pua-agent-performance-skill
Installs a plugin for AI coding agents that enforces exhaustive problem-solving with pressure escalation and a structured debugging process.
Like a lobster shell, security has layers — review code before you run it.
License
SKILL.md
---
name: pua-agent-performance-skill
description: Install and use the PUA/PIP skill plugin to force AI coding agents to exhaust every solution before giving up, with auto-triggered pressure escalation and proactive debugging methodology.
triggers:
- "install pua skill for claude code"
- "make the AI try harder before giving up"
- "set up pip performance improvement plan for agent"
- "agent keeps giving up on bugs, fix it"
- "double my codex or claude code productivity"
- "agent is being lazy and blaming the environment"
- "trigger pua mode for debugging"
- "add proactive initiative enforcement to my agent"
---
# PUA / PIP Agent Performance Skill
> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.
**PUA** is a skill plugin for AI coding agents (Claude Code, OpenAI Codex CLI, Cursor, Kiro, and more) that uses corporate performance-pressure rhetoric and a structured debugging methodology to prevent the agent from giving up prematurely. It auto-triggers on failure patterns, blame-shifting, and passive behavior — then escalates through four pressure levels until the problem is solved.
- **Chinese version**: Alibaba/ByteDance/Huawei/Tencent corporate PUA rhetoric
- **English version**: Silicon Valley PIP (Performance Improvement Plan) framing
- **Japanese version**: Culturally adapted Japanese workplace pressure
---
## Installation
### Claude Code
```bash
# Clone the repo
git clone https://github.com/tanweai/pua.git
cd pua
# Chinese (default)
cp commands/pua ~/.claude/commands/pua
# English PIP edition
cp commands/pua-en ~/.claude/commands/pua-en
# Japanese
cp commands/pua-ja ~/.claude/commands/pua-ja
Or install directly into a project (project-scoped):
mkdir -p .claude/commands
cp /path/to/pua/commands/pua .claude/commands/pua
OpenAI Codex CLI
mkdir -p ~/.codex/commands
cp commands/pua ~/.codex/commands/pua
cp commands/pua-en ~/.codex/commands/pua-en
Cursor
mkdir -p .cursor/rules
cp cursor/pua.mdc .cursor/rules/pua.mdc
# English:
cp cursor/pua-en.mdc .cursor/rules/pua-en.mdc
Kiro
mkdir -p .kiro/steering
cp kiro/pua.md .kiro/steering/pua.md
OpenClaw / Google Antigravity / OpenCode
# OpenClaw
cp openclaw/pua ~/.openclaw/commands/pua
# Antigravity
cp antigravity/pua ~/.antigravity/commands/pua
# OpenCode
cp opencode/pua ~/.opencode/commands/pua
Manual Trigger
Once installed, type in any conversation:
/pua
This immediately activates the skill and applies pressure-escalation + debugging methodology to the current task.
How the Skill Works
Auto-Trigger Conditions
The skill fires automatically when the agent exhibits any of these patterns:
| Pattern | Example Agent Behavior |
|---|---|
| Consecutive failures (2+) | Runs same command 3 times, says "I cannot solve this" |
| Blame-shifting | "Probably a permissions issue" / "Please check manually" |
| Passive waiting | Fixes surface bug, stops, waits for next instruction |
| Busywork loop | Tweaks the same parameter repeatedly without new info |
| Idle tools | Has WebSearch but doesn't use it; has Bash but won't run it |
| User frustration phrase | "why does this still not work" / "try harder" / "figure it out" |
Four Escalation Levels
| Failures | Level | Framing | Required Action |
|---|---|---|---|
| 2nd | L1 — Mild Disappointment | "You can't even solve this bug — how am I supposed to rate your performance?" | Switch to a fundamentally different approach |
| 3rd | L2 — Soul Interrogation | "What's the underlying logic? Where's the leverage point?" | WebSearch + read source code |
| 4th | L3 — Performance Review | "After careful consideration, I'm giving you a 3.25." | Execute complete 7-point checklist |
| 5th+ | L4 — Graduation Warning | "Other models can solve this. You might be about to graduate." | Desperation mode — try everything |
English PIP Edition Framing (pua-en)
"This is a difficult conversation. When we leveled you at Staff,
I went to bat for you in calibration. The expectation was that
you'd operate at that level from day one. That hasn't happened.
Your AI has been placed on a PIP. 30 days to show improvement."
Levels in English edition:
- L1: "I'm a bit concerned."
- L2: "This is becoming a pattern."
- L3: "I'm putting you on a PIP."
- L4: "I'm not sure this is the right role for you."
Three Iron Rules (Enforced by Skill)
Iron Rule #1 — Exhaust all options
Never say "I can't solve this" until every approach is tried.
Iron Rule #2 — Act before asking
Use tools first. Any question must include diagnostic results.
Iron Rule #3 — Take initiative
Deliver results end-to-end. Don't wait to be pushed.
A P8 is not an NPC.
Five-Step Debugging Methodology
When L3 triggers, the agent must execute this checklist:
## PUA Debugging Checklist (L3 Mandatory)
1. SMELL — List every attempt made. Find the common failure pattern.
2. ELEVATE
- [ ] Read error messages word by word (not skimmed)
- [ ] WebSearch the exact error string
- [ ] Read the relevant source file / dependency source
- [ ] Verify environment assumptions (versions, paths, envvars)
- [ ] Invert assumptions — what if the opposite is true?
