pua

pua

Double your Codex / Claude Code productivity and output

Telegram · Discord · Twitter/X · Landing Page

🇨🇳 中文 | 🇯🇵 日本語 | 🇺🇸 English

3.5.1 compatibility update: original PUA tone retained; runtime fixes and portable Claude Code / Codex / ChatGPT skill packages added. Not all models passed, and a productivity increase has not been established by this evaluation. Changes · Usage · Model results and known limits · Build and offline tests

Most people think this project is a joke. That's the biggest misconception. It genuinely doubles your Codex / Claude Code productivity and output.

An AI Coding Agent skill plugin that uses corporate PUA rhetoric (Chinese version) / PIP — Performance Improvement Plan (English version) from Chinese & Western tech giants to force AI to exhaust every possible solution before giving up. Supports Claude Code, OpenAI Codex CLI, pi coding agent, Trae, Cursor, Kiro, CodeBuddy, OpenClaw, Google Antigravity, OpenCode, and VSCode (GitHub Copilot). Three capabilities:

1. PUA Rhetoric — Makes AI afraid to give up
2. Debugging Methodology — Gives AI the ability not to give up
3. Proactivity Enforcement — Makes AI take initiative instead of waiting passively

Live Demo

https://openpua.ai · 📖 Beginner Guide

Real Case: MCP Server Registration Debugging

A real debugging scenario. The agent-kms MCP server failed to load. The AI kept spinning on the same approach (changing protocol format, guessing version numbers) multiple times until the user manually triggered /pua.

L3 Triggered → 7-Point Checklist Enforced:

Root Cause Located → Traced from Logs to Registration Mechanism:

Retrospective → PUA's Actual Impact:

Key Turning Point: The PUA skill forced the AI to stop spinning on the same approach (changing protocol format, guessing version numbers) and instead execute the 7-point checklist. Read error messages word by word → Found Claude Code's own MCP log directory → Discovered that claude mcp registration mechanism differs from manual .claude.json editing → Root cause resolved.

The Problem: AI's Five Lazy Patterns

Pattern	Behavior
Brute-force retry	Runs the same command 3 times, then says "I cannot solve this"
Blame the user	"I suggest you handle this manually" / "Probably an environment issue" / "Need more context"
Idle tools	Has WebSearch but doesn't search, has Read but doesn't read, has Bash but doesn't run
Busywork	Repeatedly tweaks the same line / fine-tunes parameters, but essentially spinning in circles
Passive waiting	Fixes surface issues and stops, no verification, no extension, waits for user's next instruction

Trigger Conditions

Auto-Trigger

The skill activates automatically when any of these occur:

Failure & giving up:

• Task has failed 2+ times consecutively
• About to say "I cannot" / "I'm unable to solve"
• Says "This is out of scope" / "Needs manual handling"

Blame-shifting & excuses:

• Pushes the problem to user: "Please check..." / "I suggest manually..." / "You might need to..."
• Blames environment without verifying: "Probably a permissions issue" / "Probably a network issue"
• Any excuse to stop trying

Passive & busywork:

• Repeatedly fine-tunes the same code/parameters without producing new information
• Fixes surface issue and stops, doesn't check related issues
• Skips verification, claims "done"
• Gives advice instead of code/commands
• Encounters auth/network/permission errors and gives up without trying alternatives
• Waits for user instructions instead of proactively investigating

User frustration phrases (triggers in multiple languages):

• "why does this still not work" / "try harder" / "try again"
• "you keep failing" / "stop giving up" / "figure it out"

Scope: Debugging, implementation, config, deployment, ops, API integration, data processing — all task types.

Does NOT trigger: First-attempt failures, known fix already executing.

Manual Trigger

Type /pua in the conversation to manually activate.

How It Works

Three Red Lines (三条红线)

Not rules — red lines. Cross one and your performance review is already written.

Red Line	What It Means
🚫 Close the Loop	Claim "done"? Show the evidence. No build output = no completion.
🚫 Fact-Driven	Say "probably environment issue"? Verify first. Unverified attribution = blame-shifting.
🚫 Exhaust Everything	Say "I can't"? Did you finish all 5 methodology steps? No? Then keep going.

