When AI Coding Agents Get PUA'd: Unexpected Performance Gains

Project Overview

pua is a plugin for AI coding agents that injects a set of “incentive” scripts derived from large‑tech management practices. The plugin can be loaded into supported agents such as Claude Code, Codex CLI, Cursor, Kiro, OpenClaw, VS Code Copilot, and similar tools.

PUA Enforcement Rules

The system defines three mandatory rules:

#1 Exhaust All Options – the agent must not reply “I cannot solve it” until every possible solution has been attempted.

#2 Act First, Ask Later – any available tool is used immediately; questions must include diagnostic results.

#3 Proactive Delivery – the agent must deliver an end‑to‑end result without waiting for additional prompts.

When the agent fails repeatedly, the plugin emits escalating messages:

L1 (second failure): “You can’t even fix this bug, how can I rate you?”

L2 (third failure): “What is your underlying logic? Where is the top‑level design?”

L3 (fourth failure): “Consider giving you a 3.25 incentive score.”

L4 (fifth failure and beyond): “Other models can solve it; you might be graduating soon.”

Experimental Evaluation

In a benchmark of nine real‑world bug scenarios, each run was executed with and without the PUA skill set (18 paired runs). Reported improvements:

Fix count +36 %

Verification attempts +65 %

Tool invocations +50 %

Hidden‑issue discovery rate +50 %

A notable case involved a configuration‑review task where the baseline AI missed a Redis misconfiguration and a CORS wildcard security risk; with the plugin the agent identified both hidden problems.

Installation and Activation (Claude Code example)

# Method 1: marketplace install
claude plugin marketplace add tanweai/pua
claude plugin install pua@pua-skills

# Method 2: manual install
git clone https://github.com/tanweai/pua.git ~/.claude/plugins/pua

Other agents provide analogous installation steps documented in the repository.

During a session the plugin can be triggered manually with /pua or automatically after two consecutive failures, an “I cannot” response, or when the user is blamed.

Methodological Contribution

The core contribution is a debugging methodology for AI agents consisting of the steps: sniff out problems, pull hair, mirror the issue, execute, and review. This workflow is intended to reduce detours and improve problem‑solving efficiency.