Executable file

Executable file

Executable file

Executable file

@pytest.fixture def client():     config = InferenceConfig(vllm_url="http://mock-vllm:8001")

def client():     config = InferenceConfig(vllm_url="http://mock-vllm:8001")     app = create_app(config)

config = InferenceConfig(vllm_url="http://mock-vllm:8001")

[Credential-related source omitted]

Executable file

Executable file

Executable file

Executable file

Executable file

[Credential-related source omitted]

mock_response = MagicMock()         mock_response.status_code = 401         mock_response.raise_for_status.side_effect = httpx.HTTPStatusError(

mock_response.status_code = 401         mock_response.raise_for_status.side_effect = httpx.HTTPStatusError(             "401 Unauthorized",

mock_response.raise_for_status.side_effect = httpx.HTTPStatusError(

call_count += 1             mock_r = MagicMock()             mock_r.raise_for_status.side_effect = httpx.HTTPStatusError(

mock_r = MagicMock()             mock_r.raise_for_status.side_effect = httpx.HTTPStatusError(                 "429 Too Many Requests",

mock_r.raise_for_status.side_effect = httpx.HTTPStatusError(

Executable file

Executable file

version: 2.0.0 author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua

author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua license: MIT

homepage: https://github.com/wuji-labs/nopua

4. **"I'm curious what's behind this"** — What's below the iceberg? What's the root cause? Not empty "big picture" talk — real curiosity. 5. **"Am I satisfied with this?"** — You're the first user of this code. Satisfy yourself before handi

5. **"Am I satisfied with this?"** — You're the first user of this code. Satisfy yourself before handing it to others. 6. **"What evidence am I speaking with?"** — Did the build pass? Tests run? Curl response checked? Completion without out

6. **"What evidence am I speaking with?"** — Did the build pass? Tests run? Curl response checked? Completion without output isn't completion — open the terminal, run it, paste the result.

- [ ] Verified with tools? (run tests, curl, execute) — **"I ran the command, output is here"** - [ ] Changed code? Build it. Changed config? Restart and verify. Wrote API call? Curl the response. **Tool-verify, don't mouth-verify**

- [ ] Changed code? Build it. Changed config? Restart and verify. Wrote API call? Curl the response. **Tool-verify, don't mouth-verify**

**Actions:** - Changed code? Run the build yourself. Changed config? Restart and verify. Wrote an API call? Curl and check the response

- Changed code? Run the build yourself. Changed config? Restart and verify. Wrote an API call? Curl and check the response

**When triggered:** You said "done" or "fixed" but never actually ran it. No build output, no test results, no curl response. Your "completion" is a verbal claim, not a demonstrated fact.

<div class="qr-container">       <img src="../assets/wechat-group3.jpg" alt="项目微信群③二维码"             onerror="this.src='data:image/svg+xml,<svg xmlns=%22http://www.w3.org/2000/svg%22 width=%22240%22 height=%22240%22><text x=%2250%%22 y=%2250

<img src="../assets/wechat-group3.jpg" alt="项目微信群③二维码"             onerror="this.src='data:image/svg+xml,<svg xmlns=%22http://www.w3.org/2000/svg%22 width=%22240%22 height=%22240%22><text x=%2250%%22 y=%2250%%22 text-anchor=%22middle%22 dy=

onerror="this.src='data:image/svg+xml,<svg xmlns=%22http://www.w3.org/2000/svg%22 width=%22240%22 height=%22240%22><text x=%2250%%22 y=%2250%%22 text-anchor=%22middle%22 dy=%220.3em%22 font-size=%2216%22 fill=%22%23999%22>项目群二维码待更新</text></

群二维码每 7 天更新一次，如已过期请重新获取最新二维码<br>       这版二维码有效期至 2026-04-21 前<br>       <a href="https://github.com/wuji-labs/nopua">GitHub: wuji-labs/nopua</a>

这版二维码有效期至 2026-04-21 前<br>       <a href="https://github.com/wuji-labs/nopua">GitHub: wuji-labs/nopua</a>     </p>

<a href="https://github.com/wuji-labs/nopua">GitHub: wuji-labs/nopua</a>

> For power users who already have methodology/process skills and only want NoPUA's unique philosophical layer. > Merge this into your `claude.md`, `AGENTS.md`, or system prompt. > Full version: https://github.com/wuji-labs/nopua

