nuwa-skill: Distilling a Person's Thinking into an Agent Skill

The problem nuwa-skill targets: a perspective you can call by name

Generic model output tends toward the average. Ask an unmodified assistant how to cut SaaS acquisition cost and you get a list of funnel optimizations. The README's example has the Musk skill refuse that framing and ask what the theoretical minimum path from awareness to payment is, then treat anything more than three times that minimum as removable steps. The difference is not information. It is the order in which the problem gets attacked.

nuwa-skill exists to make that order installable. The README states the goal plainly: input a name, and the project runs research, extraction and validation automatically, producing a skill that reasons through that person's cognitive framework rather than reciting their quotes. The pitch line is 「你想蒸馏的下一个员工，何必是同事」, and it is aimed at people who want Steve Jobs, Munger, Feynman or Musk as an analytical lens inside their agent. The intended user is someone already working in an agent runtime who wants a switchable viewpoint, not someone looking for a biography or a quote database.

Five layers of extraction, and the honest-boundaries section

The project describes five layers it pulls from source material: how the person speaks (expression DNA: tone, rhythm, word preferences), how they think (mental models and cognitive frames), how they judge (decision heuristics), what they refuse to do (anti-patterns and value floors), and where they stop (an explicit honesty boundary). The README's own framing is that work habits transfer through process documents, but what makes Munger and Musk reach different conclusions on the same question is the cognitive frame. So the extraction target is the cognitive operating system, not the biography.

Each generated skill carries a 诚实边界 section listing what it cannot do. The README names three: intuition cannot be distilled (frames can be extracted, inspiration cannot), abrupt change cannot be captured (the skill is a snapshot as of the research date), and public expression is not the same as private belief. The line the project uses to justify this is 「一个不告诉你局限在哪的Skill，不值得信任」. That is a design commitment, and it is the first thing to check in any skill you install, because it tells you which questions the skill will answer badly.

Installing it: one command, or a clone into a runtime-specific path

The recommended route assumes you already have an agent open. You tell it to install the skill from the repository URL, or you use the generic CLI installer from vercel-labs/skills, which the README says supports 55+ runtimes:

npx skills add alchaincyf/nuwa-skill

The installer detects the current runtime and places the skill in the right directory. To force a target, the README shows flags such as -a claude-code, -a codex, -a cursor and -a openclaw.

Manual installation means cloning into a runtime-specific path. The README's table gives these: ~/.claude/skills/nuwa-skill/ for Claude Code, ~/.codex/skills/nuwa-skill/ for Codex CLI, ~/.cursor/skills/nuwa-skill/ for Cursor, and ~/.openclaw/workspace/skills/nuwa-skill/ for OpenClaw. Hermes Agent is different from the rest: you run tools/install_hermes_skill.py rather than cloning by hand. For any other runtime, clone into that runtime's skills/ directory.

The third route matters if your runtime has no skill loader at all. The README points out that SKILL.md is markdown with YAML frontmatter, so you can paste its contents into a conversation directly. That is a useful fallback, though it also means you lose whatever the runtime does with skill metadata.

Invoking a distilled skill, and what the examples actually show

Once installed, you create a skill by naming a person in natural language: distill a Paul Graham, build a Zhang Xiaolong perspective skill, make a Duan Yongping skill. Then you invoke it the same way: analyze this investment decision from Munger's perspective, or ask how Feynman would explain quantum computing, or switch to Naval while you are stuck between three options.

The README's sample exchanges are the clearest statement of intent. In the Naval example, a user says they cannot split energy across media, a book and indie development. The reply reframes three desires as three contracts signed with unhappiness, and prescribes serial compounding rather than parallel drain. In the Jobs example, asked whether OpenAI or Anthropic has the right direction, the reply rejects the framing as a taste contest rather than a direction contest, contrasts OpenAI with 1985 Microsoft and Anthropic with early Apple, and lands on hardware control as the thing neither does. The README is explicit that this is not role-play: the Jobs output uses the focus-means-saying-no and end-to-end-control mental models, the Naval output uses desires-as-contracts, the Musk output uses asymptotic-limit reasoning. Whether that distinction holds up is exactly what a blind test is meant to measure, and the project says it ran one.

