yupi-skill: Packaging a Person's Decision Style as an Agent Skill

The Problem: Generic Coding Advice With No Position

The README opens with four complaints rather than a feature list. Ask an AI for a programming learning route and you get a textbook with no point of view. Ask it to review a resume and you get correct but useless filler. Ask whether to take a job offer and you get "many factors need to be considered." Ask what to learn in the AI era and you get a comprehensive list with no stance. That is the gap yupi-skill targets. It is not trying to make an agent smarter about Java or system design. It is trying to make the agent commit to an answer. The README's comparison table makes the intended difference concrete: where a plain model says "AI and backend each have advantages, it depends on personal interest," the skill is meant to say that a resume without AI content gets filtered out by many HR screens. The audience is Chinese-speaking programmers at the learning, job-hunting or early-career stage who already consume the author's content and want that framing available inside their editor. Anyone looking for a general-purpose coding assistant is not the target, and the skill does not pretend to be one.

What Actually Got Distilled: Mental Models and Decision Rules

The repository separates persona into layers. identity.md holds seven named mental models and a set of decision rules. The models are stated as one-line heuristics: judge the direction first and then commit fully, or cut losses if the direction is wrong; ship a minimum viable shell but never compromise the core value; secure survival before chasing the ideal; learn what enterprises need rather than what interests you; let the audience determine the content form; avoid over-promising early; and reduce future anxiety by iterating on what is in front of you. The decision rules are the operational half: have a fallback before reaching higher, act before over-planning (one month to launch beats three months of planning), do not skip steps, treat verbal promises as non-binding, and drop jargon when plain speech will do. This is the part worth reading even if you never install the skill. It is a compact statement of a working philosophy, and it is specific enough to disagree with. The rule that only contracts and already-happened facts count, for instance, is a defensible position in an offer-negotiation conversation and a bad one in a research conversation where provisional findings matter.

Repository Layout and the Progressive Loading Mechanism

The skill follows the AgentSkills open standard and is built for progressive loading. The tree is small: SKILL.md at the root, plus references/identity.md, references/voice.md and references/knowledge-sources.md. The README describes the flow as three stages. The agent reads metadata first to decide whether the skill should trigger at all. If it triggers, the agent reads SKILL.md, which the README says contains trigger conditions, workflow and answer modes. Only then does it load files from references/ as needed. That ordering matters for cost: a skill that pulls three reference documents into context on every unrelated question would be a tax on the whole session, and the standard exists to avoid that. The trade-off is that the agent has to make a correct trigger decision from metadata alone. A question that is adjacent to the skill's domain but not clearly inside it may never reach SKILL.md, and the user will not necessarily notice the skill stayed dormant. The README offers a manual override: tell the AI to answer in 鱼皮's style.

Installation Across Claude Code, Cursor and OpenClaw

Installation is a git clone into a tool-specific directory, and the README gives the paths. For Claude Code in a single project, run mkdir -p .claude/skills followed by git clone https://github.com/liyupi/yupi-skill.git .claude/skills/yupi-skill from the repository root. For a global install available to every project, the target is ~/.claude/skills/yupi-skill. Cursor uses mkdir -p .cursor/skills and clones into .cursor/skills/yupi-skill. OpenClaw clones into ~/.openclaw/workspace/skills/yupi-skill. There is no package manager, no build step and no configuration file to edit. That is the whole setup. The README states the AI will recognize and use the skill automatically in appropriate situations, with the manual trigger as a fallback. Two practical notes follow from the commands. First, the project-level Claude Code install must be run at the git repository root, so the skill lands inside the project rather than somewhere else. Second, because these are plain directories, updating means pulling or re-cloning; there is no version pinning, and the README lists no releases.

Built-In Sources and What Happens When They Are Unreachable

knowledge-sources.md maps question types to six external sites the agent is instructed to consult: codefather.cn for learning routes and project tutorials, ai.codefather.cn for AI tools, mianshiya.com for interview questions, laoyujianli.com for resume optimization, dogyupi.com for the product catalog, and github.com/liyupi for open source work. The README's stated behavior is that the agent searches these before answering technical questions rather than improvising. This is the most fragile part of the design. The skill is a set of markdown instructions, so its factual freshness is bounded by network access and by whether those pages still say what the instructions assume. The README concedes the first half of this in its limitations section: the material is current as of April 2026 and newer developments require online search. It does not address the second half. If a source changes structure or the agent cannot reach it, the fallback is the distilled persona, which will still produce a confident-sounding answer. That is the failure mode to watch for, and it is not detectable from the output alone.

Stated Limitations and the Cases Where This Is the Wrong Tool

The README lists four limitations, and they are more honest than most. The knowledge cutoff is April 2026. Video expressions, tone and body language cannot be reproduced. On private or commercial matters the skill is instructed to say plainly that the topic has not been discussed publicly. Technical depth questions get a search first, then an answer in the author's style, rather than an invented one. Those constraints define the boundary. If your question is about a specific library's behavior, a compiler error, or a performance characteristic, this skill adds a stylistic layer over a search you could run yourself, and the persona may push toward a confident conclusion that the underlying material does not support. The same applies to any domain the author does not publicly cover. The refusal behavior is a feature, but it means the skill will go quiet exactly where a general assistant would still help. Treat it as a second opinion on career and learning decisions, not as a technical reference.

Where a Plain System Prompt Still Wins

The obvious alternative is a hand-written system prompt or a custom instruction file describing the tone you want. The difference is scope and maintenance. A system prompt lives in your tool's settings and applies to every conversation; it is one block of text you edit directly. yupi-skill is a directory of four markdown files that loads conditionally, which means it stays out of unrelated sessions but also means you cannot tune it by editing one line in a settings panel. The README's update table is the clearest statement of the trade-off: new articles and video scripts go into voice.md, new consultation dialogues go into the sample section of voice.md, new products go into knowledge-sources.md, and new in-depth interviews go into identity.md. That is a real maintenance surface, and it is larger than a prompt. If you want a consistent tone across all your work, a system prompt is simpler. If you want a persona that activates only for career and learning questions and stays silent otherwise, the skill structure is the better fit. The README also points to colleague-skill as a separate project for automating the distillation of a colleague's knowledge and personality, which is a different problem.

Licence, Maintenance and What to Check Before Adopting

The project is MIT licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. This is not legal advice; read the LICENSE file in the repository if the terms matter to your situation. Maintenance is the open question. The README documents a six-step distillation process (collecting 20+ articles, 13 real consultation dialogues, 3 video scripts and one internal document, then categorizing, cross-validating, running 12 follow-up questions, writing to the AgentSkills format, and validating against three sample questions), but it does not describe a release cadence, and no releases were retrieved. The update path is manual: add material to references/ and re-clone or pull. For a skill whose value depends on staying current with a fast-moving field, that is a commitment the user takes on, not one the repository automates. Check the last commit date, read SKILL.md to confirm the trigger conditions match your questions, and verify the source sites are reachable from your network before you rely on the search step.

Editorial conclusion

Adopt yupi-skill if you want an agent that gives opinionated, conclusion-first answers about Chinese-language programming careers, learning routes and AI-era skill choices, and you accept that the opinions are one person's and dated to April 2026. Do not adopt it if you need neutral advice, verifiable technical depth, or anything outside the author's public scope; the README itself says private and commercial topics get a refusal. Before installing, open SKILL.md and confirm the trigger conditions match the questions you actually ask, then check that the six built-in sources in knowledge-sources.md are reachable from your network, because the skill's freshness depends on them.

Official sources

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