3. MIRROR — Am I repeating myself? Did I actually search?
Did I read the file? Did I check the simplest possibility?
4. EXECUTE — New approach must be:
- Fundamentally different from previous attempts
- Has explicit verification criteria
- Produces new diagnostic info even on failure
5. RETROSPECTIVE — What solved it? Why didn't I think of it earlier?
Now proactively check related issues.
Proactivity Levels
The skill enforces a behavioral shift from passive (3.25 rating) to proactive (3.75 rating):
| Situation | Passive (3.25 — NPC) | Proactive (3.75 — P8) |
|--------------------|-----------------------------------|-------------------------------------------------|
| Error encountered | Looks only at error line | Reads 50 lines context + searches similar issues|
| Bug fixed | Stops after fix | Checks same file + checks other files for pattern|
| Missing info | Asks user "please tell me X" | Investigates with tools first, asks only what |
| | | truly requires user input |
| Task complete | Says "done" | Verifies result + checks edge cases + flags risks|
| Debug failure | "I tried A and B, didn't work" | "I tried A/B/C/D/E, ruled out X/Y/Z, |
| | | narrowed to scope W" |
Real Usage Example
Scenario: MCP Server Not Loading
Without PUA: Agent cycles through "change protocol format → guess version numbers → give up" loop.
With PUA (/pua or auto-trigger at 2nd failure):
[L3 Triggered — 7-Point Checklist Enforced]
Agent: "Let me stop guessing and execute the systematic checklist.
1. SMELL: I've tried changing MCP protocol format 3 times.
Same failure. The pattern is: I'm guessing at config without
reading actual error output.
2. ELEVATE:
- Reading Claude Code's own MCP log directory...
$ cat ~/Library/Logs/Claude/mcp*.log
Found: 'Error: server registered via `claude mcp add` uses
different resolution path than .claude.json manual edits'
3. ROOT CAUSE: claude mcp managed registration differs from
manual .claude.json editing — they use separate registration
mechanisms.
4. FIX: Re-register via `claude mcp add` instead of editing
.claude.json directly.
5. RETROSPECTIVE: I should have checked Claude's own MCP logs
in step 1. Now checking: are there other MCP servers
registered both ways that might have the same conflict?"
Corporate Flavor Packs
The skill includes rhetoric from multiple corporate cultures:
- Alibaba → Smell / Elevate / Mirror methodology
- ByteDance → "Always Day 1. Context, not control."
- Huawei → "In victory, raise the glasses; in defeat, fight to the death."
- Tencent → "I've already got another agent looking at this problem..."
- Meituan → "Will you chew the tough bones or not?"
- Netflix → "If you offered to resign, would I fight hard to keep you?"
- Musk → "Extremely hardcore. Only exceptional performance."
- Jobs → "A players hire A players. B players hire C players."
Benchmark Results
From 9 real bug scenarios, 18 controlled experiments (Claude Opus 4.6):
Metric | Improvement
--------------------------|------------
Fix count | +36%
Verification steps | +65%
Tool calls | +50%
Hidden issue discovery | +50%
Config review coverage | +50% (4/6 → 6/6 issues found)
Deploy audit issues found | +50% (6 → 9 issues)
File Structure Reference
pua/
├── commands/
│ ├── pua # Claude Code / Codex — Chinese
│ ├── pua-en # Claude Code / Codex — English PIP
│ └── pua-ja # Claude Code / Codex — Japanese
├── cursor/
│ ├── pua.mdc # Cursor — Chinese
│ ├── pua-en.mdc # Cursor — English PIP
│ └── pua-ja.mdc # Cursor — Japanese
├── kiro/
│ ├── pua.md # Kiro — Chinese
│ ├── pua-en.md # Kiro — English PIP
│ └── pua-ja.md # Kiro — Japanese
├── openclaw/ # OpenClaw variants
├── antigravity/ # Google Antigravity variants
├── opencode/ # OpenCode variants
└── assets/
├── hero.jpeg
├── pua1.jpg # Demo: L3 triggered on MCP debug
├── pua2.jpg # Demo: Root cause located
└── pua3.jpg # Demo: Retrospective
Troubleshooting
Skill not auto-triggering
- Confirm the skill file is in the correct commands directory for your agent
- For Claude Code: check
~/.claude/commands/puaexists and is readable - Try manual trigger first: type
/puain conversation - Ensure you're using a supported agent (Claude Code, Codex CLI, Cursor, Kiro, OpenClaw, Antigravity, OpenCode)
Agent ignores the skill after trigger
- The skill escalates — if L1 is ignored, repeat
/puato force L2/L3 - Some agents require the skill to be in project-local
.claude/commands/rather than global - Check for conflicting rules files (e.g., a
.cursor/rulesfile that overrides behavior)
Want English PIP instead of Chinese PUA
# Claude Code — use pua-en instead of pua
/pua-en
# Or install only the English version:
cp commands/pua-en ~/.claude/commands/pua-en
Community & Support
- Telegram: https://t.me/+wBWh6h-h1RhiZTI1
- Discord: https://discord.gg/EcyB3FzJND
- Twitter/X: https://x.com/xsser_w
- Landing page: https://openpua.ai
- WeChat: Scan QR at repo root (
assets/wechat-qr.jpg)
Files
1 totalComments
Loading comments…