The fidelity scorecard and its blind-test methodology

The README states that all 15 official skills passed an independent two-agent blind test across five dimensions: stance consistency, style recognizability, edge-case honesty, source transparency and structural completeness. Methodology is documented in references/fidelity-scorecard.md, and each skill directory carries its own FIDELITY.md. The reported scores range from 89 (Musk) to 97 (MrBeast, Naval, Taleb, Jobs, Karpathy, Paul Graham, Zhang Xuefeng), with Munger, Feynman and X Mentor at 96, Trump at 95, Ilya at 94, Zhang Yiming at 93 and Sun Yuchen at 91. The README says all of them reached grade A at 85 or above.

Treat those numbers as the project's own reported results, not as an external benchmark. The README does not say who the blind-test agents were, how many rounds ran, or whether the graders were independent of the author. The five dimensions are a reasonable rubric, and publishing per-skill FIDELITY.md files is more than most persona projects do, but a self-administered blind test is still a self-administered test. The scorecard is useful for comparing skills within this family. It is not evidence that a nuwa skill will outperform a different approach on your task.

Where it fails: snapshot drift, public-only sources, and no intuition

The stated limitations are the real ones. A distilled skill is a snapshot as of its research date, so anything the person said or did afterward is invisible, and a skill built on a pre-2026 corpus will answer 2026 questions with pre-2026 frames. Sources are public, so the skill models the public persona, which is a different object from the person. And the project concedes outright that intuition cannot be extracted, which means the output is strongest on questions that fit an articulable framework and weakest on questions where the real answer was a gut call.

There is a second failure mode the README does not dwell on. A skill that is confident by construction will sound equally confident when the frame does not apply. The 诚实边界 section is the counterweight, but it only helps if you read it and if the runtime surfaces it. If you are using a distilled skill to produce claims about a living person, a company or a market, the skill's fluency is a liability: it will produce a Jobs-flavored judgment about hardware strategy with the same cadence it uses for a product critique, and nothing in the output marks which parts are grounded and which are frame extrapolation. Use it to stress-test your own reasoning. Do not use it as a source.

How this differs from the colleague-skill approach it cites

The README positions nuwa-skill against titanwings/colleague-skill, which it credits with proving that distilling a person is feasible. The difference is target selection. colleague-skill distills a coworker, which means the source material is internal and current: meeting notes, ticket comments, code review habits, the way that person actually decides under your team's constraints. The output is narrow and perishable, tied to a role and an organization.

nuwa-skill goes the other direction. It distills public figures, which makes the source material abundant, durable and external, but strips out everything private. You get Munger's latticework of mental models from decades of public writing; you do not get how Munger would read your specific cap table. The trade is breadth and reusability against grounding. A colleague skill can be wrong about your codebase and right about the person. A nuwa skill can be right about the person and have no access to your context at all. Which one you want depends on whether the bottleneck is a missing frame or missing facts.

Licence, maintenance surface, and what to verify before adopting

The repository is MIT licensed, which permits commercial use, modification and redistribution provided the copyright notice and licence text are retained. That covers the nuwa-skill repository itself. It does not settle the status of the distilled output: a skill built from a living person's public interviews, books and posts carries its own questions about quotation, attribution and personality rights, and the README's source-transparency dimension suggests the project treats provenance as part of the artifact. That is a question for your own counsel, not something the MIT badge answers.

The maintenance surface is larger than a single repo. The README lists 14 person skills plus one topic skill, each in its own repository with its own npx skills add command, plus one example shipped inside examples/sun-yuchen-perspective/. Upgrading means tracking those repositories individually, and because every skill is a dated snapshot, the useful maintenance work is re-distillation rather than version bumps. There are no releases retrieved for the main repository, so pinning to a tag is not currently an option; you are tracking main.

Before adopting, read references/fidelity-scorecard.md to see how the scores were produced, read one skill's FIDELITY.md alongside its 诚实边界 section, and confirm your runtime appears in the installation table or can at least accept a pasted SKILL.md. If your runtime is not listed and you cannot paste the file, the project has nothing to offer you yet.

Editorial conclusion

Adopt nuwa-skill if you already run an Agent Skills-compatible runtime and want a named perspective you can invoke mid-conversation, and if you accept that the output is a snapshot of public material rather than the person. Skip it if you need verifiable factual claims about a living person, or if your runtime cannot load SKILL.md at all. Before installing, open references/fidelity-scorecard.md and one FIDELITY.md file, then read the 诚实边界 section of the skill you plan to use: that section is where the project states what the distillation cannot do, and it is the part most worth checking against your use case.

Official sources

1. alchaincyf/nuwa-skill on GitHub
2. Issues
3. License: MIT
4. README