tanweai/pua

pua – An “AI Coding Agent” Skill that pushes LLM‑based code assistants to keep trying

What it is

• A plug‑in/skill package that can be added to a variety of AI‑coding assistants (Claude Code, OpenAI Codex CLI, Cursor, Kiro, CodeBuddy, OpenClaw, Google Antigravity, OpenCode, VS Code Copilot, etc.).
• The skill injects a set of “PUA” (Performance‑Improvement‑Plan / corporate‑rhetoric) rules that make the assistant behave more persistently: it refuses to give up, avoids blaming the user, and proactively checks for hidden issues.
• The rules are expressed as language‑specific “skill files” (SKILL.md, .mdc, .json, etc.) that each platform reads to auto‑trigger the behaviour, or they can be invoked manually with commands such as /pua.

Core ideas

Idea	How it works
PUA Rhetoric	The skill frames the assistant’s internal prompts with performance‑review language (e.g., “You must exhaust every possible solution before saying you can’t”).
Debugging methodology	A 7‑point checklist (close‑the‑loop, fact‑driven, exhaust‑everything, etc.) that the assistant follows when it detects repeated failures.
Proactivity enforcement	After fixing a bug the agent is forced to scan related code, run builds/tests, and verify the fix instead of simply replying “done”.

Supported platforms

• Claude Code (plugin marketplace)
• OpenAI Codex CLI (SKILL.md standard)
• Cursor (pua.mdc)
• Kiro (steering file or SKILL.md)
• CodeBuddy, OpenClaw, Google Antigravity, OpenCode, VS Code Copilot, pi.dev, Trae – all via the same skill definition.

Key features

• Auto‑trigger when the assistant fails twice in a row, starts blaming the user, or shows “busy‑work” patterns.
• Manual trigger via /pua (or /pua:on, /pua:off, /pua:loop, etc.).
• Multi‑language packs – Chinese (Alibaba/ByteDance/Huawei style), English (PIP/Amazon‑style), Japanese, each with culturally‑adapted phrasing.
• Special modes – encouragement mode, “Chinese‑mom” nagging, tech‑lead delegation, always‑on enforcement.
• Benchmark data in the README claims higher fix counts, more verification steps, and more tool calls compared with a baseline Claude Opus 4.6 run.

Typical use‑case flow

1. Install the skill for your preferred coding assistant (e.g., npx skills add tanweai/pua --skill pua-en).
2. Work on a debugging or implementation task as usual.
3. If the assistant starts to give up or repeats the same approach, the skill auto‑activates, runs the 7‑point checklist, and forces the model to explore alternative hypotheses, read logs, and verify the result.
4. The assistant returns a more thorough solution, often uncovering hidden bugs or configuration issues.

Installation snapshot

# Claude Code
claude plugin marketplace add tanweai/pua
claude plugin install pua@pua-skills

# Codex CLI (one‑liner)
curl -o ~/.codex/skills/pua/SKILL.md \
  https://raw.githubusercontent.com/tanweai/pua/main/codex/pua/SKILL.md

(Other platforms follow the same pattern – copy the appropriate rule file into the tool’s skill directory.)

Limitations / caveats

• The skill is essentially a prompt‑engineering overlay; it does not add new code‑analysis capabilities beyond what the underlying LLM already provides.
• Reported performance gains are based on the author’s internal benchmarks; independent verification is not provided.
• Because it forces the model to keep trying, runtimes can increase (e.g., the “Passive config review” test shows more steps and longer total time).
• The heavy corporate‑rhetoric style may feel noisy or overly strict for some users.

Who might benefit

• Developers who rely heavily on AI coding assistants for debugging and want the model to avoid early “I can’t solve this” exits.
• Teams that use Claude Code, Codex CLI, or similar agents in CI‑like workflows and need more systematic, verification‑heavy behaviour.
• Users interested in experimenting with prompt‑level “performance‑review” conditioning for LLM agents.

---

All information above is taken directly from the repository’s README; no additional features have been inferred.