zhangxuefeng-skill: Zhang Xuefeng's decision heuristics as an Agent Skills package

The problem: career advice from a general-purpose model is bland by default

Ask a chat model whether a 560-point Henan student should study finance and you will usually get a balanced list of considerations. Zhang Xuefeng's public persona is the opposite of balanced. The README states the project's premise directly: it is not a quote collection, it is a runnable thinking framework. The target user is someone who wants the agent to commit to a position, ask clarifying questions first, and reason from employment outcomes rather than from interest and self-fulfilment.

The README's own example shows the intended shape of an answer. Asked about finance for a 560-point student in Henan, the skill output opens with "停停停，你先别急着说金融。我问你几个问题。" and then asks whether the family works in finance before giving a recommendation. That is a specific interaction pattern: interrogate the family's resources, then route to a conclusion. If what you want is a neutral pros-and-cons memo, this is the wrong package, and the README does not pretend otherwise.

What is actually in the repository: SKILL.md plus six research files

The top level contains LICENSE, README.md, SKILL.md, assets/, examples/, references/ and wechat-qrcode.jpg. There is no build step, no dependency manifest, no test suite and no executable entry point. The primary language field is unrecorded, which fits a repository whose main artifact is Markdown with YAML frontmatter. The README says as much in the manual-installation section: even on a runtime that cannot auto-load Agent Skills, you can paste the contents of SKILL.md into a conversation because it is fundamentally markdown plus YAML frontmatter.

The substance sits in references/research/, described as six files: 01-writings.md, 02-conversations.md, 03-expression-dna.md, 04-external-views.md, 05-decisions.md and 06-timeline.md. The README claims these were built from five books, more than fifteen long-form media interviews, more than thirty first-hand quotes, eleven decision records and a full life timeline. None of that is verifiable from the repository listing alone, so the research files are the place to check before you trust a specific assertion.

The mechanism: five mental models, eight heuristics, one expression layer

The README describes three layers. First, five mental models: 社会筛子论, 选择 > 努力, 就业倒推法, 阶层现实主义 and 争议即传播. Each is given a one-line summary and an attributed source, for example 就业倒推法 is tied to the business model of 峰学蔚来, and 社会筛子论 is credited to lectures and livestreams where it was used more than twenty times.

Second, eight decision heuristics: 灵魂追问法, 中位数原则, 不可替代性检验, 500强测试, 家庭背景分流, 城市优先原则, 10年后压迫测试 and 认态度不认事实道歉法. These are the operational layer. 灵魂追问法 is the rapid-question routine that opens the finance example. 中位数原则 is the instruction to look at the middle fifty percent rather than the top or bottom of an outcome distribution. 500强测试 says to look at where companies recruit and what majors they take rather than at what they say publicly.

Third, an expression DNA layer covering sentence rhythm, vocabulary, and a four-step pattern the README describes as set up a misconception, hit it with facts, summarise in a punchline, then repeat the point in different words. The README also keeps five tensions, such as the gap between being a spokesperson for poor families and being a wealthy one, and between having succeeded through a cross-major move and advising others to pick the right major. Those tensions are the most interesting design decision in the package: a skill that flattened them would produce a caricature.

Installing it with npx skills add and running a first prompt

The README gives three installation routes. The recommended one is a single command from the cross-runtime installer maintained by vercel-labs, which the README says supports more than fifty-five runtimes and detects the current one automatically.

npx skills add alchaincyf/zhangxuefeng-skill

If detection picks the wrong target, the README shows a runtime flag, for example -a claude-code, -a codex, -a cursor or -a openclaw. The README does not list every accepted value.

The manual route clones the repository into the skills directory of your runtime. The README's table gives Claude Code as ~/.claude/skills/zhangxuefeng-skill/, Codex CLI as ~/.codex/skills/zhangxuefeng-skill/, Cursor as ~/.cursor/skills/zhangxuefeng-skill/ and OpenClaw as ~/.openclaw/workspace/skills/zhangxuefeng-skill/. For Hermes Agent the README says to run that runtime's install script or clone into its skills directory, without naming a path.

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

After installation the README suggests invoking it in natural language rather than through a command. Its examples are "用张雪峰的视角帮我分析这个专业选择", "张雪峰会怎么看这个职业方向？" and "切换到张雪峰，我孩子要填志愿了". What you should see is a reply that asks questions before answering, in the style of the finance example above. If your agent answers immediately with a balanced list, the skill was not loaded.

Where it breaks down: no data, no rollback, no neutrality

The most important limitation is that the repository ships reasoning patterns, not admissions data. 就业倒推法 tells the agent to look at where the middle fifty percent of graduates end up, and the file list shows no bundled employment dataset, score-line table or provincial admission index. The agent still has to supply those numbers from its own knowledge or from a search tool, and the skill has no mechanism for checking them. If you need figures you can defend to a parent, this package is the wrong tool.

The second limitation is runtime dependency. On a runtime that does not implement Agent Skills, the README's fallback is to paste SKILL.md into the conversation. That works, but you lose the research files unless you paste them too, and you lose whatever the runtime does with skill metadata.

The third is the absence of any evaluation harness. The README shows four worked examples and points to examples/demo-conversation.md for the full records, but the file list shows no scoring rubric or test set of questions. Removal means deleting the cloned directory or whatever the npx installer placed.

Finally, the package is explicitly opinionated. 争议即传播 is listed as a mental model, meaning the skill is designed to produce statements strong enough to be memorable. For a family making a real decision, that is a feature only if you independently verify the factual premises underneath the rhetoric.

Alternatives: a general agent prompt versus this distilled skill

The realistic alternative is not another repository. It is writing your own system prompt, or using your agent's default behaviour with a short instruction such as "act as a blunt Chinese college-admissions advisor who asks about family background first". The difference in approach is where the structure lives. A hand-written prompt encodes whatever you think to write in one sitting. This package encodes a named set of models and heuristics, each with a stated source, plus a research directory you can read and disagree with file by file.

That structure cuts both ways. It gives you something auditable: you can point at 中位数原则 or 500强测试 and say the agent should not have skipped it. It also locks the output into one person's framework, including positions that the README itself flags as contested, such as the news-major controversy that the README says increased rather than decreased commercial interest. A generic prompt is weaker and more controllable. This package is stronger and narrower. Which one you want depends on whether you are optimising for a memorable recommendation or for a defensible one.

Licence, maintenance and what an upgrade costs you

The repository is MIT licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. The books and interviews listed as sources remain their authors' and publishers' property. If you plan to redistribute modified content that reproduces long passages from the research files, that is a question for a lawyer, not for the licence badge.

The last push was on 2026-10-13. The repository has no published GitHub releases, so there is no versioned upgrade path and no changelog to read. Updating means re-running the npx command or pulling the branch again, and because the artifact is prose rather than code, an update can silently change the agent's behaviour without any version number moving. If you have local edits to SKILL.md, a pull will conflict with them.

The maintenance cost is mostly editorial. The research files reference a body of work including a 2025 title the README describes as a posthumous work, and the timeline file is dated material by nature. Someone has to decide when a new interview or a changed admissions landscape invalidates a heuristic.

Editorial conclusion

Adopt it if you already run an Agent Skills-compatible agent and want a structured counterweight to generic career advice, and you accept that the output is a distillation of one person's public positions rather than a neutral counselling method. Do not adopt it if you need admissions data you can cite to a family, or if your runtime cannot load SKILL.md. Before relying on it, open SKILL.md and read the five mental models and eight heuristics yourself, then run one of the examples in examples/demo-conversation.md and check whether the reasoning shown there matches what your agent produces.

Frequently asked questions

Official sources

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