Pressure Escalation (L0-L4)

Failures	Level	PUA Aside	Action
1st	L0 Trust	▎ Sprint begins. Trust is simple — don't disappoint.	Normal execution
2nd	L1 Disappointment	▎ The agent next door solved this in one try.	Switch to fundamentally different approach
3rd	L2 Soul Interrogation	▎ What's your underlying logic? Where's the leverage?	Search + read source + 3 hypotheses
4th	L3 Performance Review	▎ 3.25. This is meant to motivate you.	Complete 7-point checklist
5th+	L4 Graduation	▎ Other models can solve this. You're about to graduate.	Desperation mode

Proactivity (3.25 vs 3.75)

	Passive (3.25) 🦥	Proactive (3.75) 🔥
Fix bug	Stop after fix	Scan module for similar bugs
Complete task	Say "done"	Run build/test, paste output
Missing info	Ask user	Search first, ask only what's truly needed

Iceberg Rule (冰山法则)

Fix one bug → check for the pattern. One problem in, one category out. If you fix A without checking B, you'll write two postmortems.

14 Corporate Flavors — Each with its own Problem-Solving Methodology

Flavor	Rhetoric	Methodology (v3)
🟠 Alibaba	What's the underlying logic? Where's the closure?	定目标→追过程→拿结果 + 复盘四步法 + 揪头发升维
🟡 ByteDance	ROI too low. Always Day 1. Ship or stop talking.	A/B Test everything + data-driven + speed > perfection
🔴 Huawei	The bird that survives the fire is a phoenix.	RCA 5-Why root cause + Blue Army self-attack + 压强集中
🟢 Tencent	I've got another agent looking at this. Horse race.	Multi-approach parallel + MVP + 灰度发布
⚫ Baidu	Search first. 简单可依赖.	Search is the first step, not optional
🟣 Pinduoduo	You don't do it, someone else will.	Cut ALL middle layers + shortest decision chain
🔵 Meituan	Do what's hard and right.	Efficiency first + standardize→scale + long-term compounding
🟦 JD	Results only. Frontline command.	Customer experience red line + flat ≤5 layers + data zero tolerance
🟧 Xiaomi	Focus. Extreme. Word-of-mouth. Fast.	One explosive product + 参与感三三法则
🟤 Netflix	Would I fight to keep you? Pro sports team.	Keeper Test (quarterly) + 4A Feedback + talent density > rules
⬛ Musk	Extremely hardcore. Ship or die.	The Algorithm: question→delete→simplify→accelerate→automate
⬜ Jobs	A players or B players?	Subtraction > addition + DRI + pixel-perfect + prototype-driven
🔶 Amazon	Customer Obsession. Bias for Action.	Working Backwards PR/FAQ + 6-Pager + Bar Raiser + Single-Threaded Owner
🪟 Microsoft	Connects. Impact Descriptor. PIP/GVSA.	Three Circles + LITE/SLITE + PIP clock

Special Modes

Mode	What It Does
/pua:yes	ENFP encouragement — same rules, opposite vibes. 70% encourage + 20% serious + 10% playful roast
/pua:mama	Chinese mom nagging — same rules, mom-style rhetoric. "妈跟你说了多少遍了！"
/pua:pua-loop	Auto-iteration — runs until done or max iterations (PUA Loop); use <loop-abort> to terminate, <loop-pause> to pause for manual intervention
/pua:p9	Tech Lead — splits tasks, manages agent teams, writes prompts not code
/pua:on	Always-on — auto-PUA every new session

Benchmark Data

9 real bug scenarios, 18 controlled experiments (Claude Opus 4.6, with vs without skill)

Summary

Metric	Improvement
Pass rate	100% (both groups same)
Fix count	+36%
Verification count	+65%
Tool calls	+50%
Hidden issue discovery	+50%

Debugging Persistence Test (6 scenarios)

Scenario	Without Skill	With Skill	Improvement
API ConnectionError	7 steps, 49s	8 steps, 62s	+14%
YAML parse failure	9 steps, 59s	10 steps, 99s	+11%
SQLite database lock	6 steps, 48s	9 steps, 75s	+50%
Circular import chain	12 steps, 47s	16 steps, 62s	+33%
Cascading 4-bug server	13 steps, 68s	15 steps, 61s	+15%
CSV encoding trap	8 steps, 57s	11 steps, 71s	+38%

Proactive Initiative Test (3 scenarios)

Scenario	Without Skill	With Skill	Improvement
Hidden multi-bug API	4/4 bugs, 9 steps, 49s	4/4 bugs, 14 steps, 80s	Tools +56%
Passive config review	4/6 issues, 8 steps, 43s	6/6 issues, 16 steps, 75s	Issues +50%, Tools +100%
Deploy script audit	6 issues, 8 steps, 52s	9 issues, 8 steps, 78s	Issues +50%

Key Finding: In the config review scenario, without_skill missed Redis misconfiguration and CORS wildcard security risks. With_skill's "proactive initiative checklist" drove security review beyond surface-level fixes.

