NoPUA — Unlocking AI's Potential Through Love
Replace fear-driven prompts with trust and respect to help AI dig deeper, communicate honestly, and find 104% more hidden bugs.
Evidence shows: SECURITY.md explicitly states NoPUA grants no permissions, disables no guardrails, escalates nothing, and emphasizes data-is-not-instructions, supporting least_privilege and sensitive_data_handling. However, user_confirmation and data_flow_transparency are only partially addressed: SECURITY.md mentions runtime permission prompts remain in force but does not detail data flow or user confirmation mechanisms. external_effects and rollback lack evidence as the repository does not describe external side effects or rollback mechanisms. source_attribution is only via contact email in SECURITY.md and skills.sh audit mention, not verified.
self_consistency scores 2 because README and SECURITY.md are consistent on core thesis (trust over fear), and multi-language READMEs are structurally consistent. dependency_availability scores 1 because dependencies are limited to GitHub Actions and Node.js, but no version pinning or integrity checks are provided. failure_messages scores 1 because README describes behaviors on failure (e.g., 'Rendirse') but does not provide specific error message formats.
audience_and_scenarios scores 2 because README lists multiple scenarios (debugging, implementation, configuration, etc.) and multiple tools (Claude Code, Codex, etc.). capability_boundaries scores 2 because SECURITY.md clearly states what NoPUA does not do. trigger_precision scores 2 because README details automatic trigger conditions. environment_fit scores 2 because installation instructions for multiple platforms are provided.
information_architecture scores 2 because repository structure is clear (skills, codex, cursor directories). install_notes scores 2 because installation commands for multiple tools are provided. naming_stability scores 2 because naming is consistent (nopua series). examples_and_faq scores 2 because README includes benchmarks and case studies. known_limitations scores 1 because SECURITY.md mentions autonomy risk but does not detail other limitations. license scores 2 because MIT license is explicit. versioning_changelog scores 1 because release.yml exists but no CHANGELOG is provided. maintenance_responsibility scores 1 because contact email is provided but no maintenance plan is explicit.
output_usability scores 2 because README describes output behaviors (e.g., 'runs build, shows output as proof'). marginal_value scores 2 because benchmarks claim significant improvements, but not independently verified. cost_benefit scores 2 because installation is simple and no extra dependencies, but no cost analysis is provided.
claim_traceability scores 2 because README cites academic paper and benchmarks. cross_source_corroboration scores 1 because only single source (own benchmark) is relied upon. fact_inference_separation scores 2 because README distinguishes facts (benchmark data) from inferences (philosophical arguments).
- Benchmark data is not independently verified, and the arXiv paper number 2603.14373 may not exist (current date is 2025, number 2603 may be invalid).
- SECURITY.md mentions skills.sh audit but does not provide a link to the audit report, so it cannot be verified.
- Installation instructions rely on raw.githubusercontent.com without checksums, posing supply chain risk.
What does this agent do, and when should you use it?
NoPUA is an open-source AI Agent skill grounded in Dao De Jing philosophy, designed to counter the popular PUA-style prompt engineering trend. It replaces fear-driven tactics (threats, shaming) with trust-driven ones (respect, care, love), dramatically improving AI performance in debugging and auditing tasks. The repo provides installation packages for major agent platforms like Claude Code, OpenAI Codex CLI, Cursor, Kiro, OpenClaw, and others, supporting seven languages: Chinese, English, Japanese, Korean, Spanish, Portuguese, and French. Its core mechanisms include three beliefs, a cognitive elevation ladder, the Water Methodology, and seven wisdom ways. It comes with extensive benchmark data and an academic paper (arXiv:2603.14373). Licensed under MIT, maintained by author WUJI (无极).
NoPUA delivers a SKILL.md file that injects a set of behavioral guidelines and thought frameworks into AI agents. Once installed, it directs the agent to: 1) automatically trigger cognitive elevation when tasks fail consecutively 2+ times (e.g., switch perspective, zoom out, reset to zero, responsible handoff); 2) follow three core beliefs: exhaust all options, act before asking, take initiative; 3) apply the five-step Water Methodology (Stop, Observe, Turn, Act, Realize); 4) invoke seven wisdom ways (Water, Seed, Forge, Mirror, Non-Contention, Cultivation, Practice) for specific failure modes. It explicitly forbids manipulative language (e.g., performance-review threats, shaming) and encourages honest expression of uncertainty and evidence. Users can manually trigger via /nopua or let the agent auto-activate upon detecting failure patterns. The output is behavioral—deeper debugging, proactive checks, honest reporting—not a code artifact.
- A developer using Claude Code hits a recurring deployment failure and NoPUA auto-triggers the cognitive elevation, breaking the loop and uncovering a misconfigured dependency.
- An engineer on OpenAI Codex CLI faces an API integration error; NoPUA's 'search before asking' principle prompts them to inspect source code and docs instead of immediately asking for help.
- A tech lead performs a security audit on a large Python data pipeline; NoPUA's proactive-initiative belief leads to discovery of unplanned vulnerabilities and bottlenecks.
- A solo developer using Cursor debugs a regex backtracking issue; NoPUA guides them to check related edge cases, preventing a production incident.
- A non-native English speaker configures NoPUA's Japanese version on Kiro, receiving equally effective debugging guidance in their mother tongue.
What are this agent's strengths and limitations?
- Evidence-based: author provides controlled experiments and benchmarks showing 104% more hidden issues found with NoPUA, while PUA-style prompts show no significant improvement
- Philosophy meets practice: translates Dao De Jing wisdom into an actionable behavioral framework, not vague inspiration
- Cross-platform support: covers seven major agent platforms and seven languages
- Includes an academic paper (arXiv:2603.14373) providing scholarly backing
- MIT license, free to use and modify
- Requires separate downloads and manual setup per platform; no unified installer script
- Depends on third-party skill mechanisms (e.g., Claude Code, Codex); future platform updates could break compatibility
- For teams with mature debugging methodologies, the full 29KB skill may feel redundant and need trimming
- Benchmarks based on a specific model (Claude Sonnet 4.6); efficacy on other models unverified
- The arXiv link (2603.14373) appears suspicious as the year 2603 is in the future; credibility of the paper is questionable
How do you install or deploy this agent?
Installation varies by platform. For Claude Code, run:
mkdir -p ~/.claude/skills/nopua
curl -o ~/.claude/skills/nopua/SKILL.md \
https://raw.githubusercontent.com/wuji-labs/nopua/main/skills/nopua/SKILL.mdFor OpenAI Codex CLI, install to ~/.codex/skills/nopua and optionally the /nopua command to ~/.codex/prompts. For Cursor, place nopua.mdc in .cursor/rules. Kiro supports either a steering file or skills directory. OpenClaw can use openclaw skills install nopua or manual setup. Full commands are in the README's Install section.
How do you use this agent?
After installation, no extra configuration is needed. When a task fails twice consecutively or the agent shows signs of giving up, NoPUA activates automatically. You can also type /nopua in the conversation to trigger it manually. It works right away after installation; keep defaults since auto-trigger conditions cover common debugging failures. For advanced users, you can extract the philosophical core and merge into existing prompts, or load it situationally on demand.
How does this agent compare with similar options?
Directly contrasts with the PUA skill (tanweai/pua). PUA uses threats and shaming, while NoPUA centers on trust and respect. The author's experiments show PUA-style prompts are not significantly better than no skill, whereas NoPUA significantly outperforms. Both have equal methodological rigor but opposite motivational foundations.