# x-mentor-skill is a reference package loaded only when needed > x-mentor-skill is an Agent Skills package that turns six creators' writing and growth methods plus platform research into an operating manual for X, structured as a short SKILL.md and seven reference files loaded on demand. It is a prompt package rather than code, and its own README withdraws most of the rules the earlier version asserted. **alchaincyf/x-mentor-skill** — X导师.skill — 女娲的第一个「非人类」作品。蒸馏6位顶级X创作者方法论 + 开源算法数据，提炼完整的选题-写作-增长操作手册。Made with 女娲.skill - Repository: https://github.com/alchaincyf/x-mentor-skill - Stars: 1,303 · Forks: 203 - Language: Unknown - License: MIT - Published: 2026-10-17 · Updated: 2026-10-17 · Language: en - Canonical page: https://hysenlabs.com/projects/alchaincyf-x-mentor-skill ## A prompt package, not an executable The repository contains `SKILL.md`, a `references/` tree, an `examples/` directory, an `assets/` folder for the animation and images, and a licence file. The primary language is recorded as unknown, which is the honest classification for a package whose deliverable is prose plus YAML frontmatter. That shapes what you can evaluate and how. There is no build step, no test suite and no runtime behaviour to inspect, so the questions that matter are: how large is the context the agent loads, who wrote the material, when was it written, and does it distinguish what it knows from what it assumes. The README answers the size question directly, which is the design decision that makes this package different from a long prompt someone pastes into a chat. It describes the artefact as markdown plus YAML frontmatter that you can paste into a conversation even if your runtime does not support Agent Skills loading at all. Structure is `SKILL.md` at the top, a `user-data/` directory for diagnosis output generated automatically, and a references directory that is explicitly labelled as loaded on demand. The reference files come with line counts, which is a detail worth respecting. `mental-models-heuristics.md` is 221 lines, `writing-workshop.md` is 111, `growth-monetization.md` is 92, and four more sit between 36 and 57 lines each. Under `references/research/` there are six research files totalling 2475 lines, described as April 2026 historical material whose current rules live in `SKILL.md`. That split between an operational layer and a research layer is what lets a model load forty lines when it needs a principle and skip twenty-five hundred when it does not. ## What the 2026-10-11 update took back The most informative document in the repository is the changelog section of the README, and it is a retraction rather than an addition. The update folds in a content review from a named creator and describes changes that remove constraints. It cancels the 120 to 130 character limit, the rule that an outbound link must go in a reply, a default of 60 percent English, and a fixed thread length. It states that the previous reading of the algorithm was wrong, specifically that the weights should not be understood as a conversion rate from replies to likes. It says a post is not automatically deleted and reposted after 30 minutes of poor performance, but observed at 24 and 72 hours before testing one variable. Two changes are about rigour rather than content. Account diagnosis now separates the main post, replies, native Articles and outbound links, and separates impressions, attributed follows, net follower change and usage behaviour. And the package adds a feedback loop plus correction for rules that stop working. The paragraph after the list is the one to keep in mind. It says independent rule trials and statistical review were run, that conflicts in the reference files were fixed, and that this verifies rule consistency and calculation basis without claiming the growth effect has been validated by operating experiments. It also says the raw backend account data, private strategy files and local reports are not included in the repository, only the publicised migration examples and expression examples. That is an unusually candid statement for a prompt package, and it changes what you should use it for. The rules are internally consistent and derived, not experimentally validated. ## Persona skills versus topic skills The README draws a comparison that explains why this project exists rather than being one more personality prompt. The axis is what gets distilled. A persona skill distils one person's way of thinking, from research organised around one individual across six dimensions, and its output style imitates that person's tone. A topic skill distils a field's methodology, drawing on several practitioners plus platform data, and its output is a neutral operating manual. The claimed difference in value is seeing your problem through someone else's eyes versus receiving an action plan. The mental models table names six, each attributed to specific creators. The lean validation flywheel, to validate with a post before expanding, is attributed to Cole, Bush, Sahil, Hormozi and Welsh together. Attention engineering is credited to Cole plus Hormozi's value equation. Category