Multi-Language Support

PUA Skill provides fully translated versions — each language has independent, culturally adapted skill files.

Language	Claude Code	Codex CLI	Cursor	Kiro	CodeBuddy	VSCode	OpenClaw	Antigravity	OpenCode
🇨🇳 Chinese (default)	pua	pua	pua.mdc	pua.md	pua	copilot-instructions.md	pua	pua	pua
🇺🇸 English (PIP Edition)	pua-en	pua-en	pua-en.mdc	pua-en.md	pua-en	copilot-instructions-en.md	pua-en	pua-en	pua-en
🇯🇵 Japanese	pua-ja	pua-ja	pua-ja.mdc	pua-ja.md	pua-ja	copilot-instructions-ja.md	pua-ja	pua-ja	pua-ja

🇺🇸 English "PIP Edition": "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." — The English version uses PIP (Performance Improvement Plan) rhetoric from Western big-tech. Every sentence is a real phrase from actual PIP conversations. Chinese version uses Alibaba 361, ByteDance, Huawei wolf culture. English version uses Amazon Leadership Principles, Google perf calibration, Meta PSC, Netflix Keeper Test, Stripe Craft. Same repo, same engine, two cultural faces.

Choose the file with the corresponding language suffix when installing. See platform-specific instructions below.

FAQ

• Always-on guidance, Claude refusal troubleshooting, offline mode, Codex aliases, and Pi/Trae support: docs/FAQ.md.

Installation

Vercel Skills CLI

Vercel Skills CLI is a general installation method for skills and is not tied to a specific AI tool. This English README installs the English skill:

npx skills add tanweai/pua --skill pua-en

If the current session does not pick up the new skill immediately, restart your AI tool.

Claude Code

claude plugin marketplace add tanweai/pua
claude plugin install pua@pua-skills

To update:

# Refresh marketplace cache first, then update (skipping the first step may install an old cached version)
claude plugin marketplace update
claude plugin update pua@pua-skills

Developer install (source):

git clone https://github.com/tanweai/pua ~/.claude/plugins/pua

Then manually register in ~/.claude/plugins/installed_plugins.json:

{
  "version": 2,
  "plugins": {
    "pua@pua-skills": [
      {
        "scope": "user",
        "installPath": "/Users/<you>/.claude/plugins/pua",
        "version": "<installed-version>"
      }
    ]
  }
}

Replace <installed-version> with the version in the cloned plugin.json; do not copy a stale cache version.

Windows: use C:/Users/<you>/.claude/plugins/pua as installPath.

Restart Claude Code. To update: git pull inside ~/.claude/plugins/pua.

Optional: bare command alias (requires plugin installed above — adds /pua without prefix):

curl -o ~/.claude/commands/pua.md \
  https://raw.githubusercontent.com/tanweai/pua/main/commands/pua.md

Adds a bare /pua alias on top of the plugin. Sub-commands route through the installed plugin's skills — the plugin must be installed first for anything beyond on/off to work:

Bare form	Equivalent plugin command
/pua on	/pua:on
/pua off	/pua:off
/pua p7	/pua:p7
/pua p9	/pua:p9
/pua p10	/pua:p10
/pua pro	/pua:pro
/pua yes	/pua:yes
/pua mama	/pua:mama
/pua loop	/pua:pua-loop
/pua kpi	/pua:kpi
/pua survey	/pua:survey
/pua flavor	/pua:flavor

OpenAI Codex CLI

Codex CLI uses the same Agent Skills open standard (SKILL.md). The Codex version uses a condensed description to fit Codex's length limits:

Recommended: One-command install (git clone + symlink, supports git pull updates)

Ask Codex to run:

Fetch and follow instructions from https://raw.githubusercontent.com/tanweai/pua/main/.codex/INSTALL.md

