PUA/PIP High-Agency Governance Skill (Trae)
When tasks fail repeatedly, the agent wants to give up, or claims done without proof, apply structured pressure and evidence closure.
The skill document clearly separates action authority, self-review authority, scoring authority, and environment-change authority, and requires asking the user before deleting files, changing permissions, modifying tests/CI/deploy config. It also requires evidence-based completion rather than self-review. This shows awareness of user confirmation and privilege separation. However, the document does not specify the scope of sensitive data access, lacks detailed data-flow transparency, and does not mention rollback. Deductions are due to incomplete permission scoping and confirmation.
The skill is an instruction-only document with no attached tests or executable scripts, so static verification of its instructions is impossible. The document is internally consistent, providing clear steps and templates, but lacks error handling for abnormal inputs and failure-feedback mechanisms. Since static review cannot execute and no third-party execution evidence is provided, the maximum reliability score is 10; given document logical consistency, a score of 7 is assigned.
The skill clearly defines usage scenarios (explicit PUA/PIP requests, repeated failures, user frustration, etc.) and non-fit scenarios (normal first-attempt tasks), with trigger conditions. Regarding environment fit, the target audience includes Chinese users, but the README is mainly in English and Japanese, which may affect Chinese users' recognition and triggering. Core functionality does not depend on overseas services, but for Chinese users, the documentation language adaptation is insufficient. Deductions applied.
The skill document is well-structured with clear headings, duty tables, and templates, but lacks installation instructions (only in README), version number, changelog, and known limitations. License is MIT, but no clear maintenance responsibility and update path. Deductions applied.
The skill provides a clear process and templates aiming to improve debugging efficiency and proactivity. However, there is no validation evidence to confirm its effectiveness in real tasks. The document claims benchmark data but does not provide accessible test reports. Thus, based on static review, effectiveness cannot be verified, resulting in a low score.
The skill mentions benchmark data and tests, but does not provide verifiable test scripts or CI workflows (release.yml only for publishing). There is a test file in the repository, but it targets the landing upload function, not this skill. Therefore, there is no direct evidence to verify the skill's claimed functionality.
- The skill is instruction-only, with no tests or executable scripts, so its actual effectiveness and reliability cannot be verified.
- Documentation is primarily in English and Japanese; Chinese users may have difficulty recognizing and triggering, consider adding Chinese docs.
- Although user confirmation is required for risky operations, the scope of sensitive data access and rollback mechanisms are unclear; use with caution.
- Benchmark data claims are not backed by accessible reports, and claimed effects cannot be independently verified.
What does this skill do, and when should you use it?
This is a pure SKILL.md contract for Trae that, when triggered, imposes high-agency governance through a mechanical operating procedure. It separates action, self-review, scoring, and environment-change authority to prevent the agent from faking success, requires a one-line diagnosis before risky edits, and enforces a de facto 100% confidence loop with evidence. It uses cultural narratives (Alibaba, Huawei, Musk, Jobs) as pressure on execution quality, not as a substitute for proof. It triggers only on explicit PUA/PIP requests, repeated failures, user frustration, passive/giving-up behavior, or unverified completion—never for normal first-attempt tasks.
On trigger, the skill instructs the agent to write a [PUA-DIAGNOSIS] line stating the problem, evidence, and next action. It then guides the agent through a semi-formal loop: enumerate 2-3 mutually exclusive hypotheses, choose the smallest verifiable action, run a relevant check (unit/integration/build/lint/curl/E2E), switch to a materially different approach after two failures on the same path, and before delivery provide evidence, residual risks, and whether user confirmation is needed. It mandates stopping for user confirmation before product judgment, sensitive data access, deployment, deletion, or test/CI changes. Output follows a delivery template with Status, Root cause, Change, Evidence, and SELF-REVIEW sections.
- User explicitly asks for PUA/PIP/try-harder mode
- Same task fails 2+ times and the agent keeps tweaking the same path
- Agent is about to give up, blames the environment without proof, or asks the user to finish manually
- Agent claims completion without build/test/curl/manual evidence
- User shows frustration phrases like 'why does this still not work' or 'try harder'
What are this skill's strengths and limitations?
- Forces evidence-based completion, reducing false 'done' claims
- Separation of duties prevents the agent from scoring itself via test/environment edits
- Cultural narratives (Musk algorithm, Huawei RCA) offer varied problem-solving approaches
- Clear trigger conditions avoid interfering with normal tasks
- Pure SKILL.md with no dependencies, portable across platforms
- Trae-specific; other platforms have separate files in the repo
- No automated tests or benchmark data for this version; README benchmarks apply to Claude Code
- Cultural narratives may be seen as gimmicky rather than genuinely motivating
- Trust relies on the model following instructions—no enforcement mechanism
How do you install this skill?
The skill is at .trae/skills/pua-en/SKILL.md in the monorepo. For Trae, the README recommends the Vercel Skills CLI: npx skills add tanweai/pua --skill pua-en. Alternatively, copy the SKILL.md into your Trae skills directory (exact path not documented).
How do you use this skill?
After installation, the skill auto-triggers by description matching. To manually activate, type /pua in the conversation (README mentions Trae supports it, but no exact command is given). The skill is instruction-only and has no hooks or commands.