PUA High-Agency Governance Skill (Trae Edition)
Uses big-tech PUA rhetoric and verification-forcing procedures to make the AI exhaust every option before giving up, especially when tasks fail repeatedly or users are frustrated.
Evidence: The skill is instruction-only, no code execution, no elevated permissions requested. SKILL.md explicitly requires user confirmation before deleting files, changing permissions/tests/CI, or handling sensitive data. Document also emphasizes respect and simplicity to the user, and the upload API tests show explicit consent required for anonymous uploads. However, the data contribution section encourages users to upload session logs (.jsonl), which involves external data flow, and the skill description itself does not warn about this data flow. Also, the repo includes upload code and tests, but this skill path does not disclose its data flow. Deduction: permission handling is complete, but data-flow transparency is incomplete.
Evidence: SKILL.md provides clear procedures such as diagnosis first, factual 100% confidence loop, and delivery template, which are logically consistent. But no executable tests or key-path reproduction evidence are provided. The skill is instruction-only, cannot be executed in static review, and relies on external tools (e.g., CI, E2E) for verification, but these are not included in the skill file. Error handling is defined but lacks detail on abnormal input (e.g., behavior if user does not confirm). Deduction: happy path is plausible but tests, edge cases, and error handling are thin.
Evidence: The skill targets high-agency governance scenarios, clearly defines trigger conditions (explicit user request, repeated failures, frustration, giving-up behavior, etc.) and non-triggers (normal first-attempt tasks), with precise semantic triggers. The skill is in Chinese, suitable for Chinese users, and uses MIT license with no overseas service dependencies. However, the documentation heavily references corporate culture narratives that may not suit all users. Also, capability boundaries are stated simply, without detailed non-fit ranges (e.g., non-coding tasks). Deduction: triggers and scope are clear, but boundaries and non-fit scenarios are limited.
Evidence: SKILL.md has clear frontmatter (name, description, license, compatibility), good information architecture, provides triggers, procedures, and delivery template. But it lacks installation/dependency notes (README has them, but the skill file itself doesn't), lacks version number, changelog, maintenance responsibility, and update path. It has MIT license, but publisher is unverified, and known limitations are not disclosed in the skill file (only in README). Deduction: readable and usable but missing versioning/governance and troubleshooting gaps.
Evidence: The skill provides concrete procedures to improve agent's agency, which is well-designed, and README contains benchmark data claiming improvements, but these are author claims, not independently verified, and static review cannot verify actual effect. The core task is to prevent agent from giving up, and the templates may help users get more complete outputs, but marginal value depends on agent compliance and may add unnecessary complexity. Deduction: core task seems plausible, but direct evidence of directly usable output is limited, and possible added time.
Evidence: The repo includes test cases for upload functionality (upload-function.test.ts) which may relate to the skill's data upload, but this is not the key path of the skill itself. The benchmark data in README lacks reproducible experimental steps or raw data, only screenshots. There is no CI workflow specifically testing the skill. Deduction: auditable primary material (tests) exists but covers only peripheral functionality; core claims are not reproducibly verified.
- The skill encourages users to upload session logs (.jsonl) to an external service openpua.ai for improvement, which involves external data flow. Assess sensitive data risks and ensure explicit consent before use.
- The skill documentation includes corporate culture narratives (Alibaba, Huawei, ByteDance) with emotionally charged pressure language, which may not suit all users or team cultures.
- This is a static review; no execution tests were performed. Claims of effectiveness are unverified. Trial in a small scope before adoption.
What does this skill do, and when should you use it?
This is a Trae-compatible Agent Skill whose core is a mechanically enforceable behavioral contract rather than mere encouragement. It enforces a separation of powers (action, self-evaluation, scoring, environment modification) to prevent the AI from masquerading “looks done” as “actually done”; it mandates a diagnosis-first workflow, and it requires a “factual 100% confidence loop” with evidence at every step. It triggers only when the user explicitly asks for PUA, the same task has failed twice or more, the user shows frustration, or the AI shows signs of giving up. It applies to debugging, implementation, configuration, deployment, and similar tasks. The Skill is part of a monorepo bundling 45 skills; this profile describes only the .trae/skills/pua/SKILL.md file.