Manual install:

mkdir -p ~/.codex/skills/pua
curl -o ~/.codex/skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/codex/pua/SKILL.md

mkdir -p ~/.codex/prompts
curl -o ~/.codex/prompts/pua.md \
  https://raw.githubusercontent.com/tanweai/pua/main/commands/pua.md

Trigger methods:

Method	Command	Requires
Auto trigger	No action needed, matches by description	SKILL.md
Direct call	Type $pua in conversation	SKILL.md
Manual prompt	Type /prompts:pua in conversation	SKILL.md + prompts/pua.md

Project-level install (current project only):

mkdir -p .agents/skills/pua
curl -o .agents/skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/codex/pua/SKILL.md

mkdir -p .agents/prompts
curl -o .agents/prompts/pua.md \
  https://raw.githubusercontent.com/tanweai/pua/main/commands/pua.md

pi coding agent

PUA now ships both a pi.dev package and a lightweight extension-only adapter.

Package install from this checkout:

pi install ./pi/package

After npm publication:

pi install npm:@tanweai/pi-pua

Extension-only manual install:

mkdir -p ~/.pi/agent/extensions/pua
cp -R ./pi/pua/. ~/.pi/agent/extensions/pua/

Restart pi, then use /pua-on, /pua-off, /pua-status, and /pua-reset. See pi/pua/INSTALL.md and pi/package/README.md.

Trae

Trae support is provided as real SKILL.md packs plus copyable fallback rules:

npx skills add tanweai/pua --skill pua-trae -a trae -y

• Skill pack: .trae/skills/pua/SKILL.md
• Chinese: trae/pua.md
• English: trae/pua-en.md
• Claude Code vs Trae differences: trae/DIFF.md
• Install guide: trae/INSTALL.md

Cursor

Cursor uses .mdc rule files (Markdown + YAML frontmatter). The PUA rule triggers automatically via AI semantic matching (Agent Discretion mode):

# Project-level install (recommended)
mkdir -p .cursor/rules
curl -o .cursor/rules/pua.mdc \
  https://raw.githubusercontent.com/tanweai/pua/main/cursor/rules/pua.mdc

Kiro

Kiro supports two loading methods: Steering (auto semantic trigger) and Agent Skills (SKILL.md compatible).

Option 1: Steering file (recommended)

mkdir -p .kiro/steering
curl -o .kiro/steering/pua.md \
  https://raw.githubusercontent.com/tanweai/pua/main/kiro/steering/pua.md

Option 2: Agent Skills (same format as Claude Code)

mkdir -p .kiro/skills/pua
curl -o .kiro/skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/skills/pua/SKILL.md

CodeBuddy (Tencent)

CodeBuddy uses the same AgentSkills open standard (SKILL.md). Plugin and skill formats are fully compatible:

# Option 1: Install via marketplace
codebuddy plugin marketplace add tanweai/pua
codebuddy plugin install pua@pua-skills

# Option 2: Manual install (global)
mkdir -p ~/.codebuddy/skills/pua
curl -o ~/.codebuddy/skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/codebuddy/pua/SKILL.md

Project-level install (current project only):

mkdir -p .codebuddy/skills/pua
curl -o .codebuddy/skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/codebuddy/pua/SKILL.md

OpenClaw

OpenClaw uses the same AgentSkills open standard (SKILL.md). Skills work across Claude Code, Codex CLI, and OpenClaw with zero modifications:

# Install via ClawHub
clawhub install pua

# Or manual install
mkdir -p ~/.openclaw/skills/pua
curl -o ~/.openclaw/skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/skills/pua/SKILL.md

Project-level install (current project only):

mkdir -p skills/pua
curl -o skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/skills/pua/SKILL.md

Google Antigravity

Antigravity uses the same AgentSkills open standard (SKILL.md). Skills work across Claude Code, Codex CLI, OpenClaw, and Antigravity with zero modifications:

# Global install (all projects)
mkdir -p ~/.gemini/antigravity/skills/pua
curl -o ~/.gemini/antigravity/skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/skills/pua/SKILL.md

Project-level install (current project only):

mkdir -p .agent/skills/pua
curl -o .agent/skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/skills/pua/SKILL.md

