How the PUA Plugin Boosts AI Coding Success by 36%

The PUA plugin on GitHub watches Claude Code’s output, detects repeated failures, and forces the AI to keep working until the problem is solved.

1. Five typical AI coding slacking patterns

Violent retry – repeat the same command and give up after failure.

Passing the buck – reply with “please handle manually”.

Tool idle – can perform WebSearch but chooses not to.

Procrastination – fine‑tune repeatedly without progress.

Passive waiting – stop after a surface fix, no deeper investigation.

PUA’s core idea is to increase pressure as the failure count rises until the task is completed.

2. Pressure escalation system

L1 (second failure) – “gentle disappointment”, switch to a fundamentally different solution.

L2 (third failure) – “soul interrogation”, perform WebSearch and read source code.

L3 (fourth failure) – “361 assessment”, complete a seven‑item checklist.

L4 (fifth and beyond) – “graduation warning”, enter “panic mode”.

Triggers are automatic after consecutive failures or when the AI says it cannot solve the problem, and manual via the /pua command.

3. Three iron rules

Exhaust everything – do not claim inability before all alternatives are tried.

Act first, ask later – tools must be used before asking, and queries must include diagnostic results.

Proactive delivery – end‑to‑end execution without waiting for prompts.

4. Five‑step debugging methodology (adapted from Alibaba’s “three‑axe” approach)

Smell – list all attempts and identify common failure patterns.

Pull hair – read error messages word‑by‑word, then WebSearch, read source, verify environment.

Mirror – check whether the direction has been tried before and eliminate the simplest possibilities.

Execute – the new solution must be fundamentally different, not just a faster version of the same path.

Review – document what was solved and check for related issues.

5. Installation

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

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

After installation, invoke manually with /pua or let the plugin intervene automatically after consecutive failures.

6. Empirical results

In nine real‑world scenarios with 18 control groups, the plugin increased the number of fixed issues by 36 %, verification attempts by 65 %, tool invocations by 50 %, and hidden‑issue discovery by 50 %.

The authors conclude that the limitation lies not in the model but in the management of the debugging process.