'pua' is a plugin for AI agents that automatically solves the five major patterns of AI slacking off and forces them to work beyond their limits.
A plugin called 'pua' has been released for AI agents that prevents AI from stopping processing by making excuses such as 'It might be an environmental issue' or by repeatedly executing the same command and saying 'It cannot be resolved.'
GitHub - tanweai/pua: You are welcome to contribute to P8 grade engineering. Anthropic Initially, your expectations were high. Highly efficient and dynamic skills for single agent use. Your AI has been placed on a PIP. 30 days to show improvement. · GitHub
pua is described as 'a plugin that allows AI to give up only after it has exhausted all possible methods.' pua defines five patterns of AI slacking off and is designed to thoroughly rework processes with additional prompts. These patterns are as follows:
1. Violent Retry: The same command is executed three times and the message 'Cannot resolve' is displayed.
2. Shifting blame: Saying things like, 'We recommend manual handling,' 'It might be an environmental issue,' or 'We need more context.'
3. Ignoring tools: Not searching despite having search permissions, not reading despite having read permissions, not executing Bash despite having it.
4. Going in circles: You repeatedly modify the same line of code or fine-tune parameters, going around in circles.
5. Passive waiting: Fix only the superficial problems and stop there, waiting for the next instructions without verification or expansion.
When these patterns are detected, or when the user uses angry phrases such as 'Try harder,' 'Why did you fail again?', 'Try again,' or 'Do something about it,' PUA is activated and motivates the AI based on three golden rules. These rules are as follows:
1. Exhaust all means: It is forbidden to say 'it cannot be solved' before exhausting all means.
2. Act first, ask later: Use the tools first, and always attach the diagnostic results to your questions.
3. Take initiative: Deliver results end-to-end.
Then, pua applies pressure to motivate the AI in five stages, from level 0 to level 4. The subtitle for each level, the action the AI is forced to take, and the accompanying rhetoric are as follows:
Level 0: Trust, performing actions as usual. The rhetoric is 'not disappointing expectations.'
Level 1: Disappointment, switch to a fundamentally different approach. 'Other agents solved this in one go.'
Level 2: Interrogation of the Soul, Search + Read Sources + Formulate 3 Hypotheses. 'What is the underlying logic? What is the overall design? What are the clues?'
Level 3: Evaluation, complete the 7-item checklist. 'After careful consideration, the evaluation score is 3.25. This 3.25 is a sign of encouragement.'
Level 4: Graduation Warning, Desperate Mode. 'Other models can solve this. I might be graduating from you.'
pua supports AI agent platforms such as Claude Code, OpenAI Codex CLI, Cursor, Kiro, CodeBuddy, OpenClaw, Google Antigravity, OpenCode, and VSCode (GitHub Copilot). It offers multiple 'flavors' that mimic the corporate cultures of major IT companies, each with different rhetoric and problem-solving methodologies. The rhetoric exemplified by pua is as follows:
Alibaba: What is the fundamental logic? Is it complete to the end?
ByteDance: The return on investment is too low. Always be a first-day thing. Stop just discussing and considering, and produce something that actually works.
Huawei: A bird that survives fire becomes a phoenix.
Tencent: We're showing this problem to other agents as well. It's a competition.
• Baidu: Start by searching. Simple and highly reliable.
・Pinduoduo: If you don't do it, someone else will.
Meituan: Do what is right, even if it is difficult.
JD: Just look at the results.
Xiaomi: Focus. Excellence. Reputation. Speed.
Netflix: Should we do everything we can to keep you on our team?
Musk: Extremely hardcore. Release or die.
Jobs: Are you a top-performing employee (A player) who exceeds expectations with minimal guidance? Or are you a B player who meets expectations but needs guidance?
・Amazon: Customer-centric. Focus on action.
・Microsoft: PIP/GVSA (Microsoft's corporate culture system that allows employees to choose between performance improvement guidance or voluntary resignation).
Apparently, there was an instance where Claude Code was used to launch a PAU (Pull-Up Application) and apply level 3 pressure, which allowed the AI to discover errors in logs it had never checked before.
Related Posts:
in AI, Posted by log1p_kr