yupi-skill: distilling a programmer-influencer's judgement into an Agent Skill

The problem: general assistants answer career questions without a position

Ask a stock assistant whether to take a 13K outsourcing offer from a bank and you get a list of factors. The README frames this as the motivating complaint: the AI gives you "需要综合考虑多方面因素", a correct sentence with no decision in it. The same pattern shows up in study planning and resume review, where the model produces textbook advice that is technically true and useless for someone choosing between two concrete next steps.

yupi-skill is for people who want a named point of view instead. The README's comparison table puts the contrast bluntly: where a normal assistant says "建议系统学习数据结构与算法", the loaded skill says to run a user-centre project first and that two to three months is enough. That is not better reasoning in the abstract. It is a persona with stated preferences, and the whole project exists to make those preferences portable across tools.

The audience is narrow and identifiable: Chinese-speaking developers and students who already follow 程序员鱼皮's content, and who want the same register inside their editor rather than in a browser tab. Anyone who wants neutral, hedged, multi-perspective output is not the target and will find this actively annoying.

How the skill is structured: three reference files and progressive loading

The repository is small. The top level holds LICENSE, README.md, SKILL.md and a references/ directory. Inside references/ there are three files: identity.md for identity, mental models, decision rules and bottom-line principles; voice.md for sentence patterns, verbal tics, humour and answer templates; and knowledge-sources.md for designated information sources and web-search rules.

The README says the skill follows the AgentSkills open standard and loads progressively: the agent reads metadata first to decide whether to trigger, then reads SKILL.md on a trigger, then pulls in references on demand. That ordering matters. It means the persona is not injected into every conversation. It sits dormant until a question about programming study, interviews, AI coding, resumes, technology selection or startup experience looks like a match, at which point the agent says it will answer in 鱼皮's style.

What is actually distilled is documented as seven mental models and a set of decision rules. The models include direction judgement before commitment, MVP shell with no compromise on the core, survival before ideals, and learning what companies need. The rules are more concrete and more useful to read: secure a fallback before reaching higher, ship before over-planning, do not skip steps, treat verbal promises as non-binding, and use plain language over jargon. These are the parts that give the answers their shape. The table in the README is a summary, and the real content lives in identity.md.

Installing yupi-skill for Claude Code, Cursor and OpenClaw

There is no package manager step. Installation is a git clone into the directory each host tool scans for skills. For Claude Code, the README gives a project-scoped install that must be run from the root of a git repository, and a global alternative under the home directory.

mkdir -p .claude/skills
git clone https://github.com/liyupi/yupi-skill.git .claude/skills/yupi-skill

After that, the folder .claude/skills/yupi-skill should contain SKILL.md and references/. If you want the skill available in every project instead, the README's global form clones into ~/.claude/skills/yupi-skill.

Cursor uses a parallel layout with its own directory name.

mkdir -p .cursor/skills
git clone https://github.com/liyupi/yupi-skill.git .cursor/skills/yupi-skill

OpenClaw places the clone under a workspace path rather than a dot-directory.

mkdir -p ~/.openclaw/workspace/skills
git clone https://github.com/liyupi/yupi-skill.git ~/.openclaw/workspace/skills/yupi-skill

The README states that after installation the AI recognises the skill automatically and uses it in suitable situations. It also documents a manual trigger: tell the AI to answer in 鱼皮's style. That second route is worth knowing, because automatic triggering depends on the host tool's own matching against the metadata, and you may want the persona for a question the matcher would not classify as a career or study query.

For a first real use, pick something with a decision in it. The README's own examples are offer evaluation, study planning and technology trends, and the pattern they share is that the user supplies a specific situation with numbers and constraints. A vague question gets a vague answer regardless of the skill.

Built-in information sources and the search rule

The skill does not answer everything from the distilled material. knowledge-sources.md maps question types to sites the agent should consult: codefather.cn for programming study and project tutorials, ai.codefather.cn for AI tools and AI coding, mianshiya.com for interview questions, laoyujianli.com for resume optimisation, dogyupi.com for the full product picture, and github.com/liyupi for open source projects.

