# tong-jincheng-skill: a Claude Code persona skill built from 200,000 characters of livestream transcripts > The repository packages one Chinese relationship commentator's decision heuristics into a Claude Code skill, installs with a single npx command, and states its own blind spots in the README. Here is what it does, where it breaks, and who should skip it. **hotcoffeeshake/tong-jincheng-skill** — 童锦程视角 Skill — 用深情祖师爷的思维框架分析人际关系 - Repository: https://github.com/hotcoffeeshake/tong-jincheng-skill - Stars: 2,661 · Forks: 329 - Language: Unknown - License: MIT - Published: 2026-09-28 · Updated: 2026-09-28 · Language: en - Canonical page: https://hysenlabs.com/projects/hotcoffeeshake-tong-jincheng-skill ## What tong-jincheng-skill actually packages This is not a chatbot and not a fine-tune. It is a Claude Code skill: a directory of Markdown that Claude Code loads when you invoke it, so the model answers inside a fixed set of heuristics instead of its default helpful-assistant voice. The subject is Tong Jincheng, described in the README as a top emotional-content streamer on Kuaishou and Douyin whose audience calls him the affectionate grandmaster. The material behind the skill is nine first-hand video transcripts, roughly 200,000 characters, split between a livestream clip compilation of about 110,000 characters, six dating vlogs, and two third-party analyses of his approach. The intended user is someone with a concrete interpersonal problem who wants a framework, not comfort. The README is explicit that the skill does not replay quotes: it applies his way of seeing a problem. The example exchanges show the shape of that. Asked about a woman who is warm one day and silent the next, the skill answers that if you are unsure whether she likes you, she does not, and that hot-and-cold behaviour means you are not important to her. Asked about testing a boyfriend by going silent, it answers that human nature does not survive testing, and that the real problem is an unspoken need. The register is short sentences, direct address, and occasional self-deprecating asides. ## The mental models the skill loads and why they constrain answers The README lists five mental models and nine decision heuristics, all extracted from the transcript corpus. The models are: attraction beats pleasing, because people like you for what draws them, not for how much you like them; give the other person a step, meaning people need a reason they can tell themselves before they act; human nature cannot be tested, so give conditions rather than tests; showing off exposes insecurity, since what a person lacks is what they display; and success splits the world in two, because after success everyone around you is good to you. The heuristics are the operational layer. Uncertainty about being liked means not being liked. When stuck, read or train, never drink. Anger solves nothing, so work on the solution. Spend money on people who matter, not on performances for acquaintances. This is the part worth judging on design grounds. A persona skill is a constraint system: every heuristic you load narrows the answer space. That is the point here, and it is also the cost. A framework that resolves ambiguity by declaring the ambiguous case a no will give fast, confident answers and will be wrong whenever the situation is genuinely mixed. The README does not present these as probabilistic; it presents them as his framework, and the skill answers in that voice. ## Installing it and getting a first answer There is no build step and no configuration file. Installation is one command, run from a terminal, and it requires Claude Code to be present since the skill is a Claude Code skill. ```bash npx skills add hotcoffeeshake/tong-jincheng-skill ``` The README documents no flags, no version pinning and no install directory; the command is given exactly as above. After installation, the skill is not active by default. You trigger it with one of five phrases, and the README lists them verbatim: ``` > 童锦程 > 深情祖师爷 > 用童锦程的方式 > 从童锦程视角 > 景辰怎么看 ``` Once triggered, you ask your question in the same session. The README's own examples are questions like a six-month crush on a colleague that was never voiced, or a partner who is neither warm nor cold. The output you should expect is a short, declarative analysis in Chinese, in the style of the sample exchanges: a direct verdict first, then the reasoning, then a concrete action such as saying the need out loud instead of testing for it. The five triggers are the only activation mechanism the README describes, so if a session does not respond in that voice, the skill was not loaded. ## The honesty section is the most useful part of the README Most persona repositories do not ship a limitations table. This one does, and it is specific. Under what the skill cannot do, the README lists four items: business decisions, because the source material is almost entirely about relationships and the data on livestream commerce and startup strategy is thin; replacing the person, because his current state, latest thinking and private personality cannot be copied; developments after 2025, because network access is restricted and recent information may be missing; and details of his relationship with Xinba, because he never discusses it publicly and the material is insufficient. That last row is unusual and worth noting. Naming a specific topic the corpus cannot cover is more useful than a generic disclaimer, because it tells you where the model would otherwise improvise. The skill also carries a stated bias: the README closes the section with the line that a skill which does not tell you where its limits are is not worth trusting. The limitation that matters most in practice is not on the list, though. The corpus is Chinese, the triggers are