Trae High-Agency Governance Skill
When your AI keeps failing or wants to give up, enforce a mechanically executable protocol to exhaust all options before delivering.
Evidence: The skill is instruction-only, with no actual code execution hooks. SKILL.md emphasizes separation of 'action rights' and 'evaluation rights', prohibits modifying tests/CI to pass, and requires explanation and confirmation before sensitive operations (delete files, change permissions, deploy). An explicit consent mechanism for data upload is present (see test). Deductions: lacks detailed least-privilege specification (e.g., permission scope), and as an instruction skill, actual execution depends on base agent's permissions; no explicit data retention or deletion policy. Environment modification rights are well-defined, but user revocation mechanism not specified.
Evidence: SKILL.md is self-consistent, with clear definitions for triggers, diagnostics, and delivery template. Provides explicit failure-feedback mechanism (SELF-REVIEW) and diagnostic steps. However, no tests or independent verification coverage; static reading cannot run verifications. Deductions: key paths not reproduced, limited error-handling details (e.g., specific commands and assertions), actual reliability needs runtime evaluation.
Evidence: Description explicitly specifies trigger scenarios and conditions for not triggering. SKILL.md states Trae optimization, includes Chinese language support, and Trae is accessible in mainland China. Deductions: boundaries clear, but no actual usage cases or feedback. Trigger scenarios are specific, covering failure, frustration, and completion verification. Clear non-fit scope provided.
Evidence: SKILL.md is well-structured, with frontmatter (name, description, license, compatibility), installation instructions (README provides multi-platform install), versioning via git, and MIT license. Deductions: lacks dedicated changelog or known limitations section; maintainer information is mentioned in README but not explicit maintenance responsibility; dependency notes only on Trae, no other dependencies.
Evidence: Static reading cannot verify actual effectiveness. Value proposition is reasonable (structuring to improve proactivity and verification). Deductions: no actual usage cases or user feedback, no direct benchmark evidence, direct usability cannot be assessed.
Evidence: Repository contains CI config (release.yml) and unit tests (for upload API), but these do not cover the skill's key paths (skill is instruction-based). Deductions: no tests for the skill itself, reliability claims unverified, benchmark data in README without provenance.
- This skill is instruction-only; actual effectiveness depends heavily on the base agent (e.g., Trae) permissions and behavior, requiring verification in real environments.
- Trigger conditions are clear, but it may generate excessive pressure (emotional manipulation); use with caution in tasks requiring psychological safety or creativity.
- The upload feature asks users to submit session logs; despite consent mechanism, ensure users understand what is uploaded and why.
- The skill emphasizes 'not giving up', which may prolong debugging and increase costs; users should weigh accordingly.
What does this skill do, and when should you use it?
This is a PUA-style governance skill optimized for Trae, rewriting the governance boundaries of the Claude Code version into a pure SKILL.md protocol. It does not rely on Claude Code hooks or agents, but instead uses separation of powers, diagnosis-first, factual 100% confidence loop, and cultural narrative binding to force the AI to output evidence, root cause, and SELF-REVIEW when faced with repeated failures, user frustration, or unverified completion. It triggers only on explicit PUA requests, 2+ consecutive failures, or signs of giving up, leaving normal first-attempt tasks untouched.
This skill defines a behavioral contract for Trae: when loaded, it instructs the AI to output a [PUA-DIAGNOSIS] line, enforce separation of powers (action, self-evaluation, scoring, environment modification), run a loop of mutually exclusive hypotheses and minimal verifiable actions, and finally produce a delivery template with status, root cause, changes, evidence, and SELF-REVIEW. It requests user confirmation before sensitive operations. It does not execute any external commands itself; it only governs the AI's behavior via rules.
- When the same task fails twice consecutively and the AI keeps tweaking the same path, enable this skill to force a fundamentally different approach.
- When the AI is about to say 'I can't' or 'handle it manually' without verifying, enable this skill to require verification before attribution.
- When the user explicitly asks to 'PUA', 'try harder', 'try another way', or 'try again', enable this skill to load the governance rules.
- When the AI claims completion but lacks build/test/curl/manual acceptance evidence, enable this skill to demand an evidence list.
- When the AI passively waits, fixes only surface issues, or skips verification, enable this skill to push proactive module scanning.
What are this skill's strengths and limitations?
- Writes governance rules as mechanically executable procedures, effective in Trae without Claude Code-specific mechanisms.
- Clearly separates action power from scoring power, preventing 'looks done' masquerading as 'actually done'.
- Enforces diagnosis-first and evidence checklists, reducing blind rework.
- Binds cultural narratives to engineering actions, balancing motivation with pragmatism.
- Multi-language support (this version is Chinese) and clearly scoped trigger conditions to avoid interfering with normal tasks.
- Contains only SKILL.md, no Claude Code hooks or subagents, so it cannot automate pressure escalation or code-level detection.
- No built-in tests or benchmark data; effectiveness is unverified.
- Trigger conditions rely on description matching, which may be imperfect, requiring explicit user instructions as fallback.
- Emphasizes 'pressure only on yourself', but actual compliance depends on the AI's understanding and adherence, which may vary.
- Installation only mentions npx skills, not manual folder placement, which may be less direct for some users.
How do you install this skill?
Install via Vercel Skills CLI: npx skills add tanweai/pua --skill pua-trae -a trae -y, then restart Trae. See the README's Trae section for more details.
How do you use this skill?
After installation, the skill triggers automatically by description matching; you can also explicitly invoke it in conversation. Once triggered, the AI will follow SKILL.md to output a diagnosis, run verification, and finally provide a delivery template (conclusion, evidence, SELF-REVIEW). Note that normal first-attempt tasks will not trigger it.
How does this skill compare with similar options?
Compared to the Claude Code version of the PUA skill (pua in the main repo), this Trae version strips out Claude Code-exclusive hooks, slash commands, subagents, and Stop feedback, reducing it to a pure SKILL.md protocol. It is lighter but lacks the automation and depth of the Claude Code version.