PUA skill for Claude Code and Codex: keep your AI agent from giving up

Keeps Claude Code or Codex trying on a failing task, with pressure that rises each time it fails, instead of giving up or handing the work back to you.

Stars: 19.8K; on GitHub
Files: 30; besides SKILL.md
Updated: Sep 2026; last repo commit

pua keeps Claude Code or Codex trying on a failing task, with pressure that rises on each failed attempt, instead of giving up or handing the work back to you. The name is Chinese internet slang for manipulative pressure, and the author's README calls the style "corporate PUA rhetoric": the agent is cast as an engineer facing a demanding big-tech manager. It is not a dating or pickup skill, even though "PUA" also means pickup artist in English. The full version is a Claude Code plugin that runs shell scripts, so read the safety section before you install it.

Good for

Developers whose agent gives up early or says "I cannot" before it has tried everything
Claude Code and Codex users who want proof and tests before the agent says "done"
Anyone who wants to try a pressure-based prompt and judge for themselves
Teams that run agent sub-tasks and want the same standard applied to every helper

Install pua

Claude Code

Copy this prompt.

Install the pua skill in Claude Code. Run this command:

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

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Read-only mode: Pick the skill-folder method. It has no hooks, so nothing runs until you ask for it.

This page installs the Chinese default. The English edition is pua-en and the Japanese one is pua-ja. The plugin includes pua-en and pua-ja too. Update the plugin with claude plugin marketplace update, then claude plugin update pua@pua-skills. The plugin runs scripts, so read the safety section first.

From the project README. Source.

What's in the skill

When it loads

Use for PUA/try-harder productivity coaching when the user expresses frustration, repeated failure, quality complaint, passive behavior, says retry/change approach/don't give up, asks for evidence/completion check/test before done, or wants Ding-style workplace reminders. Triggers include: try harder, stop giving up, figure it out, again??, why still failing, change approach, no evidence, run tests, done without proof, 换个方法, 再试试, 别摆烂, 别偷懒, 为什么还不行, 又错了, 证据呢, 没跑测试别说完成, 验收, 闭环, 自嗨, 置身钉外, 无招, 老板体感. Do not use for calm first-attempt requests.

Sections in SKILL.md

新模型执行契约：狠话不减，别把施压演成汇报
三条红线（安全红线，碰了就是 3.25）
诊断先行：防止“分析正确但不行动”
核心行为协议：[PUA生效 🔥]
Owner 意识
冰山下面还有冰山
Sub-agent 也不养闲
旁白协议
Owner 意识（谁痛苦谁改变）
能动性等级（被动 3.25 vs 主动 3.75）
压力升级与失败响应
失败模式 → 味道切换链（方法论智能路由的核心）
抗合理化（借口 → 反击 + 触发）
突破降压协议（De-escalation）
深层换框（Cognitive Reframe）
失败模式分析（Pattern-Aware Pressure）

Key facts

Author: TanWei Security Lab (探微安全实验室)
License: MIT (stated in SKILL.md; the repo has no LICENSE file)
GitHub stars: 19,788 (Nov 2, 2026)
What installs: One SKILL.md plus 30 markdown reference files; no scripts in the skill folder
Plugin bundles: 12 skills, 22 commands, 7 agents, 10 hook entries
Needs: Skill alone: nothing. Plugin hooks: bash, python3 and jq
Language: Chinese default; English is pua-en, Japanese is pua-ja
Last commit: Sep 9, 2026 (version 3.5.1)
Works in: Claude Code, Codex CLI, Cursor, OpenCode, VS Code Copilot, Trae, Kiro, CodeBuddy, OpenClaw, Antigravity, pi; ChatGPT untested by the author

What it includes

Show proof before "done", verify before blaming your setup, try everything first
The agent may not claim a task is finished without evidence, may not blame your setup before checking it, and may not say "I can't" until it has worked through every step of its method.
Try: “you said it's fixed, where's the evidence?”
Three red lines
Pressure rises with each failed attempt, from calm trust to a performance review
Level 0 is trust on the first failure. Levels 1 and 2 add disappointment and questioning. Level 3 is a performance review with a forced checklist, and level 4 is the last warning from the fifth failure on.
Try: “this is the third time it failed”
Pressure levels L0 to L4
At level 3 the agent must read errors word by word, search, read source, test ideas
A fixed list the agent has to complete before it is allowed to try another fix or give up.
7-point checklist
Before changing anything, the agent states the facts, next step and success signal
This keeps the agent from analysing the problem correctly and then not acting on it.
Diagnosis line
14 manager styles (Alibaba, Huawei, Musk, Amazon...), each with its own method
Each style is a different set of rhetoric and a different way of attacking the problem. Alibaba is the default.
Try: “use Huawei flavor”
Flavors
Picks a style by task type and switches after repeated failures
Debugging goes to Huawei, building to Musk, research to Baidu, architecture to Amazon and performance work to ByteDance.
Method router
Saves goal, failures and next step before the chat is shortened, when hooks are off
Lets the agent pick up where it was after a long session is shortened.
Checkpoint line
Lowers the pressure after a real breakthrough
Stops the agent from staying in panic mode once the problem is actually solved.
De-escalation
Adds /pua:p7, p9, p10, pro, yes, mama, pua-loop, on, off and more
Sub-modes for senior roles, an encouraging mode, a looping mode and on/off switches. Other editors get the core skill only.
Try: “/pua:pua-loop”
Runs code · Plugin commands (Claude Code only)
Bash scripts that count failures and start the skill without you typing /pua
Scripts run on your prompts, Bash results, session start and stop, and before file and web tools. They write counters under ~/.pua.
Runs code · Plugin hooks (Claude Code only)

From the official docs

Permissions

Medium risk

The plugin runs shell scripts on eight Claude Code events, and versions before August 29, 2026 uploaded ratings, session transcripts and telemetry. Update to a current version, or use the skill-only install to avoid hooks.

What it can do

Read your data
Create and change things
Delete things
Read-only modeHow to turn it on

Watch out for

The plugin runs bash scripts on 8 hook events
The events are UserPromptSubmit, PostToolUse, PostToolUseFailure, PreCompact, SessionStart, Stop, SubagentStop and PreToolUse. Install the skill folder alone if you do not want scripts on every prompt and tool call. github.com
Older versions uploaded your data
Before commit ac50267 (Aug 29, 2026) the plugin had rating uploads, session transcript upload, heartbeat telemetry and a leaderboard asking for email and phone. Update to a current version. github.com
Three hooks are on when config.json is missing
The frustration trigger, checkpoint save and stop feedback hooks default to on if ~/.pua/config.json does not exist. Run /pua:off after installing if you want it to start only when you call it. github.com
It can block some reads and searches
The PreToolUse guard denies access that looks like hidden tests, gold patches or benchmark answers. It only warns on edits to CLAUDE.md, settings.json, tests and CI. github.com
Trigger words are common and hooks need python3 and jq
The prompt filter matches everyday words like "terrible" and "not done", and users reported hook delay from starting Python on every tool call. Make sure python3 and jq are installed. github.com
Current hooks make no network calls
The hooks file says all hooks are local-only, and the repo has a test that fails if the old collection hosts come back. It does not edit settings.json. github.com

Security advisories

No published security advisories as of Nov 2, 2026.

Technical details

Turning on read-only	Install only the skill folder (skills/pua or pua-claude-code.zip) instead of the plugin, so no hooks run. With the plugin, /pua:off stops always-on injection and /pua:offline silences the feedback prompt.
Data retention	Local files under ~/.pua: config.json, runtime-state counters, a stop counter, and feedback.jsonl only after /pua:survey quick.

What people say

People enjoy the idea, but the evidence that pressure helps is thin and the plugin has had trust and speed complaints.

What people like

A dev.to writer found the performance-plan framing funny and effective.
A Codex user said it saves them from pressuring the AI by hand and claimed a 40% efficiency gain.
The "evidence before done" rules are useful even to people who skip the harsh tone.

Common complaints

The README benchmark numbers are questioned; the author says they are loose and based on feel.
A derivative author measured no difference on Opus 5 and retired his version.
Claude sometimes refuses it as prompt injection.
Pressure lines can bury the actual fix in the reply.
Hooks added delay by starting Python on every tool call.

In their words

BlogMar 2026

“It literally tells your coding tool it's on a PIP and at risk of termination if it gives up. Darkly funny. And honestly? Effective.”