Chinese, and the sample answers are Chinese. Nothing in the README suggests an English output mode. If you do not read Chinese, the skill's answers will be of limited use even if the install succeeds. One more constraint comes from the repository layout itself. The top level holds LICENSE, README.md, SKILL.md, assets/ and references/. The research data lives under references/research/ in six files covering writings, conversations, expression DNA, external views, decisions and a timeline. Those files are readable before you install, which means you can audit the distillation rather than trusting the summary table. ## Where this approach differs from a general-purpose model The obvious alternative is asking the same question to Claude Code without the skill loaded. The difference is not knowledge, since a general model can reason about relationships perfectly well. The difference is stance. A default assistant tends to hedge, ask clarifying questions and present both sides. This skill is built to do the opposite: it resolves the ambiguity and commits, as in the sample where uncertainty about being liked is treated as an answer in itself. A second alternative is a different persona skill from the same lineage. The README states the skill was produced with the nuwa.skill methodology, which it describes as a skill that builds skills: give it a name and it runs research, distillation and validation. That matters for evaluation, because it means the quality of any given persona skill depends on the corpus behind it, not on the pipeline. Nine transcripts and 200,000 characters is a real corpus for one narrow domain; the same pipeline applied to a thinner source would produce a thinner skill, and the README's limitations table would look very different. A third comparison is simply reading the research files. If you want the frameworks without the persona voice, references/research/01-writings.md and 05-decisions.md contain the distilled material directly, and no model is involved. For some readers that is the better trade. ## Licence, maintenance and what upgrading costs you The licence is MIT, and the README states it plainly: use it, change it, build on it. For a repository that is mostly Markdown, that is the whole story. There is no bundled model, no service dependency and no API key, so there is nothing to relicense downstream beyond the text and the assets. The assets/ directory contains promotional images, including a WeChat mini-program QR code and a service-account QR code; those are brand assets, and MIT covers the repository's code and text rather than granting rights to someone else's likeness or trademarks, which is a question for a lawyer and not something this article resolves. The last push to the default branch was on 2026-06-22. The repository is not archived. There are no releases, which is consistent with a project that ships as a directory of files rather than a versioned package; the README gives no version number. Upgrading, in practice, means re-running the install command or pulling the repository, and the thing that changes between versions is the Markdown: SKILL.md, the README, and the six research files. If you have edited SKILL.md locally to adjust tone, an upgrade will overwrite it unless you keep your changes elsewhere. The README also promotes a WeChat mini-program under the name 情商练习有事问AI, where the README says you can upload chat screenshots for a conversation review or get help drafting a reply. That is a separate product from the skill, and the README does not describe how, or whether, the two share code. ## Conclusion Adopt tong-jincheng-skill if you already run Claude Code, read Chinese, and want a blunt second opinion on a relationship question rather than a sympathetic one. Do not adopt it for business decisions, for English-only workflows, or as a stand-in for the person himself; the README rules all three out. Before you install, open references/research/03-expression-dna.md and SKILL.md, since those two files decide the tone you will actually get, and check whether a persona skill fits your team's policy on sending personal chat content to a model. ## FAQ ### How do I install tong-jincheng-skill? Run npx skills add hotcoffeeshake/tong-jincheng-skill from a terminal with Claude Code available. The README gives no other install method and no configuration step. ### How do I activate the skill once it is installed? The README lists five trigger phrases: 童锦程, 深情祖师爷, 用童锦程的方式, 从童锦程视角 and 景辰怎么看. After one of them, you ask your question in the same session. ### What is tong-jincheng-skill built from? The README states it is distilled from nine first-hand video transcripts totalling roughly 200,000 characters, including a livestream clip compilation of about 110,000 characters, six dating vlogs and two third-party analyses. The underlying analysis files sit in references/research/. ### What can tong-jincheng-skill not do? The README's honesty section rules out business decisions because the corpus is almost entirely relationship content, replacing the person himself, and topics after 2025 because network access is restricted. It also states the material on his relationship with Xinba is insufficient. ### What licence does tong-jincheng-skill use? MIT. The README describes it as free to use, modify and build on, and the repository carries a LICENSE file at the top level. ## Sources - [hotcoffeeshake/tong-jincheng-skill on GitHub](https://github.com/hotcoffeeshake/tong-jincheng-skill) - [Issues](https://github.com/hotcoffeeshake/tong-jincheng-skill/issues) - [License: MIT](https://github.com/hotcoffeeshake/tong-jincheng-skill/blob/main/LICENSE) - [README](https://github.com/hotcoffeeshake/tong-jincheng-skill/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/hotcoffeeshake-tong-jincheng-skill