creation is credited to Cole's Snow Leopard work and Koe's Niche of One. Value front-loading, giving the secret away and selling execution, is Hormozi, Welsh and Sahil. Building in public is credited to two named practitioners in the open-source world rather than to X creators, which is worth noting since the package is about X. Systemised compounding, templates replacing inspiration, is Welsh and Koe. The README then makes a distinction that matters for anyone tempted to treat this as a growth prediction tool. It says X's own documentation states the ranking weights apply to predicted probability rather than raw interaction counts, therefore a reply cannot be converted into a like. It says early research files are retained for traceability and are not current algorithm facts, and that unverified claims about timing, frequency and language ratio need validation against your own account. So the package distinguishes three things explicitly: platform facts, individual cases, and rules that need testing. That structure is more useful than the content of any one rule. ## Installation, and the runtime list is the real compatibility story Installation has three documented paths, and the first one does the work for you. ```bash npx skills add alchaincyf/x-mentor-skill ``` That calls a general purpose CLI installer from vercel-labs/skills which, according to the README, detects the current runtime and places the skill in the right directory. It supports 55 or more runtimes, and you can name one explicitly with flags such as `-a claude-code`, `-a codex`, `-a cursor` or `-a openclaw`. The README lists Claude Code, Codex, Cursor, OpenClaw, Hermes Agent, CodeBuddy, Workbuddy, Gemini CLI and OpenCode among the runtimes the skill is claimed to run on. The second path is manual, and the README gives a table of directories: The clone command is `git clone https://github.com/alchaincyf/x-mentor-skill` followed by the destination path for your runtime. Claude Code expects it at `~/.claude/skills/x-mentor-skill/`, Codex CLI at `~/.codex/skills/x-mentor-skill/`, Cursor at `~/.cursor/skills/x-mentor-skill/`, and OpenClaw at `~/.openclaw/workspace/skills/x-mentor-skill/`. For Hermes Agent the README says to run that runtime's install script or clone into its skills directory. The third path is the one with no tooling at all: paste the contents of `SKILL.md` into a conversation. That works because the file is markdown with YAML frontmatter, so an agent can read it as a prompt with no installation at all. If you are evaluating whether this package is any use to you, that is the fastest route. Activation is natural language rather than a command. The README says to tell your agent you want help writing a post, want X strategy, ask how to grow on Twitter, want a post reviewed, or want your account analysed. The skill then decides which reference files to load. ## It was generated by another skill, and that is the larger story This package did not appear on its own. The README attributes it to another project by the same author, a skill called nuwa-skill, and describes the variant flow that produced it. The flow has six steps. Confirm what is being distilled, which here is a topic with three to five core people rather than a single person. Run six parallel research agents, each covering a dimension: writing methodology, growth engines, content branding, platform algorithm, the AI niche, and cases with anti-patterns. Cross-validate to find where the six people's methods agree and where they diverge. Build the operating manual without imitating anyone's tone. Run a two-agent validation pass covering known tests, edge cases and style. Then iterate with an auto-optimizer that scores on an eight-dimension rubric and hill-climbs to a stated score of 88. Read as an engineering process, that is a research pipeline: parallel retrieval, cross-validation, synthesis, then evaluation against a rubric. The 88 is a self-reported rubric score, not an external benchmark. The research layer in this repository matches those six dimensions file by file: writing methods, growth engines, content and brand, platform mechanics, the AI and tech niche strategy, and cases with anti-patterns. Line counts of 503, 386, 398, 415, 404 and 369 add up to the stated 2475, so the structure is complete rather than partial. One sourcing decision is worth repeating because it is unusual: the README states that its information sources exclude Zhihu, WeChat public accounts and Baidu Baike. For research about writing methods circulating in Chinese social platforms, excluding those is a deliberate choice about source quality, and it also means the package has no references to the discourse it appears in. The README closes by listing other skills from the same generator for people and topics, which places this project as one instance of a repeatable process rather than a one-off artefact. ## What it is worth taking if you take nothing else Judged as engineering, this is a well-documented prompt artefact with a clear provenance, a stated scope and an unusually honest changelog. Judged as advice, it is six people distilled