zhangxuefeng-skill: Zhang Xuefeng's Decision Framework as a Runnable AI Agent Skill

A Runnable Encoding of Zhang Xuefeng's Reasoning, Not a Quote Collection

Zhang Xuefeng became one of the most-followed voices in Chinese college admissions advice through a method that is blunt, data-driven, and explicit about class. zhangxuefeng-skill does not reproduce a collection of his memorable lines. The README states directly: it is not a quote collection but a runnable thinking framework. The deployable file, SKILL.md, encodes the underlying reasoning patterns so that an AI agent can apply them to new questions rather than recite past answers.

The target audience is anyone navigating China's gaokao system, weighing graduate school against direct employment, or evaluating which professional field to enter given family resources and regional job market realities. The README's demonstration conversations show the skill applied to a 560-point Henan student asking about finance majors, a double-non-985 candidate debating whether to take the entrance exam, and someone asking how AI changes earlier major recommendations. Each response demonstrates a pattern: ask diagnostic questions first, apply a specific framework, deliver a concrete verdict without hedging.

The distinction the README draws is important for evaluating the project honestly. A skill that only mimics Zhang Xuefeng's tone could reproduce his sentence rhythm while generating advice he would not endorse. zhangxuefeng-skill claims to reproduce the analytical structure itself, which makes the distillation process the central engineering question.

Five Mental Models and Eight Decision Heuristics in Detail

The README documents five core mental models extracted from five books and over forty primary sources across the project's six research files.

The 'social sieve theory' (社会筛子论) treats academic credentials, real estate, and job title as successive filters society applies to individuals and families. The 'employment backtracking method' (就业倒推法) starts from where the median graduate ends up, not where the top 10 percent land. The 'class realism' model (阶层现实主义) gives different advice depending on family resources, explicitly separating strategies for students who can absorb risk from those who cannot. The 'choice over effort' principle (选择>努力) treats direction selection as more consequential than work intensity at the margin. The fifth model, 'controversy as transmission,' describes Zhang Xuefeng's communication strategy: pushing positions to their extreme edge to make them memorable.

Eight decision heuristics accompany these models. The 'median test' looks at where the middle 50 percent of a graduating cohort ends up, not the outliers in either direction. The '500 top company test' checks where recruiters actually go rather than what a company claims about itself. The '10-year pressure test' asks whether a student can accept earning less a decade from now than peers who scored lower. The 'family background bifurcation' heuristic explicitly gives different answers to the same question depending on whether the family has industry connections or not. The 'city-first principle' weights the destination city heavily, on the reasoning that geography shapes available thinking, resources, and opportunities over time.

All eight heuristics are listed explicitly in the README with short-form descriptions and attributed sources. A reader can cross-check that a specific response is applying one of the documented heuristics before treating it as representative of Zhang Xuefeng's method.

The nuwa-skill Pipeline That Produced This Skill

zhangxuefeng-skill was generated automatically by a companion skill, nuwa-skill (alchaincyf/nuwa-skill). The README describes the production pipeline in operational terms: one input, a person's name, triggers six agents running in parallel. Each agent produces one research file covering a distinct dimension: books and systematic thinking, on-record conversations, expression style, external criticism, key life decisions, and a full biographical timeline.

Cross-validation across those six files produces the final SKILL.md. The pipeline includes quality checks: three known test cases, one edge case, and a style verification pass. This production method is visible in the repository itself. The references/research/ directory holds the six source files, labeled 01-writings.md through 06-timeline.md, so the extraction logic is auditable.

This matters for anyone evaluating the skill's reliability. The quality ceiling is set by what those six files contain. The README credits 30 direct quotes, 11 key decision records, and a complete biographical timeline, all from primary sources: five books, B-station lecture recordings, a Sina Finance CEO interview, a Jiemian News deep interview, and a China News Weekly interview. Secondary sources from Titanium Media, Huxiu, Sanlian Life Weekly, 36Kr, and 21 Economic Network were included. Zhihu, WeChat public accounts, and Baidu Baike were explicitly excluded as sources. That sourcing discipline is documented, not assumed.

Anyone who wants to produce a comparable skill for a different public figure can install nuwa-skill and run the same pipeline with a different name as input.

