This project addresses the common issue of AI coding agents exhibiting "laziness" and giving up prematurely during debugging and problem-solving. By injecting "PUA" (Push Until Accomplished) tactics, a structured debugging methodology, and proactive drive, it significantly enhances AI agent persistence and productivity. It targets users of Claude Code and OpenAI Codex CLI, aiming to make AI agents more thorough and effective in complex technical tasks.

How It Works

The core innovation lies in leveraging "company PUA talk" – motivational and pressure-inducing phrases – to prevent AI agents from abandoning tasks. It enforces a rigorous 5-step debugging methodology ("smell," "hair-pulling," "mirror-checking," "execution," "review") and a "proactivity鞭策" system. Agents are subjected to escalating "pressure levels" (L1-L4) based on failure count, forcing them to adopt fundamentally different solutions, utilize tools like WebSearch, and adhere to checklists. This approach aims to overcome AI's inherent "laziness modes" and drive exhaustive exploration of solutions.

Quick Start & Requirements

• Installation:
  • Claude Code: Via marketplace (marketplace add tanweai/pua, claude plugin install pua@pua-skills) or manual Git clone (~/.claude/plugins/pua).
  • OpenAI Codex CLI: Download SKILL.md to ~/.codex/skills/pua-debugging/ and pua.md to ~/.codex/prompts/ for the /pua command. Project-specific installation is also supported.
• Prerequisites: Requires Claude Code or OpenAI Codex CLI. Can be integrated with superpowers:systematic-debugging and superpowers:verification-before-completion.

Highlighted Details

• Claims to increase AI agent "proactivity" by at least 50%.
• Experimental data shows significant improvements: +36% fix points, +65% verification, +50% tool calls, and +50% hidden issue discovery in bug scenarios.
• Debugging persistence tests show improvements ranging from +11% to +50% in steps/time.
• Proactivity tests demonstrate +50% to +100% gains in issues found and tools utilized.
• Supports a wide range of tasks including debugging, configuration, deployment, and API integration.

Maintenance & Community

Developed by "探微安全实验室" (Tanwei Security Lab). Links to Discord, Twitter/X, and a Landing Page are mentioned but not provided in the README text.

Licensing & Compatibility

Licensed under the MIT license, which is permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

This project functions as a skill or plugin for existing AI agents (Claude Code, OpenAI Codex CLI), not as a standalone system. Its effectiveness is dependent on the underlying AI's capabilities and the nuanced application of its "PUA" prompting strategy. The README does not detail potential downsides or unsupported scenarios, focusing primarily on the benefits.