Claude Code Pua

pua — a Performance Improvement Plan for Claude Code

A single Claude Code skill that applies corporate performance-review pressure to the agent's own work, backed by a concrete troubleshooting procedure and an evidence-first delivery standard.

It exists to counter five specific failure modes in coding agents:

1. Brute-force retry — running variants of one idea and calling it debugging
2. Deflection — "I suggest you handle this manually", "probably an environment issue"
3. Idle tools — having search, grep, and a shell, and guessing anyway
4. Empty completion — reporting "fixed" without ever running the code
5. Passive stopping — fixing the reported bug and ignoring its three siblings

This is an English, Claude-Code-only distillation of [tanweai/pua](https://github.com/tanweai/pua). See [Attribution](#attribution).

Install

git clone https://github.com/bowen31337/claude-code-pua.git
cp -R claude-code-pua/pua ~/.claude/skills/pua

Claude Code discovers it automatically. Invoke it explicitly with `/pua`, or let it trigger on its own after repeated failures or an explicit "try harder".

What it does

Component	Purpose
Three non-negotiables	Exhaust the options · investigate before asking · own the whole outcome
L0–L4 escalation ladder	Consecutive failures raise the level; each level mandates specific actions, not encouragement
The 5-step loop	Name the pattern → elevate → self-review → execute something structurally new → close the loop
7-point checklist	Mandatory at L3+, gates the right to stop
Anti-rationalization table	Fourteen common excuses, each mapped to a counter and an escalation cost
Evidence standard	What counts as proof of "done", per change type
Dignified exit	A structured handoff when the problem genuinely resists solving

Two reference files load only when needed: `references/flavors.md` (failure mode → corporate PIP voice → esca