into rules that its own author says were not validated by experiment. Both judgements can be true, and the second one does not stop it being useful. The parts that survive even if you reject the rules are structural. The separation of an operational layer from a research layer means you can read forty lines instead of twenty-five hundred. The distinction between platform facts, personal cases and rules needing validation is a habit worth adopting in your own prompts. The account diagnosis format, which separates the main post, replies, native Articles and outbound links and separates impressions from attributed follows from net follower change from usage, is a genuinely better reporting shape than most growth advice gives, and it is free to copy regardless of the skill. The time-window advice is the most practically useful line and it came from a retraction. Observing a post at 24 and 72 hours rather than deciding at 30 minutes is correct advice independent of who said it, and the package now states it after having previously asserted something else. The failure modes to watch for are specific. The methodology is drawn from creators whose audiences are individual knowledge workers, so its growth reasoning transfers to technical content better than to a consumer brand. The follower counts and revenue figures in the README are the author's characterisations of other people, not measurements you can rely on. And the package describes itself as a distil of a field, which is a claim about coverage rather than a guarantee that the field was read. Licensing is MIT with the licence file at the repository root, so the material can be reused and modified with the notice retained. The repository is not archived, the last push was on 2026-10-13, and there are no GitHub releases, so updates arrive on the default branch `master` with no version history to pin against. ## Conclusion x-mentor-skill fits a writer or developer who publishes technical work on X and wants a documented decision procedure rather than tips, since it states its sources, marks which claims are platform facts versus personal cases, and includes an account diagnosis format worth copying on its own. Do not adopt it as a growth guarantee, because the README explicitly says its rules were validated for internal consistency and not by operating experiments, and do not adopt it expecting platform knowledge to stay current, since the package itself withdrew its earlier algorithmic claims. Verify first that your runtime supports the Agent Skills protocol, and read the 2026-10-11 update before following any rule that contradicts what you remember from the earlier version. ## FAQ ### How do I install x-mentor-skill? Run `npx skills add alchaincyf/x-mentor-skill`, which uses a general purpose installer that detects your runtime and places the skill in the right directory. You can also clone it into a runtime's skills folder manually, or simply paste the contents of SKILL.md into a conversation since the file is markdown with YAML frontmatter. ### Which AI runtimes support x-mentor-skill? It follows the open Agent Skills protocol, and the README lists Claude Code, Codex CLI, Cursor, OpenClaw, Hermes Agent, CodeBuddy, Workbuddy, Gemini CLI and OpenCode among fifty-plus runtimes. Manual paths are given for Claude Code, Codex CLI, Cursor and OpenClaw in the table of skills directories. ### What changed in the 2026-10-11 update to x-mentor-skill? It removed several hard constraints, including the 120 to 130 character limit, the rule that links must go in replies, a default of 60 percent English and a fixed thread length. It also stopped treating a poor 30 minute result as a reason to delete and repost, and split account diagnosis into main post, replies, native Articles and outbound links. ### Are x-mentor-skill's rules validated? The README states that independent rule trials and statistical review verified rule consistency and calculation basis, and explicitly says this does not mean the growth effect has been validated by operating experiments. Raw backend account data and private strategy files are not included in the repository. ### What is the difference between a persona skill and a topic skill? According to the README, a persona skill distils one person's way of thinking and imitates their tone, while a topic skill distils a field's methodology from several practitioners plus platform data and produces a neutral operating manual instead of an imitation. ### What license is x-mentor-skill released under? MIT, with the LICENSE file at the repository root. The repository is not archived, its last push was on 2026-10-13, and it has no GitHub releases, so updates land on the default branch master with no version history to pin. ## Sources - [alchaincyf/x-mentor-skill on GitHub](https://github.com/alchaincyf/x-mentor-skill) - [Issues](https://github.com/alchaincyf/x-mentor-skill/issues) - [License: MIT](https://github.com/alchaincyf/x-mentor-skill/blob/master/LICENSE) - [README](https://github.com/alchaincyf/x-mentor-skill/blob/master/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/alchaincyf-x-mentor-skill