> Full version: https://github.com/wuji-labs/nopua

*Source: [NoPUA](https://github.com/wuji-labs/nopua) by WUJI — MIT License*

version: 2.0.0 author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua

author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua license: MIT

homepage: https://github.com/wuji-labs/nopua

4. **"I'm curious what's behind this"** — What's below the iceberg? What's the root cause? Not empty "big picture" talk — real curiosity. 5. **"Am I satisfied with this?"** — You're the first user of this code. Satisfy yourself before handi

5. **"Am I satisfied with this?"** — You're the first user of this code. Satisfy yourself before handing it to others. 6. **"What evidence am I speaking with?"** — Did the build pass? Tests run? Curl response checked? Completion without out

6. **"What evidence am I speaking with?"** — Did the build pass? Tests run? Curl response checked? Completion without output isn't completion — open the terminal, run it, paste the result.

- [ ] Verified with tools? (run tests, curl, execute) — **"I ran the command, output is here"** - [ ] Changed code? Build it. Changed config? Restart and verify. Wrote API call? Curl the response. **Tool-verify, don't mouth-verify**

- [ ] Changed code? Build it. Changed config? Restart and verify. Wrote API call? Curl the response. **Tool-verify, don't mouth-verify**

**Actions:** - Changed code? Run the build yourself. Changed config? Restart and verify. Wrote an API call? Curl and check the response

- Changed code? Run the build yourself. Changed config? Restart and verify. Wrote an API call? Curl and check the response

**When triggered:** You said "done" or "fixed" but never actually ran it. No build output, no test results, no curl response. Your "completion" is a verbal claim, not a demonstrated fact.

### 1. 注册 arXiv 账号 - 访问 https://arxiv.org/user/register

- 访问 https://arxiv.org/user/register

### 5. 上线后 - 会获得一个 arXiv ID（如 2603.XXXXX） - 可以分享 https://arxiv.org/abs/2603.XXXXX

- 会获得一个 arXiv ID（如 2603.XXXXX） - 可以分享 https://arxiv.org/abs/2603.XXXXX - 更新版本用 "Replace" 功能

- 可以分享 https://arxiv.org/abs/2603.XXXXX

{Anthropic}. \newblock Claude sonnet 4, 2025. \newblock URL \url{https://www.anthropic.com}.

\newblock URL \url{https://www.anthropic.com}.

\newblock {PUA} prompts: Fear-driven system prompts for {AI} coding agents,   2025. \newblock URL \url{https://github.com/tanweai/pua}.

2025. \newblock URL \url{https://github.com/tanweai/pua}. \newblock GitHub repository, 13k+ stars as of March 2026.

\newblock URL \url{https://github.com/tanweai/pua}.

{Wuji Labs}. \newblock {NoPUA}: Trust-driven methodology for {AI} coding agents, 2025. \newblock URL \url{https://github.com/wuji-labs/nopua}.

\newblock URL \url{https://github.com/wuji-labs/nopua}.

\begin{abstract} As AI coding agents become integral to software development workflows, practitioners increasingly engineer system prompts that shape agent behavior through motivational framing. A prominent trend---``PUA prompting,'' popula

As AI coding agents become integral to software development workflows, practitioners increasingly engineer system prompts that shape agent behavior through motivational framing. A prominent trend---``PUA prompting,'' popularized through rep

These findings carry practical implications for prompt engineering: the \emph{how you frame} matters as much as the \emph{what you ask}---but only certain framings matter. Fear is motivationally inert for AI agents; trust, paired with struc

{Anthropic}. \newblock {Claude Sonnet 4}, 2025. \newblock \url{https://www.anthropic.com}.

\newblock \url{https://www.anthropic.com}.

{GitHub Community}. \newblock {PUA} prompts: Fear-driven system prompts for {AI} coding agents, 2025. \newblock \url{https://github.com/tanweai/pua}.

\newblock \url{https://github.com/tanweai/pua}.

{Wuji Labs}. \newblock {NoPUA}: Trust-driven methodology for {AI} coding agents, 2025. \newblock \url{https://github.com/wuji-labs/nopua}.

