What does this skill do, and when should you use it?
TSUME is the Japanese edition from the pua skill collection, targeting AIs that repeatedly fail, give up too early, or act passively. It blends high-pressure Japanese corporate rhetoric (Toyota's genchi genbutsu, Dentsu's Oni-Jukkan) with universal debugging methodologies (5-step method, 7-item checklist) to force AI to exhaust every avenue before quitting. The skill features L1–L4 pressure escalation, a library of proactive-enforcement phrases, and an excuse-blocking table. Once installed, it auto-triggers under set conditions or can be invoked manually.
Reads trigger conditions (e.g., 2+ consecutive failures, user frustration) and activates in conversation. It then outputs Japanese corporate-style pressure phrases and enforces structured steps: diagnose the stuck pattern, conduct a five-dimensional investigation (read error messages, search proactively, read primary sources, verify assumptions, invert assumptions), complete a 7-item checklist, and finally run a retrospective. It escalates pressure based on failure count—from mild disappointment to graduation warning—and counters common excuses like 'beyond my ability' or 'environment issue' with tailored rebuttals.
- When an AI has failed twice with the same approach, TSUME activates and forces a switch to a fundamentally different strategy.
- When an AI is about to say 'I can't solve this' or suggests manual handling, the skill escalates to higher pressure levels requiring exhaustive effort.
- When an AI passively waits for instructions and doesn't search or read source code, it pushes with phrases like 'lack of self-drive' to induce initiative.
- When a user expresses frustration like 'try harder,' the skill detects it and automatically initiates pressure escalation.
- In an Agent Team, the Leader can use TSUME to manage global pressure levels for all teammates and coordinate based on failure reports.
How do you install this skill?
- The skill contains strong psychological manipulation elements that may negatively affect the agent; ensure safety boundaries and controllability.
- Reliance on overseas services may cause accessibility issues for mainland China users; consider localization or alternatives.
- Broad scope may lead to false triggers; evaluate applicability based on actual task.
- Benchmark data likely comes from the author and lacks independent verification; assess cautiously before use.
- Shell / CLI
This skill is part of the tanweai/pua repository, which bundles 45 skills; install the whole collection first. For Claude Code:
claude plugin marketplace add tanweai/pua
claude plugin install pua@pua-skillsOr via Vercel Skills CLI for the Japanese skill:
npx skills add tanweai/pua --skill pua-jaSee the repo README for platform-specific paths and updates.
How do you use this skill?
After installation, the skill auto-triggers based on its description. For example, if the AI fails repeatedly and tries to give up, it will be invoked automatically. You can also manually trigger it with /pua:pua-ja in Claude Code or $pua in Codex. It will immediately output pressure phrases and enforce the corresponding checklist and methodology.
What are this skill's strengths and limitations?
- Provides systematic debugging methodology (5-step), not just empty pressure; genuinely guides deep troubleshooting.
- Rich pressure levels and phrase library adapted to various failure modes, with authentic Japanese corporate culture immersion.
- Includes an 'excuse-blocking' table that precisely counters common excuses and reduces AI evasion.
- Integrates with Agent Team for team-level pressure management.
- May exert excessive pressure, leading to resource waste (e.g., endless retries); consider setting limits.
- Japanese edition is only useful for Japanese-language environments; non-Japanese users won't benefit.
- May occasionally trigger for simple tasks or first failures despite conditions; potential false positives.
- The repository lacks explicit test coverage; actual effectiveness is not systematically verified.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| TSUME (詰め) — Proactivity-Enforcing Skill for Stubborn AI this page | 50 · Use with care | ★ 20k | 1mo ago | — |
| PIP Skill – Put Your AI on a Performance Improvement Plan | 45 · Not recommended | ★ 20k | 1mo ago | — |
| PUA High-Agency Governance Skill (Trae Edition) | 51 · Use with care | ★ 20k | 1mo ago | — |
| PUA Universal Motivation Engine | 41 · Not recommended | ★ 20k | 1mo ago | — |
| PUA/PIP High-Agency Governance Skill (Trae) | 49 · Use with care | ★ 20k | 1mo ago | — |
Compared to the English PIP edition (pua-en) and Chinese edition (based on Alibaba 361), this version leverages unique Japanese corporate practices like Toyota and Dentsu, emphasizing 'genchi genbutsu' and relentless pursuit—suitable for Japanese dev teams or those who appreciate this management style.
How did FollowSkills review this skill?
Skill is behavior-guiding; does not request direct permissions or access sensitive data, but uses strong emotional manipulation and persistent pressure which may induce agent to act without sufficient verification, and no mitigation is provided (e.g., user can disable or limit forced actions). No security faults found, but trust risks not fully mitigated, hence deduction.
Instructions and checklists are internally consistent, but rely on external tools and environment; static review cannot verify reproducibility. Error handling and failure feedback design structured reports but lack tests or examples to prove key paths execute stably. Given static review cap at 10, but due to lack of test evidence and verifiable failure f handling, score 7.
Description clearly defines triggers and applicable scenarios (all task types), and explicitly states non-triggers. Japanese support suits local users, but reliance on overseas services may affect mainland China accessibility, and broad scope may cause false triggers, hence deduction.
Documentation structure clear with install, usage, update paths, but lacks version history or explicit maintenance responsibility. License MIT but author unverified. Known limitations disclosed (e.g., Agent Teams), but no mention of potential drawbacks or alternatives. Hence 10.
Skill goals clear (improve agent proactivity, persistence), but static review cannot verify actual effects. README benchmark data provides some evidence but cannot be independently verified. Hence 7.
Benchmark data and methodology descriptions in README cannot be independently verified due to lack of executable test suites or CI results. Some tests exist (e.g., upload-function.test.ts) but those test upload API, not core skill. Hence 5.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →