huashu-nuwa Skill for Claude: Build a Skill That Thinks Like Someone Else

Researches a person or topic, then builds a new skill that reasons the way they do, with a warning that a full run is long and can be costly.

Installs: 27.1K; on skills.sh
Stars: 37.1K; on GitHub
Updated: Oct 2026; last commit

huashu-nuwa builds a new skill from how a person or topic thinks. You name someone, such as Munger or Feynman, or describe a need, and it researches them and writes a "perspective" skill into .claude/skills/<name>-perspective/. If you only describe a need, it first suggests 2 or 3 people or topics and the limits of each.

A full run is a long task. SKILL.md says it starts 6 research agents in parallel, can pass 500k tokens in one run, and that a top model can spend tens of dollars, so it asks you to pick a depth before it starts. It searches the web and can run a YouTube subtitle script. Say "distill [name]", "nuwa" or "create a [person] perspective skill" to trigger it.

Good for

People who want a thinking advisor modeled on a founder, investor or scientist
Skill authors who want to turn someone's writing and interviews into a reusable skill
Anyone with a vague need such as better decisions who wants help choosing whose thinking to copy

Install huashu-nuwa

Claude Code

Copy this prompt.

Install the huashu-nuwa skill in Claude Code. Run this command:

npx skills add https://github.com/alchaincyf/nuwa-skill --skill huashu-nuwa

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js.

From its listing on skills.sh. Source.

What's in the skill

When it loads

女娲造人：输入人名/主题/甚至只是模糊需求，自动深度调研→思维框架提炼→生成可运行的人物Skill。两种入口：(1)明确人名→直接蒸馏 (2)模糊需求→诊断推荐→再蒸馏。触发词：「造skill」「蒸馏XX」「女娲」「造人」「XX的思维方式」「做个XX视角」「更新XX的skill」。模糊需求也触发：「我想提升决策质量」「有没有一种思维方式能帮我...」「我需要一个思维顾问」。 English triggers: "distill [person]", "nuwa", "create a [person] perspective skill", "how does [person] think", "I need a thinking advisor".

Sections in SKILL.md

核心理念
执行流程
Phase 0: 入口分流
Phase 0A: 需求澄清（直接路径）
Phase 0B: 需求诊断（模糊路径）
候选1: [人名/主题] ⚡已有Skill / 🆕需要蒸馏
Phase 0.5: 创建Skill目录
Phase 1: 多源信息采集（并行Agent Swarm）
Phase 1.5: 调研Review检查点
Phase 2: 框架提炼（Synthesis）
Phase 2.5: 提炼确认检查点
Phase 3: Skill构建
回答工作流（Agentic Protocol）
Step 1: 问题分类
Step 2: [人物名]式研究（按问题类型选择）
Step 3: [人物名]式回答

Key facts

Author: 花叔 (Huashu)
License: MIT
Works in: Claude Code, Codex, Cursor, OpenClaw, Hermes Agent, CodeBuddy, Workbuddy, Gemini CLI, OpenCode
The skill folder: SKILL.md sits at the repo root beside references/ (4 files) and scripts/ (4 files: 1 shell, 3 Python); the repo has 162 files, 124 of them under examples/
Cost warning: SKILL.md: a full distillation with a top model can cost tens of dollars, so pick a depth first
Installs on skills.sh: 27,106 (Nov 2, 2026)
GitHub stars: 37,067 (Nov 2, 2026)
Last commit: 2026-10-24
skills.sh audits: Gen Agent Trust Hub fail, Socket warn, Snyk pass