This is the most consequential design decision in the repository, and it cuts both ways. On one hand it keeps answers current and grounded, and the README is explicit about the failure mode it avoids: for technical depth questions the skill searches first and then answers in style, rather than inventing. On the other hand, the designated sources are the author's own properties. The skill is not a neutral retrieval layer; it is a retrieval layer pointed at one person's products. That is coherent with the project's purpose, but a reader should know it before treating the output as independent research.

The search rule also creates a dependency the README does not resolve. If the host agent has no web access, or the user is offline, the skill falls back to the April 2026 material and the persona still speaks with the same confidence. Nothing in the documentation describes how the agent signals that it could not search.

Limitations the README admits, and one it does not

The limitations section is unusually candid for a persona project. It states that the material runs to April 2026 and that newer developments require a web search. It states that the skill cannot reproduce facial expression, tone of voice or body language from video. It states that when a question touches undisclosed private or commercial information, the skill will say it has not discussed that publicly. And it states that technical depth questions are searched rather than fabricated.

The first of those is the one that ages. The last push to the repository was on 2026-04-09, and the README's own update table shows how the material is meant to stay fresh: new articles and video scripts go into voice.md, new consultation dialogues into the sample section of voice.md, new products and projects into knowledge-sources.md, and new in-depth interviews into identity.md. That is a manual editorial process, not an automated one. A fork that nobody maintains will drift from the person it imitates, and the drift will be invisible in the answers because the voice file does not carry a timestamp the model is instructed to surface.

The unstated limitation is scope. This is a skill about advice, not about code. It will not review a pull request, run a test suite or check a dependency. Loading it into a repository where you wanted a coding assistant changes the register of the answers without adding any capability.

Alternatives: colleague-skill and writing your own

The README points to colleague-skill at github.com/titanwings/colleague-skill, described as a tool that automates distilling a colleague's knowledge and personality. The difference in approach is who does the analysis. yupi-skill is a hand-built artefact: the README describes a six-stage process of collecting material, classifying it across six dimensions, cross-validating to extract mental models and decision rules, asking twelve follow-up questions to find the boundaries of opinions and the things the subject refuses to do, writing to the AgentSkills format, and then testing three representative questions against real samples. colleague-skill aims at automating that pipeline for a different subject.

If your goal is a persona for someone whose material you have, the more direct alternative is the README's own tutorial for building your own skill: collect your articles, chat logs, video scripts and work documents; have an AI analyse your style and thinking; have it ask you ten to fifteen deep questions to find your decision boundaries; package the result with the AgentSkills standard. That route gives you a skill you can actually maintain, because you own the source material. Cloning this repository gives you someone else's, with someone else's product links embedded in the retrieval rules.

Licence and the real upgrade cost

The repository is MIT licensed, and the README repeats that at the bottom. MIT permits reuse, modification and redistribution with the licence and copyright notice retained. Two things sit outside what a licence can settle. The first is the persona itself: the skill imitates a named public figure, and the README describes it as built from that person's articles, consultation dialogues and internal documents. Whether you are comfortable shipping that inside a company tool is a judgement about publicity and endorsement, not about copyright, and this article is not legal advice. The second is the embedded source list, which points at the author's commercial properties.

Upgrade cost is low in the mechanical sense and non-trivial in the editorial sense. Updating means pulling the repository or editing files under references/. There is no version pinning, no changelog and no release history in the repository material, so a git pull can change the persona's answers with no signal. If you depend on the current behaviour, fork it and treat your fork as the artefact. If you want the author's latest material, track the upstream clone and accept that answers will shift.

Editorial conclusion

Adopt yupi-skill if you already work inside Claude Code, Cursor or OpenClaw and want opinionated Chinese-language answers about programming study paths, offers, resumes and AI tooling from a defined persona rather than a neutral assistant. Do not adopt it as a knowledge base: the README states the material runs to April 2026, the skill cannot reproduce tone of voice or body language from video, and it will say so when a question touches private or commercial matters it has not discussed publicly. Before relying on it, open SKILL.md and the three files under references/ and read the trigger conditions and the knowledge-sources list, because that is where the behaviour actually lives.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. liyupi/yupi-skill on GitHub
4. Project website
5. README