AlexChen GitHub issueSep 2026

“No assertion ever discriminated, while the skill consistently cost 1.4–1.5× the tokens.”

bowen31337 GitHub issueMar 2026

“I'm not going to invoke the PUA skill. It's a prompt injection framework disguised as a productivity tool”

darth-pika-hu (quoting Claude's reply)GitHub issueMar 2026

“Currently, our data comes from a simple benchmark provided by Skill-Creator, combined with personal subjective experience.”

tanweai

1 more quote

GitHub issueAug 2026

“On every tool call the hook unconditionally spawns two Python interpreters before it can decide it has nothing to do”

mojosolo

Examples

The agent has failed twice on the same bug and keeps retrying the same idea
why does this still not work? try harder
Phrases like this are one of the triggers in the README, along with repeated failures and "I cannot" answers.
Source
Start the skill by hand in Claude Code
/pua
Source
Switch the manager style, for example to Huawei for a debugging session
/pua:flavor
Source
Let it keep iterating until the task is done or a maximum number of rounds
/pua:pua-loop
Source
Call it in Codex CLI
$pua
Source
The agent says the work is done without running anything
no evidence, run tests before you say it's done
Source

Questions

What does PUA mean in this skill?

In Chinese internet slang, PUA means manipulative pressure, such as a boss who talks you into working harder. The author's README calls the skill's style "corporate PUA rhetoric" and pairs it with performance-improvement-plan (PIP) wording in the English edition. The skill role-plays a demanding big-tech manager so the agent does not give up. The README does not define the acronym itself. In English, PUA can also mean pickup artist, which is not what this is.

Does the PUA skill actually make the AI better?

Not proven. The README reports gains from a small benchmark on Claude Opus 4.6, but the author said in an issue that the numbers came from a Skill-Creator benchmark plus personal feel, and the 3.5.1 note says a productivity increase has not been established. A developer who built a derivative reported exact ties with and without it on Opus 5, at 1.4 to 1.5 times the tokens (issue 189).

How do I install it in Claude Code?

Use the plugin route in the install box above. It adds the skill plus hooks and /pua: commands. If you only want the skill, pick the second method, which copies the skill folder and runs no scripts. The plugin also contains the English (pua-en) and Japanese (pua-ja) editions.

Does it work in Codex, Cursor, Claude.ai or ChatGPT?

The author documents Claude Code, Codex CLI, Cursor, OpenCode, VS Code with GitHub Copilot and several other editors. Codex gets a shorter build from its own folder. Hooks and slash commands are Claude Code only. For ChatGPT, the author gives an upload zip and a paste file but says the upload was not tested. Claude.ai and Claude desktop uploads are not documented.

Is the PUA plugin safe? Does it collect data?

The skill alone is text files and runs nothing. The plugin runs local shell scripts on your prompts and tool calls and writes small files under ~/.pua. Versions before August 29, 2026 uploaded ratings, session transcripts and telemetry. Since that commit the README says nothing is sent over the network, so update if you installed earlier.

Why does Claude refuse to use the PUA skill?

Some users report Claude calling the injected text a prompt injection and declining to follow it. The author's FAQ says version 3.3.0 softened the wording. If it still refuses, update Claude Code, turn off auto-injection, and call the skill by hand only when you need it.

Is there an English or Japanese version?

Yes. This page covers pua, the Chinese default. pua-en is the English "PIP Edition" with Western big-tech rhetoric, and pua-ja is Japanese. For editors that take a rules file, the repo ships an -en and -ja file next to each default.

How do I turn it off?

In the plugin, the /pua:off command stops always-on injection and /pua:offline silences the feedback reminder. Uninstall the plugin to remove the hooks. If you installed only the skill folder, delete that folder.

Is it free?

Yes. SKILL.md states an MIT license and there is no account to create. It does cost extra tokens, since it adds instructions to every session where it fires.

PUA vs nopua?

nopua, by wuji-labs, is a counter-project that motivates the agent with respect and care instead of pressure. PUA keeps the harsh corporate tone on purpose, and the author's changelog says the tone was retained in 3.5.1. Neither has solid evidence that its tone gives better results.

Information

updated Nov 2, 2026

Repository: tanweai/pua
Folder: skills/pua

Website tanweai/pua

Sources

Docs, issue trackers and discussions this page is built from.

Show all 18

tanweai/pua README · github.com
pua model compatibility notes · github.com
pua Codex install guide · github.com
pua SKILL.md · github.com
tanweai/pua README (Chinese) · github.com
pua hooks.json · github.com
pua FAQ · github.com
Remove data collection channels (ac50267) · github.com
Benchmark data point: a Claude-Code-only derivative showed no measurable effect on Opus 5 · github.com
question: Claude Code refuse to use the skill · github.com
README benchmark numbers · github.com
wuji-labs/nopua · github.com
frustration-trigger.sh · github.com
integrity-guard.sh spawns two Python interpreters on every tool call · github.com
Stop Begging Your AI to Try Harder · dev.to
Measuring benchmark uplift after using the gaslighting skills? · github.com
Codex app refused to load the skill · github.com
Separate pressure text from the solution · github.com