Nuwa.skill: Distilling How a Person Thinks Into a Runnable AI Skill

Most "celebrity AI" products are in the business of imitation. They learn a tone of voice, memorize quotations, and stay in character. Nuwa.skill takes a different route: you give it a person's name, and what comes back is that person's cognitive operating system. The project is open-sourced by independent developer Huashu, and its core mechanism is straightforward — six agents collect public material in parallel, a triple filter removes anything that does not genuinely belong to the subject, and the output is a conversational SKILL.md. It never claims to bring anyone back. It simply puts your question through someone else's framework and lets you see what falls out.

The problem it actually solves

Reading a biography or watching a long interview usually leaves you with conclusions. What you rarely get is the reasoning that produced them. If you want to borrow someone's judgment, the useful parts are the ones nobody writes down: what they look at first, what they rule out immediately, how they place a bet when the information is incomplete. Those fragments are scattered across books, hours of recorded conversation, and old decision records. Assembling them by hand is expensive, and secondhand interpretation tends to pull the result off target.

Nuwa.skill turns that assembly work into a pipeline. Collection runs in parallel, filtering is explicit and strict, the result is structured, and every source stays on disk so you can retrace the path yourself.

Core capabilities

Six-way parallel collection. Six research agents work at the same time, each pulling a different face of the public record: books and long-form writing, deep interviews, social media activity, critical perspectives, records of key decisions, and the timeline of the person's life. The point of splitting the work six ways is coverage — a single pass tends to over-weight whatever is easiest to find.

Triple-filter distillation. A claim only becomes a formal mental model if it clears three tests at once: cross-domain recurrence (it shows up in more than one area of their work), generative power (it produces new questions rather than just restating a conclusion), and cognitive exclusivity (a generic smart person would not have arrived there). Ideas that appear in exactly one context, or that any thoughtful observer could have produced, get dropped.

Automatic skill generation. The system emits a structured cognitive model, typically 3–7 mental models, 5–10 decision heuristics, an expression DNA, core values and anti-patterns, plus an explicit statement of where the skill's honesty boundary sits.

A quality loop that closes. Known questions test directional consistency. Edge-case questions probe where inference breaks down. Style tests check whether the expression actually matches. Together they keep the generated skill from becoming either overconfident or vacuously general.

Topic-level distillation. Beyond individual people, the same pipeline can synthesize the methodologies of several experts in one field — social platform operations, for instance — into a domain-level cognitive framework.

Transparent, auditable end to end. Every skill keeps its first-hand and second-hand sources, research notes, and extraction logic in a references/ directory. The generation process can be traced and checked rather than taken on faith.

A companion evaluation suite. The project ships with Darwin.skill, an evaluation system that scores skills across 8 dimensions in batches and drives automatic optimization.

How to use it

1. Prepare the environment: install Claude Code, or use a third-party environment that supports the skills.sh protocol.
2. Run npx skills add alchaincyf/nuwa-skill in your terminal to add Nuwa to the skill library.
3. In the conversation, type "distill a XXX" or "build a skill about XXX", replacing XXX with the person or domain you have in mind.
4. Let the automated run finish: six agents collect, distill, verify, and generate, then write out SKILL.md. A single run usually takes about two to three hours, with no supervision required in between.
5. Call the skill directly afterwards — for example, "evaluate this feature from a product person's perspective."
6. Optional: run Darwin.skill for an 8-dimension score, top up the material behind weak dimensions, and distill again.
7. When you need to check a conclusion, open the references/ directory and read the original sources and extraction logic.

Where it fits

• Product decisions and startup judgment. Pull in a product-intuition or first-principles framework to evaluate feature direction, business model viability, and experience design.
• Investing and risk management. Use multi-model thinking to cross-check a target, surface tail risks, and stress-test for antifragility.
• Content creation and distribution. Work from hit structures, analogy-driven explanation, and platform algorithm logic to sharpen topics and scripts.
• Education and career planning. Extract a practitioner's decision framework for school, graduate study, or job choices, then run your own constraints through it.
• Technology trend analysis. Read an industry's direction through the lens of technical evolution, coding paradigms, and compute bottlenecks.
• Negotiation and business strategy. Borrow power analysis and game-theoretic thinking for commercial negotiation, equity splits, and competitive positioning.

Limits and caveats

Distillation quality depends heavily on how much public material exists. For a thinly documented person, the resulting framework will be visibly thinner — sometimes it degrades into common sense in nicer packaging. A skill is also a static snapshot of the moment it was researched; if the person's core thinking shifts later, you have to rerun the pipeline to update it. Two boundaries are worth holding firmly. Do not distill from private email, chat logs, or internal documents — anything non-public requires the subject's explicit consent. And do not treat the output as the person's proxy. What runs is an extracted framework, not the individual's own judgment.

FAQ

How is Nuwa.skill different from role-play AI?

Role-play imitates how someone talks. Nuwa distills how someone judges. It captures how the person thinks rather than what the person said, and it states plainly what it cannot do.

How long does one skill take?

A single distillation run typically takes two to three hours, during which the six agents handle collection, distillation, verification, and generation without anyone watching.

What do I need installed?

Claude Code, or a third-party environment compatible with the skills.sh protocol. Installation is one npx command.

Can I distill someone who is still alive?

Yes, using public material only and never private data. If you share the result publicly, say clearly that it is a simulation of a thinking framework built from public information, not the person's actual views.

More detail is available at the official site: https://nv-wa-skill.lol/zh