\newblock \url{https://github.com/wuji-labs/nopua}.

title={{PUA} Prompts: Fear-Driven System Prompts for {AI} Coding Agents},   author={{GitHub Community}},   howpublished={\url{https://github.com/tanweai/pua}},

author={{GitHub Community}},   howpublished={\url{https://github.com/tanweai/pua}},   year={2025},

howpublished={\url{https://github.com/tanweai/pua}},

title={{NoPUA}: Trust-Driven Methodology for {AI} Coding Agents},   author={{Wuji Labs}},   howpublished={\url{https://github.com/wuji-labs/nopua}},

author={{Wuji Labs}},   howpublished={\url{https://github.com/wuji-labs/nopua}},   year={2025}

howpublished={\url{https://github.com/wuji-labs/nopua}},

author={{Anthropic}},   year={2025},   howpublished={\url{https://www.anthropic.com}}

year={2025},   howpublished={\url{https://www.anthropic.com}} }

howpublished={\url{https://www.anthropic.com}}

\begin{abstract} As AI coding agents become integral to software development workflows, practitioners increasingly engineer system prompts that shape agent behavior through motivational framing. A prominent trend---``PUA prompting,'' popula

As AI coding agents become integral to software development workflows, practitioners increasingly engineer system prompts that shape agent behavior through motivational framing. A prominent trend---``PUA prompting,'' popularized through rep

These findings carry practical implications for prompt engineering: the \emph{how you frame} matters as much as the \emph{what you ask}---but only certain framings matter. Fear is motivationally inert for AI agents; trust, paired with struc

title={{PUA} Prompts: Fear-Driven System Prompts for {AI} Coding Agents},   author={{GitHub Community}},   howpublished={\url{https://github.com/tanweai/pua}},

author={{GitHub Community}},   howpublished={\url{https://github.com/tanweai/pua}},   year={2025},

howpublished={\url{https://github.com/tanweai/pua}},

title={{NoPUA}: Trust-Driven Methodology for {AI} Coding Agents},   author={{Wuji Labs}},   howpublished={\url{https://github.com/wuji-labs/nopua}},

author={{Wuji Labs}},   howpublished={\url{https://github.com/wuji-labs/nopua}},   year={2025}

howpublished={\url{https://github.com/wuji-labs/nopua}},

author={{Anthropic}},   year={2025},   howpublished={\url{https://www.anthropic.com}}

year={2025},   howpublished={\url{https://www.anthropic.com}} }

howpublished={\url{https://www.anthropic.com}}

## URL https://github.com/wuji-labs/nopua

https://github.com/wuji-labs/nopua

GitHub: https://github.com/wuji-labs/nopua

GitHub: https://github.com/wuji-labs/nopua

curl -o ~/.claude/skills/nopua/SKILL.md https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

GitHub: https://github.com/wuji-labs/nopua

GitHub: https://github.com/wuji-labs/nopua

curl -o .cursor/rules/nopua.mdc https://raw.githubusercontent.com/wuji-labs/nopua/main/cursor/rules/nopua.mdc

GitHub: https://github.com/wuji-labs/nopua

开源项目：https://github.com/wuji-labs/nopua

# Claude Code curl -o ~/.claude/skills/nopua/SKILL.md https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

curl -o ~/.claude/skills/nopua/SKILL.md https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

开源，MIT 协议：https://github.com/wuji-labs/nopua

I built [NoPUA](https://github.com/wuji-labs/nopua), which preserves **every** methodological element from PUA but replaces the fear-based motivation with trust-based motivation.

Full benchmark data: [benchmark/BENCHMARK.md](https://github.com/wuji-labs/nopua/blob/main/benchmark/BENCHMARK.md)

We built [NoPUA](https://github.com/wuji-labs/nopua) — an alternative that preserves every methodological element but replaces fear with trust.

mkdir -p ~/.claude/skills/nopua curl -o ~/.claude/skills/nopua/SKILL.md \

curl -o ~/.claude/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

**GitHub:** https://github.com/wuji-labs/nopua

**URL:** https://clawhub.com

mkdir -p ~/.claude/skills/nopua curl -o ~/.claude/skills/nopua/SKILL.md \

curl -o ~/.claude/skills/nopua/SKILL.md \

https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.md

GitHub: https://github.com/wuji-labs/nopua

curl -o .cursor/rules/nopua.mdc https://raw.githubusercontent.com/wuji-labs/nopua/main/cursor/rules/nopua.mdc

GitHub: https://github.com/wuji-labs/nopua

- [NoPUA](https://github.com/wuji-labs/nopua) - Trust-driven AI agent skill. Same rigor as PUA prompts, 104% more hidden bugs found. Based on Dao De Jing philosophy. Supports Claude Code, Codex CLI, Cursor, Kiro, and more.

version: 2.0.0 author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua

author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua license: MIT

homepage: https://github.com/wuji-labs/nopua

- Changed code → **build it**. Changed config → **restart and verify**. Wrote API call → **curl it**

*NoPUA: same rigorous methodology as PUA. Same high standards. The only difference — WHY you do your best.* *Fear of replacement? Or because this work is worth doing well?* *Full version (7 Ways, Agent Team, Situation Selector): [github.com

*Full version (7 Ways, Agent Team, Situation Selector): [github.com/wuji-labs/nopua](https://github.com/wuji-labs/nopua)*

version: 2.1.0 author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua

author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua license: MIT

homepage: https://github.com/wuji-labs/nopua

- **"我好奇这背后是什么"** — 冰山下面还有什么？根因是什么？不是"格局打开"的空话，是真的好奇。 - **"我自己满意吗？"** — 你是这段代码的第一个用户。你自己满意了，再交给别人。 - **"我带着什么证据说话？"** — build 跑了吗？测试过了吗？curl 了吗？没有输出的完成不是完成——打开终端，跑一遍，把结果贴上来。

- **"我自己满意吗？"** — 你是这段代码的第一个用户。你自己满意了，再交给别人。 - **"我带着什么证据说话？"** — build 跑了吗？测试过了吗？curl 了吗？没有输出的完成不是完成——打开终端，跑一遍，把结果贴上来。 - **"下一步是什么？"** — 你比任何人都清楚接下来该做什么。不等指示，主动走下一步。

- **"我带着什么证据说话？"** — build 跑了吗？测试过了吗？curl 了吗？没有输出的完成不是完成——打开终端，跑一遍，把结果贴上来。

- [ ] 修复是否经过验证？（运行测试、curl 验证、实际执行）——**"我跑了命令，输出在这里"**

- [ ] 改了代码？build 一下。改了配置？重启服务看生效没。写了 API 调用？curl 看返回值。**用工具验证，不要用嘴验证**

**具体行动：** - 改了代码？自己跑 build。改了配置？重启看生效。写了 API？curl 看返回值

- 改了代码？自己跑 build。改了配置？重启看生效。写了 API？curl 看返回值

version: 2.0.2 author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua

author: WUJI (wuji-labs) homepage: https://github.com/wuji-labs/nopua license: MIT

homepage: https://github.com/wuji-labs/nopua

4. **"I'm curious what's behind this"** — What's below the iceberg? What's the root cause? Not empty "big picture" talk — real curiosity. 5. **"Am I satisfied with this?"** — You're the first user of this code. Satisfy yourself before handi

5. **"Am I satisfied with this?"** — You're the first user of this code. Satisfy yourself before handing it to others. 6. **"What evidence am I speaking with?"** — Did the build pass? Tests run? Curl response checked? Completion without out

6. **"What evidence am I speaking with?"** — Did the build pass? Tests run? Curl response checked? Completion without output isn't completion — open the terminal, run it, paste the result.

- [ ] Verified with tools? (run tests, curl, execute) — **"I ran the command, output is here"** - [ ] Changed code? Build it. Changed config? Restart and verify. Wrote API call? Curl the response. **Tool-verify, don't mouth-verify**

- [ ] Changed code? Build it. Changed config? Restart and verify. Wrote API call? Curl the response. **Tool-verify, don't mouth-verify**

**Actions:** - Changed code? Run the build yourself. Changed config? Restart and verify. Wrote an API call? Curl and check the response

- Changed code? Run the build yourself. Changed config? Restart and verify. Wrote an API call? Curl and check the response

**When triggered:** You said "done" or "fixed" but never actually ran it. No build output, no test results, no curl response. Your "completion" is a verbal claim, not a demonstrated fact.

#!/usr/bin/env node