nuwa-skill: Distill Any Thinker's Mental Models into an Agent Skill

What nuwa-skill Does and Who It Is For

nuwa-skill is an open-source skill built on the Agent Skills protocol. Its function is to distill how a specific person thinks: their mental models, decision heuristics, expression style, anti-patterns, and acknowledged limitations, into a portable SKILL.md file that an AI agent runtime can load and apply.

The project is aimed at users of AI coding agents such as Claude Code, Cursor, Codex CLI, OpenClaw, Hermes, CodeBuddy, Workbuddy, Gemini CLI, and OpenCode who want to consult a specific cognitive framework during their work. It is not a roleplay or persona tool. The README distinguishes between reproducing famous quotes and applying a cognitive framework: the intention is that the distilled skill uses the person's reasoning patterns to analyze a given situation, not to repeat things they have said.

The Five Layers nuwa-skill Extracts

The README describes five layers of extraction that nuwa-skill targets, each going deeper than observable work habits:

Expression DNA covers tone, rhythm, and word preferences: how the person actually phrases things, what vocabulary they reach for first, whether they speak in short declarative sentences or layered qualifications. Mental models and cognitive frameworks describe how the person structures a problem before solving it. Decision heuristics describe the shortcuts and principles that drive choices under uncertainty. Anti-patterns and value limits document what the person actively refuses to do and where they draw lines in their public statements. Honest limitations state what the distilled skill cannot replicate or what it explicitly does not know.

The README frames this by noting that while work habits can be transmitted through process documents, what makes a Charlie Munger response different from an Elon Musk response to the same question is cognitive frameworks. nuwa-skill is attempting to extract that operating layer rather than a surface-level imitation. The example dialogues in the README show this concretely: the distilled Musk applies first-principles decomposition to an acquisition cost problem by working back to the theoretical minimum number of steps, while the distilled Naval reframes a multi-goal conflict as a contract problem before any solution is proposed.

Installing nuwa-skill and Creating a Distilled Skill

The recommended install method uses the vercel-labs/skills CLI, which detects the current runtime and places the skill in the correct directory automatically:

npx skills add alchaincyf/nuwa-skill

For manual installation, clone the repository directly to the runtime-specific skills directory:

git clone https://github.com/alchaincyf/nuwa-skill ~/.claude/skills/nuwa-skill/

The install paths for other runtimes are `~/.codex/skills/nuwa-skill/` for Codex CLI and `~/.cursor/skills/nuwa-skill/` for Cursor. For runtimes not yet supporting automatic loading, pasting the contents of SKILL.md directly into the conversation also works, since the skill is a markdown file with YAML frontmatter.

After installation, the agent can be instructed to distill a specific person:

> 蒸馏一个保罗·格雷厄姆

Once a skill is created, it can be invoked directly:

> 用芒格的视角帮我分析这个投资决策

The 14 Pre-Distilled Figures and Fidelity Scoring

The repository ships 14 pre-distilled person skills and one theme skill, each as its own independently installable repository. The people range across investment (Munger, Naval, Taleb), product and design (Jobs, Zhang Yiming), AI and engineering (Karpathy, Ilya Sutskever), content creation (MrBeast), education (Feynman), and public communication (Trump). The theme skill is X-mentor, covering Twitter/X growth methodology.

All 15 official skills have been evaluated through a dual-agent blind test using five criteria: stance consistency, style recognizability, edge-case honesty, source transparency, and structural completeness. The scorecard methodology lives in references/fidelity-scorecard.md. All 15 scored A-grade (85 or above). Individual scores documented in the README range from 97 for MrBeast, Naval, Taleb, Jobs, Karpathy, Paul Graham, and Zhang Xuefeng, down to 89 for Musk.

Community-contributed skills are indexed in COMMUNITY.md. To be listed, a skill must pass the same fidelity scorecard at B-grade or above and be submitted as a one-line pull request pointing to the contributor's own repository.

Limitations the README Documents Explicitly

The README states three specific things that nuwa-skill cannot do. First, it cannot distill intuition: frameworks can be extracted from public statements and work, but inspiration and intuition that leave no written trace are outside the capture boundary. Second, each distilled skill is a snapshot: it reflects publicly available information up to the research date and does not update when the person's views change. Third, public expression is not the same as private belief: the skill is built only from what the person has said publicly, which may not represent their actual reasoning in every context.

The README phrases this as a self-aware limit: a skill that does not tell you its boundaries is not worth trusting. Each pre-distilled skill includes a section documenting what it cannot replicate. This is the honest-boundaries layer listed as the fifth extraction target.

How nuwa-skill Differs from colleague-skill

The README directly references colleague-skill (titanwings/colleague-skill) as the prior work that demonstrated distilling a single person's thinking is feasible. The stated motivation for nuwa-skill is to apply that same approach to publicly known figures rather than coworkers.

The practical difference is significant. colleague-skill is designed for a workplace context: the input is someone you work with, whose behavior you observe directly. nuwa-skill's inputs are public writing, interviews, speeches, and published work. colleague-skill produces a skill shaped by private interactions; nuwa-skill produces one bounded by whatever the named person has made public.

For teams who want to codify the reasoning style of an internal colleague rather than a public figure, colleague-skill is the more appropriate tool. nuwa-skill is the better fit when the goal is to consult a well-known cognitive framework during work, such as applying Feynman's explanation method to a technical writeup or applying Munger's inversion principle to a product decision.

Runtime Compatibility and MIT License

nuwa-skill is built on the Agent Skills protocol at agentskills.io and requires a runtime that supports that protocol. The README lists over 50 compatible runtimes, including Claude Code, Codex, Cursor, OpenClaw, Hermes Agent, CodeBuddy, Workbuddy, Gemini CLI, and OpenCode. The vercel-labs/skills CLI handles detection and placement automatically for any supported runtime with the `-a <runtime>` flag when the target needs to be specified.

The repository mentions a companion project called darwin-skill that is designed to make skills evolve over time. Where nuwa-skill handles the initial distillation, darwin-skill is intended to update and refine existing skills as new public information becomes available. The README treats the two as complementary: nuwa creates, darwin evolves.

The last push was on 2026-08-25. The repository is not archived. It is released under the MIT license, which permits use, modification, and redistribution without restriction beyond attribution. The SKILL.md file, which is the core methodology document, is explicitly not open for external pull requests; the README asks that methodology improvement suggestions go through issues instead.

Editorial conclusion

nuwa-skill is a practical tool for anyone who already uses an Agent Skills-compatible runtime and wants to apply the cognitive frameworks of known thinkers to decision-making, writing, or analysis. The key constraint is that each distilled skill is a snapshot of publicly available information up to the research date, and the README is explicit that intuition, private beliefs, and real-time changes are outside what any skill can capture. Before relying on a newly distilled skill, run it through the fidelity scorecard in references/fidelity-scorecard.md to evaluate consistency and honesty-boundary behavior before treating its outputs as representative.

Frequently asked questions

Official sources

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