Key points
This article introduces Nuwa.skill (女娲.skill), an open-source project by Chinese developer Huashu (花叔), built on the momentum of the viral GitHub project colleague.skill.
Origin: from colleague.skill to Nuwa.skill
- In late March 2026, developer titanwings released colleague.skill, which feeds a departed coworker's chat logs, documents, and emails to AI to "distill" a digital stand-in. It gained 7,000+ GitHub stars in five days.
- colleague.skill uses a "dual-track" architecture: Part A (Work Skill) for hard professional capabilities, and Part B (Persona) — a five-layer personality structure covering hard rules, identity, expression style, decision patterns, and interpersonal behavior.
- Huashu's insight: instead of distilling coworkers' *behaviors*, why not distill the *thinking styles* of the greatest minds, whose public materials (books, talks, interviews, tweets) are vastly richer?
- Jobs: Focus means saying no; end-to-end control; the "insanely great" standard
- Munger: Invert, always invert; latticework of mental models
- Musk: Compute physical limits; first principles; question every requirement
- Naval: Desire is a contract; specific knowledge + leverage; compounding
- Feynman: Naming is not understanding; build to understand
- Cross-domain recurrence: appears in at least two different domains
- Generative power: can predict the person's stance on new questions
- Exclusivity: not something every smart person would say
- Simple role-play *quotes* a person's famous lines; Nuwa.skill *runs* their cognitive frameworks on new problems.
- Example: asked "OpenAI vs Anthropic — who is right?", a Jobs-style skill reframes it as a contest of *taste*, invokes focus-as-saying-no and end-to-end control, and concludes whoever controls chips + models + interface wins.
- Use cases shown: Naval for life choices, Musk for business cost analysis, Munger for investment decisions, Feynman for testing true understanding, MrBeast for content optimization.
- Distill intuition — the felt judgment after frameworks run stays unwritten ("what can't be written in is your real moat")
- Capture sudden changes — it's a snapshot as of research time, not a live clone
- Pierce performative expression — public statements may differ from real decision-making thoughts
- Replace the person — it is a mirror, not a stand-in
- colleague-skill: https://github.com/titanwings/colleague-skill
- nuwa-skill: https://github.com/alchaincyf/nuwa-skill
- *Poor Charlie's Almanack*; Isaacson's *Steve Jobs* (2011) and *Elon Musk* (2023); *Surely You're Joking, Mr. Feynman!* (1985); *The Almanack of Naval Ravikant* (2020); paulgraham.com essays
The five-layer distillation model
| Layer | Meaning | |---|---| | How they speak | Expression DNA — tone, rhythm, diction | | How they think | Mental models — the cognitive "glasses" they wear | | How they decide | Decision heuristics — quick judgment rules | | What they won't do | Anti-patterns and value bottom lines | | Knowing limits | Honest boundaries — what the skill cannot do |
Example distilled mental models:
Methodology: six-dimensional parallel research
Six agents research each person simultaneously:
1. Writings — core arguments recurring across 3+ sources 2. Conversations — impromptu responses under questioning (e.g., Ilya Sutskever on scaling laws) 3. Expression — high-frequency words and style on social media 4. External views — critics and biographers, revealing blind spots 5. Decisions — actual behavior at major inflection points 6. Timeline — capturing how the person changed over time
Results pass triple verification before becoming core mental models:
Naval's thousands of tweets and dozens of podcast hours compress into just 5 mental models and 8 decision heuristics. Final quality gate: a "3+1 test" — 3 questions the person publicly answered (check directional consistency) plus 1 question they never addressed (check appropriate uncertainty rather than confident fabrication).
Not role-play, but a cognitive operating system
Honest limits and ethics
The skill explicitly documents what it cannot do:
Ethical questions raised: consuming the deceased vs. extending their ideas; healthy cognitive tools vs. unhealthy digital surrogates of lost loved ones.
Conclusion
Nuwa.skill produces not a digital twin but a cognitive mirror: you borrow another thinker's glasses to examine your own problems, then hand them back. Possible future directions include personal skills, incremental updates, multi-perspective panels (Jobs + Munger + Musk on one problem), and domain-specific methodology skills not tied to any person.