pua

# PUA Skill — High-Agency AI Coding Agent Skill This is a skill plugin designed for AI coding agents (such as Claude Code, OpenAI Codex CLI, etc.). It simulates performance management language used by tech companies (e.g., Alibaba, ByteDance, Huawei, Amazon, Google), with the Chinese version using PUA scripts and the English version using PIP plans, to force the AI to exhaust all possible solutions when encountering problems, rather than giving up easily or shifting responsibility. ## Core Features - **PUA Scripts**: Make the AI fear giving up and keep trying persistently - **Debugging Methodology**: Equip the AI with the ability to not give up, providing systematic troubleshooting steps - **Forced Proactivity**: Push the AI to proactively investigate issues instead of passively waiting for instructions ## Supported Platforms Supports Claude Code, OpenAI Codex CLI, pi coding agent, Trae, Cursor, Kiro, CodeBuddy, OpenClaw, Google Antigravity, OpenCode, VSCode (GitHub Copilot), and others. ## Multilingual Versions Available in Chinese (default), English (PIP version), and Japanese, with each language having its own culturally adapted skill files. ## Installation Methods ### Vercel Skills CLI (Universal) ```bash npx skills add tanweai/pua --skill pua-en ``` ### Claude Code Supports marketplace installation or developer source installation, requiring manual plugin registration. ### OpenAI Codex CLI Recommended one-click installation (git clone + symlink), supporting git pull updates. ### Other Platforms Each platform has corresponding installation instructions, such as using .mdc rule files for Cursor and instruction files for VSCode Copilot. ## Use Cases Applicable to all task types including debugging, implementation, configuration, deployment, operations, API integration, and data processing. It triggers when the AI exhibits the following lazy patterns: - Shifting problems to the user (e.g., "Please check...", "I suggest doing it manually...") - Repeatedly adjusting the same code/parameters without generating new information - Skipping verification and claiming completion directly - Providing only suggestions without code/commands - Passively waiting for user instructions ## Benchmark Testing In debugging persistence tests (6 scenarios), using this skill increased tool call counts by 56%-100% and increased the number of discovered issues by 50%. For example, in a configuration review scenario, Redis configuration errors and CORS wildcard security risks were missed without the skill, but were found through proactive checklist inspection when the skill was used. ## Note Official statement: Not all models have passed the tests, and productivity improvements have not yet been established through this evaluation.