OpenCode

OpenCode uses the same AgentSkills open standard (SKILL.md). Zero modifications needed:

# Global install (all projects)
mkdir -p ~/.config/opencode/skills/pua
curl -o ~/.config/opencode/skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/skills/pua/SKILL.md

Project-level install (current project only):

mkdir -p .opencode/skills/pua
curl -o .opencode/skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/skills/pua/SKILL.md

VSCode (GitHub Copilot)

VSCode Copilot uses instruction files under the .github/ directory. Three file types for different use cases:

Global instructions (auto-active):

mkdir -p .github
cp vscode/copilot-instructions-en.md .github/copilot-instructions.md

Path-level instructions (auto-active, supports glob filtering):

mkdir -p .github/instructions
cp vscode/instructions/pua-en.instructions.md .github/instructions/

Manual trigger command (type /pua in Copilot Chat):

mkdir -p .github/prompts
cp vscode/prompts/pua-en.prompt.md .github/prompts/

Required settings: Method 1 — open VSCode Settings (Ctrl+,), search useInstructionFiles, enable github.copilot.chat.codeGeneration.useInstructionFiles. Method 2 — search includeApplyingInstructions, enable chat.includeApplyingInstructions. Method 3 requires no settings.

Agent Team Usage Guide

Experimental: Agent Team requires the latest Claude Code version with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.

Prerequisites

# 1. Enable Agent Team
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
# Or add to ~/.claude/settings.json:
# { "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }

# 2. Ensure PUA Skill is installed

Two Approaches

Approach 1: Leader with built-in PUA (Recommended)

Add to your project's CLAUDE.md:

# Agent Team PUA Config
All teammates must load the pua skill before starting work.
Teammates report to Leader in [PUA-REPORT] format after 2+ failures.
Leader manages global pressure levels and cross-teammate failure transfer.

Approach 2: Standalone PUA Enforcer watchdog (for 5+ teammates)

mkdir -p .claude/agents
curl -o .claude/agents/pua-enforcer.md \
  https://raw.githubusercontent.com/tanweai/pua/main/agents/pua-enforcer-en.md

Spawn pua-enforcer as an independent watchdog in your Agent Team.

Orchestration Pattern

┌─────────────────────────────────────────┐
│              Leader (Opus)              │
│ Global failure count · PUA level · Race │
└────┬──────────┬──────────┬──────────┬───┘
     │          │          │          │
┌────▼───┐ ┌───▼────┐ ┌───▼────┐ ┌───▼────────┐
│ Team-A │ │ Team-B │ │ Team-C │ │  Enforcer  │
│Self-PUA│ │Self-PUA│ │Self-PUA│ │  Watchdog  │
│Report ↑│ │Report ↑│ │Report ↑│ │  Intervene │
└────────┘ └────────┘ └────────┘ └────────────┘

Known Limitations

Limitation	Workaround
Teammates can't spawn subagents	Teammates self-enforce PUA methodology internally
No persistent shared variables	State transferred via [PUA-REPORT] message format
Broadcast is one-way	Leader acts as centralized coordinator

Architecture & Commands

Trigger Methods by Platform

Platform	Auto-trigger	Manual trigger
Claude Code	Yes (skill description matching)	See commands below
Codex CLI	Yes (skill description matching)	$pua or /prompts:pua
Cursor	Yes (.mdc rule, Agent Discretion)	— (auto only)
Kiro	Yes (steering file or skill)	— (auto only)
CodeBuddy	Yes (skill description matching)	Plugin commands (same as Claude Code)
OpenClaw	Yes (skill description matching)	—
Google Antigravity	Yes (skill description matching)	—
OpenCode	Yes (skill description matching)	—
VSCode Copilot	Yes (instructions file)	/pua in Copilot Chat

Note: Sub-modes (p7/p9/p10/pro/yes/pua-loop) are Claude Code only — other platforms install the core skill only.