Installation: The npx One-Liner and Manual Paths

The recommended installation uses a single command that auto-detects the current agent runtime:

npx skills add alchaincyf/zhangxuefeng-skill

This uses the vercel-labs/skills universal CLI installer, which the README says supports over 55 runtimes. Supported runtimes explicitly named in the README include Claude Code, Codex, Cursor, OpenClaw, Hermes Agent, CodeBuddy, Workbuddy, Gemini CLI, and OpenCode. To target a specific runtime without auto-detection, add the runtime flag:

npx skills add alchaincyf/zhangxuefeng-skill -a claude-code

For manual installation, clone the repository directly to the skills directory for the target runtime. For Claude Code, the correct destination is the skills subdirectory under ~/.claude/:

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

After installation, the skill is not active by default. Each session requires an explicit invocation. The README shows example prompts: 'Use Zhang Xuefeng's perspective to analyze this major choice,' 'Switch to Zhang Xuefeng, my child is filling out the gaokao form,' or 'How would Zhang Xuefeng evaluate this career direction?' The skill file is essentially a markdown document with YAML frontmatter. For runtimes that do not support automatic SKILL.md loading, the README says the file's content can be pasted directly into any conversation to achieve the same effect.

Repository Structure and What the Six Research Files Contain

The repository has a compact structure. The top-level entries are README.md, SKILL.md, the references/ directory, the examples/ directory, a .gitignore, and a LICENSE file. A WeChat QR code image is also in the top level.

SKILL.md is the deployable artifact. It encodes the five mental models, eight decision heuristics, expression DNA, and five documented internal tensions in a format the Agent Skills protocol can load. The expression DNA section specifies the rhetorical rhythm (set up the misconception, refute it with data, land a memorable phrase, rephrase and repeat), vocabulary choices (survival, employment, salary, sieve, stepping stone, dead-end field), sentence cadence (short sentences, fast rhythm, high information density), and characteristic markers from Zhang Xuefeng's Northeast dialect.

The references/research/ directory holds the six source files: 01-writings.md, 02-conversations.md, 03-expression-dna.md, 04-external-views.md, 05-decisions.md, and 06-timeline.md. A reader who wants to trace a specific claim in a skill response can work backward through these files to the named primary source.

The examples/demo-conversation.md file records actual conversations with the installed skill. These are the full demonstrations the README quotes: the financial advice to the Henan student, the graduate school discussion, the AI-era major update, and the case for choosing a livable profession before pursuing art. The README notes that all four deployed specific mental models explicitly, with citations in the response text to which model is being applied.

Limitations: Scope, Runtime Dependency, and the Skill's Embedded Assumptions

The skill is focused entirely on the Chinese gaokao, domestic graduate admissions, and career planning within China's job market. Questions about education systems outside China, non-Chinese professional hiring, or fields outside Zhang Xuefeng's documented commentary are outside its scope. The README does not claim otherwise, but users should be aware that the 'employment backtracking method' depends on Chinese labor market data that Zhang Xuefeng tracked closely. The skill cannot update that data after the cutoff of its training sources.

A second limitation is runtime dependency. If an agent runtime does not load SKILL.md automatically, the skill file has no effect. The universal CLI installer handles this for supported runtimes, but users must verify their specific runtime appears on the supported list.

Third, the skill applies Zhang Xuefeng's documented frameworks, which carry explicit assumptions. The 'class realism' model gives different advice based on family wealth. The 'social sieve theory' frames education credentials as filtering mechanisms. Users who find these framings reductive will receive advice shaped by them regardless.

The README documents five internal tensions in Zhang Xuefeng's thinking that the skill deliberately preserves rather than resolves: his identity as a champion of the working class alongside his personal wealth; his successful cross-field career pivot versus his advice to students to choose practical majors; a stated concern for health alongside documented extremely long work hours; a stated intention to hold back versus a faster-than-thinking delivery style; and the question of whether his controversy-generating approach is strategy or character. These preserved contradictions mean a skill session may produce inconsistent guidance across different framings of the same underlying question.

The repository was last pushed on 2026-08-25 and has no GitHub releases.

Editorial conclusion

Engineers who want a Zhang Xuefeng persona in an AI workflow and can confirm their runtime supports the Agent Skills protocol will find this the most direct installation path. Students expecting a general-purpose chatbot should verify first that their runtime loads SKILL.md automatically; without that loading step, the file is just text. The MIT licence allows reuse and modification. Readers who want to audit the distillation can trace every mental model to the six source files in references/research/, which name each primary interview and publication the extraction drew on.

Frequently asked questions

Official sources

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