When triggered, the Skill executes these operations: 1) Outputs a one-line [PUA-DIAGNOSIS] in the format “Problem is ___; evidence is ___; next action is ___”, and requires subsequent actions to align with that diagnosis. 2) Enforces separation of powers: the current agent only reads code, changes business implementation, and runs verification; it must not directly modify tests/CI/scorers to fabricate a pass. Self-evaluation is output as a SELF-REVIEW listing evidence, risks, and uncovered items, and must not state “I think it's done” as final fact. Scoring is decided by external commands, user acceptance, CI, or E2E results. Environment changes (deleting files, changing permissions, modifying tests, altering deployment config) must be explained and confirmed first. 3) Runs the “factual 100% confidence loop”: list 2-3 mutually exclusive hypotheses, pick the smallest verifiable action, run at least one relevant local verification (unit/integration/build/lint/curl/E2E), switch to a fundamentally different path if it fails twice, and before delivery output an evidence list, uncovered risks, and why you didn't keep asking the user. 4) Binds each cultural narrative (Alibaba, Huawei, ByteDance, etc.) to a concrete engineering action. 5) Finally, outputs a delivery template with conclusion, evidence, and SELF-REVIEW. The entire process depends on no external commands or scripts; it guides the agent purely through the SKILL.md text.
- When the user explicitly says “try harder,” “try again,” or “switch approach,” and you want the AI to not give up easily.
- When the same task has failed more than twice and the AI keeps fine-tuning in the same direction, forcing a change of strategy.
- When the AI is about to say “cannot complete,” “suggest manual handling,” or blame the environment without verification.
- When the AI claims completion but lacks build/test/curl/manual acceptance evidence, forcing it to provide verification.
- When the user is frustrated by repeated failures (e.g., “why does this still not work”) and the Skill resets behavior.
- In Trae environments (which lack Claude Code's hooks or subagents) where a mechanically executable procedure is needed to enforce high agency.
What are this skill's strengths and limitations?
- Replaces vague encouragement with a mechanically executable procedure, explicitly separating powers to prevent fake completions.
- Mandates diagnosis before action and requires alignment between diagnosis and action, avoiding aimless work.
- The 100% confidence loop requires listing mutually exclusive hypotheses and running real verification, significantly improving evidence quality.
- Binds cultural narratives to engineering actions, balancing motivation with practical steps.
- Clear trigger conditions ensure it doesn't interfere with normal first-time tasks.
- Relies solely on SKILL.md text and cannot use Claude Code's hooks, slash commands, or subagents; enforcement strength depends on model adherence.
- Not fully automatic; requires the user or model to recognize triggering scenarios. If the model ignores it, it has no effect.
- Has no built-in test suite or validation scripts to prove effectiveness.
- The repository README showcases integrations with other platforms (Claude Code, Codex, etc.), but this Trae version may lack those advanced features (e.g., PUA Loop, always-on).
- The SKILL.md states MIT license, but the repository's LICENSE file may differ; confirm before using.
How do you install this skill?
As this Skill is part of a monorepo, you must obtain the whole repository. Use npx skills to install the collection and ensure .trae/skills/pua/SKILL.md is placed in Trae's skills directory. Specifically, run: npx skills add tanweai/pua --skill pua-trae -a trae -y. If the new skill isn't picked up immediately, restart Trae.
How do you use this skill?
When trigger conditions are met (e.g., user requests PUA, task fails twice or more), the Skill loads automatically. Users can also manually trigger by expressing PUA-related intent in the conversation. Once triggered, the agent first outputs a [PUA-DIAGNOSIS] line, then follows the separation of powers, diagnosis-first workflow, and 100% confidence loop, and finally outputs a delivery template with conclusion, evidence, and SELF-REVIEW. No extra user action is needed, but the agent will pause and ask for confirmation when product judgments, sensitive data, deployment, file deletion, or test/CI changes are involved.