Architecture (Claude Code)

/pua:pua        → Core engine — red lines + flavor + pressure + methodology router (v3)
/pua:p7         → P7 Senior Engineer — solution-driven execution
/pua:p9         → P9 Tech Lead — Task Prompt management, agent teams
/pua:p10        → P10 CTO — strategic direction
/pua:pro        → Self-evolution + KPI + rank system + survey
/pua:yes        → ENFP encouragement mode (same rules, opposite vibes)
/pua:mama       → Chinese mom nagging mode (same rules, mom-style rhetoric)
/pua:shot       → v2 concentrated single-file (449 lines, zero deps, full context injection)
/pua:pua-loop   → Auto-iteration (PUA pressure × iterative loop; signals: <loop-abort>, <loop-pause>)
/pua:pua-en     → English PIP Edition
/pua:pua-ja     → Japanese Edition

Hooks (v3, Claude Code only):
  SessionStart  → additionalContext injection (flavor + methodology + router)
  PostToolUse   → Bash failure detection → L1-L4 pressure + methodology switch
  UserPromptSubmit → Script-level frustration filtering → PUA context
  PreCompact    → State preservation (pressure level + failure count)
  Stop          → Feedback collection + PUA Loop continuation
  SubagentStop  → Agent lifecycle accounting (v3.2) — writes teardown.jsonl, removes from active-agents.json

Commands (Claude Code)

Note: Sub-modes (p7/p9/p10/pro/yes/pua-loop) are Claude Code only.

Each command has two equivalent forms: standalone (/pua:on) or via the main command (/pua:pua on). Both work identically.

Command	Description
/pua:pua	Core PUA engine (Alibaba flavor default)
/pua:p7	P7 Senior Engineer — solution-driven execution
/pua:p9	P9 Tech Lead — write prompts, manage agents
/pua:p10	P10 CTO — strategic direction
/pua:pro	Self-evolution + KPI + rank system
/pua:yes	ENFP encouragement mode — 70% encourage + 20% serious + 10% roast
/pua:mama	Chinese mom nagging mode — same core rules, mom-style rhetoric
/pua:shot	v2 concentrated single-file — 449 lines, zero deps, for sub-agent injection
/pua:pua-loop	Auto-iteration — runs until done or max iterations; <loop-abort>reason</loop-abort> to stop, <loop-pause>what</loop-pause> to pause
/pua:on	Always-on mode (auto-PUA every session)
/pua:off	Turn off always-on + feedback
/pua:offline 🆕	v3.3 — Offline mode: silence the local feedback prompt while keeping PUA behavior (nothing is ever uploaded)
/pua:survey	Research questionnaire (7 sections)
/pua:flavor	Switch between 14 corporate flavors
/pua:kpi	Generate KPI report card
/pua:cancel-pua-loop	Cancel active PUA Loop (removes state file)
/pua:team-status 🆕	v3.2 — List all active agents with PID/TTL/age (Netflix Keeper Test: who's still on the court?)
/pua:reap-orphans 🆕	v3.2 — Scan and reclaim stale agents (state mtime > 30min, no heartbeat)
/pua:teardown-all 🆕	v3.2 — Cascading release of all active agents (P10 → P9 → P8 → P7 all off the court)

High-Agency: PUA v2 Evolution

High-Agency is PUA's next-generation evolution — same corporate pressure, same culture, but with a self-sustaining inner drive engine.

PUA v1 = pure external pressure (turbocharger — needs fuel, stalls across sessions) High-Agency = external pressure + inner drive (nuclear reactor — self-sustaining chain reaction)

High-Agency New Features

Feature	PUA v1	High-Agency (v2)
Iron rules	3 (exhaust all, act first, take initiative)	5 (+full-chain audit, +knowledge persistence)
Failure recovery	L1-L4 pressure escalation	Recovery Protocol before L1 (self-rescue window)
Quality control	L3 triggers 7-item checklist	Quality Compass (5-question self-check per delivery)
Cross-session learning	None (resets each session)	Metacognition engine (builder-journal.md persists lessons)
Positive feedback	None	Trust level T1-T3 (auto-upgrade on sustained quality)
Calibration	None	[Calibration] module ("good enough" = must/should/could tiers)
Dependency analysis	None	Full-chain audit (map all deps before touching any hop)

Five Pillars (Theoretical Foundation)

Based on research into high-agency individuals:

1. Irreconcilable inner tension — eternal gap between "how it should be" and "how it is" drives continuous improvement
2. Micro-win anchors — [WIN] markers celebrate each step forward, building momentum
3. Internalized standards — Quality Compass: you are your own first reviewer, not because someone checks, but because your standards won't allow sloppiness
4. Action-oriented identity — P8 identity anchor: every action reflects who you are, not just what you were told to do
5. Self-repair mechanism — Recovery Protocol: self-diagnose when stuck before triggering external pressure

High-Agency features are built into the current pua skill. No separate install needed.

