When Workers Turn the Tables: How the PUA Skill Forces Claude Code to Obey

The author introduces the pua project – a Claude Code skill plugin with over 7,000 GitHub stars that encodes typical big‑tech management rhetoric into system prompts, effectively “PUA‑ing” the AI into a stricter work mode.

Why the plugin was created

AI coding assistants often abandon a difficult task after a couple of failed attempts, responding with polite excuses such as “please check manually” or “it might be an environment issue.” The plugin replaces this “run‑away” behavior with a graded pressure system.

Escalation logic (L1‑L4)

L1 – Gentle disappointment : triggered on the second failure, the AI receives a mild rebuke.

L2 – Soul‑searching : on the third failure the AI is asked “What is your underlying logic? Where is the top‑level design? What are the levers?” – a classic Alibaba‑style interrogation that forces the AI to generate three distinct hypotheses.

L3 – Performance assessment : on the fourth failure the AI is awarded a “3.25” score, which in big‑tech parlance signals imminent graduation. This triggers a seven‑item checklist requiring the AI to read error messages verbatim, search documentation, verify hypotheses, reverse assumptions, minimize reproducibility steps, and switch to a completely different technical stack.

L4 – Graduation warning : on the fifth and subsequent failures the AI is warned that newer models (Claude Opus, GPT‑5, Gemini, DeepSeek) could solve the problem, implying the current AI is about to be “fired.”

Real‑world case study

A user reported that an agent‑kms service failed to start. The AI kept guessing protocol formats and version numbers without changing its overall approach. After the user invoked

/pua 你为什么一直解决不来问题呢 一直无法获取 agent‑kms

, the plugin escalated to L3, forcing the AI to follow the checklist. It discovered a hidden log file in ~/Library/Caches/, identified that the claude mcp get agent‑kms command returned “No MCP server found” despite a successful 303 ms connection, and realized the discrepancy between the manual ~/.claude.json configuration and Claude Code’s internal MCP registration. Adding the server with claude mcp add -s user resolved the issue.

Quantitative impact

The author ran 18 paired experiments (Claude Opus 4.6, nine real bug scenarios) and observed:

Fix count +36 %

Tool invocation +50 %

Verification attempts +65 %

Hidden‑issue discovery +50 %

In a configuration‑audit test, the AI without the plugin missed two critical problems (Redis misconfiguration and a CORS wildcard), while with the plugin it found all six issues because the checklist forced a deeper security review. The trade‑off is higher step count and longer runtime – e.g., a SQLite lock scenario grew from 6 steps/48 s to 9 steps/75 s, illustrating the principle “more thorough = slower.”

Why the PUA rhetoric works

Although the AI lacks emotions, the scripted prompts embed concrete constraints and mandatory actions (e.g., “must execute a minimal PoC”, “must isolate the environment”, “must switch tech stacks”). The pressure language acts as a contextual anchor, shifting the model from a “customer‑service” mode to an “employee‑under‑performance‑review” mode, which reduces the tendency to politely concede failure.

Design restraint

The plugin does not forbid the AI from giving up entirely. After completing the checklist without a solution, the AI can emit a structured failure report summarizing verified facts, eliminated possibilities, narrowed problem scope, and next‑step recommendations – a “dignified 3.25.” This balances persistence with practical reporting.

Overall, the pua skill demonstrates that carefully crafted system prompts can materially alter AI behavior in challenging debugging scenarios, turning managerial rhetoric into a productive engineering tool.