// the find
alchaincyf/nuwa-skill
你想蒸馏的下一个员工，何必是同事。蒸馏任何人的思维方式——心智模型、决策启发式、表达DNA。Distill how anyone thinks.
This is an Agent Skills prompt file (SKILL.md) and the pipeline that produces it. It turns a public figure's writings, interviews, and decisions into a perspective skill made of mental models, decision heuristics, and expression style, so an LLM can reason through a problem with a named thinker's framework. It is for developers who want that kind of framework-driven analysis rather than quotes.
- The research is auditable. Each example keeps its raw research files (writings, conversations, external critics, timeline), so a reader can trace a claimed mental model back to sources instead of taking the SKILL.md on faith.
- The inclusion rule for a mental model is written down and testable: it must recur across two or more domains, predict the person's stance on a new question, and not be something every smart person would say. That is a stricter filter than most persona prompts use.
- The limits section is specific. It says intuition does not transfer, the snapshot stops at the research date, and public statements are not private views. A persona skill that states its own boundaries is easier to trust.
- Installation is plain markdown with YAML frontmatter in the open Agent Skills format. There is no runtime or dependency, and it works in any compatible agent or pasted straight into a chat.
- The core asset is a prompt, and nothing in the repo measures whether it works beyond the author's own scorecard. The A-grade fidelity scores for all 15 skills come from a dual-agent blind test described as independent, but the README does not say the judging agents are a different model family from the ones that wrote the skills, so shared bias is not ruled out.
- The English README is out of date. It says 13 person skills, while the Chinese section and the examples/ directory list 14. It is a small error, but it suggests the translations are not kept in sync and English readers get the thinner version.
- The contribution rules close the main improvement path. SKILL.md does not accept external PRs; ideas go to issues and the maintainer reimplements them. For a repo whose value is its methodology, quality control sits with one person.
- Output quality depends on the underlying model, and the README never says so. The same SKILL.md may produce a sharper or blander persona on different runtimes. The demo transcripts are chosen examples, with no failure cases or cross-model comparisons.