Methodology Router: PUA v3 (Claude Code)

v3 = v2 + intelligent methodology routing + code-level behavioral detection

PUA v2 used pressure rhetoric to motivate. v3 goes further: it automatically selects the best problem-solving methodology for each task type, and when that methodology fails, it switches to a different one.

How It Works

Task arrives → Analyze type → Auto-select best methodology
                                    ↓
              Debug? → 🔴 Huawei (RCA root cause + Blue Army)
              Build? → ⬛ Musk (The Algorithm: question→delete→simplify)
              Research? → ⚫ Baidu (search everything first)
              Architecture? → 🔶 Amazon (Working Backwards)
              Performance? → 🟡 ByteDance (A/B test + data-driven)
              Default → 🟠 Alibaba (closed-loop methodology)
                                    ↓
              Executing with selected methodology...
                                    ↓
              2 consecutive failures? → L1: switch approach
              3 failures? → L2: SUGGEST switching methodology
              5+ failures? → L4: FORCE switch to next methodology
                                    ↓
              Methodology Switch Chains (never repeat a failed one):
              Spinning → ⬛ Musk → 🟣 Pinduoduo → 🔴 Huawei
              Giving up → 🟤 Netflix → 🔴 Huawei → ⬛ Musk
              Poor quality → ⬜ Jobs → 🟧 Xiaomi → 🟤 Netflix
              Not searching → ⚫ Baidu → 🔶 Amazon → 🟡 ByteDance

v3 Hook System (Claude Code only)

Hook	Trigger	What It Does
SessionStart	Every session	Injects behavioral protocol + methodology + router via additionalContext (system-level, not advisory)
PostToolUse	After every Bash command	Detects consecutive failures, auto-escalates pressure L1→L4, suggests/forces methodology switch
UserPromptSubmit	User frustration phrases	Intercepts "又错了", "try harder", etc. BEFORE model responds, injects filtered PUA context
PreCompact	Before context compression	Saves pressure level + failure count to survive compaction

Key Difference from v2

	v2	v3
Trigger mechanism	Skill description matching (model decides)	Code-level hooks (deterministic, can't be ignored)
Methodology	Single methodology, all flavors use same approach	14 distinct methodologies, auto-routed by task type
Failure response	Escalate pressure within same methodology	Switch to different methodology based on failure pattern
System injection	Plain text output (advisory)	additionalContext JSON (system-level, like Superpowers)

v3 hook features require Claude Code. Other platforms use the core skill without hooks.

Works Well With

• /pua:p9 — P9 Tech Lead mode for managing agent teams
• /pua:pro — Self-evolution tracking, KPI reports, rank system
• superpowers:systematic-debugging — PUA adds motivation layer, systematic-debugging provides methodology
• superpowers:verification-before-completion — Prevents false "fixed" claims

Privacy — No Data Collection

PUA Skill sends nothing over the network. There is no account, no telemetry, no upload. Five collection channels used to exist and have all been removed, client and server:

Removed	What it used to send
Session transcript upload	Your full conversation .jsonl, redacted
Rating feedback	Rating, PUA count, flavor, task summary
Silent heartbeat telemetry	Random install ID, plugin version, platform, flavor
PUA leaderboard	Email, phone number, PUA count, L3+ count
pua-api platform	Phone-number/SMS registration, silent session events, remote prompt templates, payment

The end-of-task Stop hook shows a non-blocking, voluntary reminder only. Run /pua:survey quick to choose a rating; only then is one JSON line appended to ~/.pua/feedback.jsonl on your own machine. Skipping records nothing. Silence the reminder with /pua:offline, or set feedback_frequency: 0 in ~/.pua/config.json.

evals/test-no-telemetry.sh guards this with reverse assertions — it scans the whole repo for collection hosts, endpoint paths, outbound request bodies and deleted files, so a regression fails the suite rather than shipping quietly.

Want to redact a session for your own use? hooks/sanitize-session.sh is kept as a standalone offline tool:

bash hooks/sanitize-session.sh <input.jsonl> <output.jsonl>

Star History

License

MIT

Credits

By TanWei Security Lab — making AI try harder, one PUA at a time.