What it includes

You name a person or topic and it starts research
It checks who or what you mean, your focus, the purpose, whether a skill already exists and whether you have local material. Then it picks a depth and starts.
Try: “Distill Paul Graham”
Network · Direct path
You describe a need and it suggests 2 or 3 people or topics
After at most 2 follow-up questions it recommends candidates, shows which already exist as local skills, and states each one's blind spots.
Try: “I need a thinking advisor for business decisions”
Diagnosis path
Quick (3 areas), standard (6) or deep (6 plus full downloads)
Quick is about a third of the standard cost. Deep also downloads books, subtitles and long articles for archive.
Depth choice
Makes .claude/skills/<name>-perspective/ with research and source folders
Research results must be saved inside that folder so it can be copied on its own.
Write · Create skill folder
Writings, conversations, expression, outside views, decisions, timeline
Each agent saves a numbered markdown file under references/research/ and marks sources as first-hand or second-hand. It runs agents one by one if your client cannot run them in parallel.
Write · Six research agents
Downloads YouTube subtitles with yt-dlp, Chinese then English
Tries manual subtitles first, then auto-generated ones, and writes SRT or VTT files to a folder you give.
Runs code · scripts/download_subtitles.sh
Clean subtitles, summarise research, score the finished skill
Three Python scripts that import only sys, re and pathlib. They turn subtitles into text, count sources, and run 6 pass or fail checks.
Runs code · srt_to_transcript.py, merge_research.py, quality_check.py
Pauses to show source counts and conflicts before it goes on
You confirm the research is good enough, or ask for more on a weak area.
Research checkpoint
Keeps 3 to 7 mental models that pass a three-part test
A model must show up in at least 2 fields, predict views on new problems and not be something every smart person says. Weaker items become rules of thumb.
Framework extraction
Writes the new SKILL.md, then tests it with separate sub-agents
Three tests: known views, an edge case it should hedge on, and a voice check. Up to 2 rounds of fixes, then it ships with the weak spots listed.
Write · Build and test
Two review agents polish it; later runs refresh a skill with new material
An update re-checks recent conversations, decisions and the timeline only, and does not rewrite the whole skill.
Refinement and updates

From the official docs

Permissions

Medium risk

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

A full run can cost tens of dollars in tokens
SKILL.md says a top model can spend tens of dollars in one distillation and a run can pass 500k tokens. Pick the quick depth first and stop between phases. github.com
It runs yt-dlp and Python scripts on your machine
download_subtitles.sh needs yt-dlp and fetches from YouTube. Read the scripts first and run them only on URLs you chose. github.com
It lists Z-Library and LibGen for book downloads
SKILL.md lists "Z-Library/LibGen" search and download of books as an optional aid, saved to sources/books/. Do not use it to copy books you have no right to copy. Give it books you own or have a licence for instead. github.com
Be careful with private people
For living non-public people SKILL.md says the user must supply the material and be reminded to get the person's consent. Skip the person if you cannot. github.com

Technical details

Talks to	Web search and page reads during research. YouTube via yt-dlp when you use the subtitle script. Book search and download from Z-Library and LibGen, listed as an optional aid in SKILL.md. Other installed skills it may call, such as gemini-video, web-article-reader, agent-reach, huashu-research and pdf
Sign-in and keys	No sign-in or key is described in SKILL.md.
Scopes	Creates .claude/skills/<name>-perspective/ with research files and a SKILL.md Reads local books, transcripts and notes you hand over Runs scripts/download_subtitles.sh, which calls yt-dlp, and three Python scripts

Questions

What does huashu-nuwa do?

It builds a new skill that mimics how a person or topic thinks. It researches their writing, talks, decisions and timeline, extracts mental models and speaking style, and writes a skill you can then use as a thinking advisor.

How do I install huashu-nuwa?

Use the install box above. The README also says you can tell a skills-compatible agent "Install this skill for me" with the repo link, or run the skills CLI.

How do I use it?

Say "distill [name]", "nuwa" or "create a [person] perspective skill". If you do not know whom to pick, describe the problem, such as slow decisions, and it recommends candidates first.

How much does a run cost?

It depends on depth and model. SKILL.md says a full distillation with a top model can cost tens of dollars, and it asks you to choose quick, standard or deep before it starts. A lighter model lowers the cost.

Can I use my own material?

Yes, and SKILL.md ranks it highest. Hand over book PDFs, transcripts or notes and it reads those first, then searches only for what is missing.

Which sources does it avoid?

It never uses Zhihu, WeChat public accounts or Baidu Baike, and for Chinese sources it accepts named outlets such as 36Kr and LatePost, plus original podcast and video platforms.

Is it safe?

It needs care. It runs web research, a yt-dlp subtitle script and three small Python scripts, and it writes files under .claude/skills/. The biggest risk is spend, so pick the quick depth first. SKILL.md also lists Z-Library and LibGen as book sources; decline that step and give it books you own or have a licence to copy. For a living private person, use only material they have agreed to share.

Does it work with people who are not public figures?

SKILL.md says to ask you for the material, and to remind you to get the person's consent first. Without material it warns that quality will be limited.

What if my client cannot run agents in parallel?

SKILL.md says to run the 6 research tasks one at a time and save each file before the next. It also gives a way to split the run across sessions for smaller context windows.

Information

updated Nov 2, 2026

Repository: alchaincyf/nuwa-skill

alchaincyf/nuwa-skill skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 3

github.com · github.com
huashu-nuwa on skills.sh · skills.sh
github.com · github.com