> "The next employee you want to distill — why must it be a colleague?"
Background: From colleague-skill to Nuwa.skill
In late March 2026, developer titanwings released colleague-skill on GitHub: feed it a departed coworker's chat logs, work documents, and emails, and the AI distills a digital stand-in that can write code, reply to messages, and mimic the person's tone — including their blame-shifting habits. It earned 7,000+ stars in five days, spawning derivatives like ex.skill, boss.skill, and mentor.skill.
colleague-skill uses a dual-track architecture:
- Part A — Work Skill: extract hard skills (systems owned, code style, code review habits, workflows, lessons from past mistakes)
- Part B — Persona: a five-layer personality structure (hard rules, identity, expression style, decision patterns, interpersonal behavior)
- Why distill, not just read? Public corpora run to hundreds of thousands of words. Three costs motivate the tool: time to read, effort to filter signal from noise, and difficulty turning scattered knowledge into an operable cognitive framework. Nuwa.skill converts "read it, vaguely remember it" into "input a question, get analysis in that person's perspective."
- Five-layer distillation model (outer to inner): 1. Expression DNA — tone, rhythm, word choice; the outer manifestation of how someone organizes thought. 2. Mental models — the core layer: 3–7 cognitive frameworks, e.g. Jobs's "focus means saying no" and end-to-end control; Munger's inversion and latticework of models; Musk's first-principles and physical-limit estimation; Naval's "desire is a contract" and leverage; Feynman's "naming is not understanding." 3. Decision heuristics — 5–10 fast rules of thumb (tactics to the mental models' strategy). 4. Anti-patterns and values — explicit "never do" lists, which often define a person better than what they do. 5. Honest boundaries — every skill states what it *cannot* do: it cannot distill intuition, capture belief changes after the research snapshot, pierce performative public expression, or replace the real person.
- Six-dimensional parallel research: six agents gather (1) writings for recurring core arguments (a claim in 3+ sources is likely a true belief), (2) interviews/podcasts for unrehearsed reasoning under questioning, (3) social media for expression DNA, (4) critics' external views for blind spots and contradictions, (5) major decisions — "watch what they did, not what they said," and (6) a full timeline, since people change.
- Triple validation: a candidate belief becomes a core mental model only if it (a) recurs across at least two domains, (b) is generative — it can predict stances on new questions, and (c) is exclusive — not something every smart person would say. Partial matches are demoted to heuristics. Naval's thousands of tweets thus compress into 5 mental models and 8 heuristics.
- Quality check — the 3+1 test: the skill must match the person's answers on 3 questions they publicly answered, and show appropriate uncertainty (not fabricate) on 1 question they never addressed.
- Intuition cannot be distilled: the felt judgment after the framework runs is not in any text.
- No capture of change: the skill is a snapshot as of research time, not a live sync.
- Public expression ≠ private thought: the tool cannot see behind strategic or performative statements.
- Ethics: does building skills from the deceased "consume" them? Huashu argues that distilling voluntarily published ideas is continuation, not consumption — while acknowledging the question is open. If a skill convincingly imitates a lost loved one, healthy mourning can shade into avoidance.
Huashu saw this and asked a bolder question: if we can distill coworkers, why not distill Jobs, Musk, Munger, or Feynman? These figures leave behind vastly richer public material, and what makes them great is not their habits but their thinking styles. That is the premise of Nuwa.skill (GitHub: https://github.com/alchaincyf/nuwa-skill).
Key points
Not roleplay — a cognitive operating system
The key distinction: roleplay *quotes* (borrowed phrases, the model's own reasoning underneath); Nuwa.skill *runs* the person's frameworks on new problems.
Example: asked "OpenAI vs Anthropic — who has the right direction?", the Jobs-perspective skill reframes it as a contest of taste and focus, likens OpenAI to 1985-era Microsoft (doing everything, winning share, never insanely great), and stresses that whoever controls chips, models, and interface together wins. Asked "what is blockchain?", the Feynman skill explains it via trusted-less strangers keeping shared ledgers, where tampering costs make fraud mathematically impractical — "replacing trust with math" — instead of reciting a definition.
Usage in Claude Code is conversational: "analyze this from Naval's perspective," "switch to Munger, I'm stuck on an investment decision."
Limitations and ethics
Conclusion: a mirror, not a stand-in
Nuwa the goddess made humans from clay; Nuwa.skill makes mirrors from public information. The output is not a digital double that lives or decides for you, but a cognitive mirror — borrow someone's glasses, look at your problem, then hand them back. Possible next steps include personal skills distilled from one's own digital footprint, incremental live updates, multi-perspective panels (Jobs + Munger + Musk debating one question), and domain-specific methodology skills unbound from any individual.
References
1. colleague-skill — titanwings. GitHub: https://github.com/titanwings/colleague-skill 2. nuwa-skill — Huashu. GitHub: https://github.com/alchaincyf/nuwa-skill 3. *Poor Charlie's Almanack* — ed. Peter Kaufman 4. *Steve Jobs* — Walter Isaacson, 2011 5. *Surely You're Joking, Mr. Feynman!* — Richard Feynman, 1985 6. *The Almanack of Naval Ravikant* — Eric Jorgenson, 2020 7. paulgraham.com essays — Paul Graham, 2001–2025 8. *Elon Musk* — Walter Isaacson, 2023 9. Ilya Sutskever interview collection — Lex Fridman, Dwarkesh Patel podcasts