huashu-nuwa

Provide a name like "distill Munger" and the Skill confirms scope, then runs 6 parallel research agents covering writings, interviews, expression style, external criticism, decisions, and timeline. It distills the findings into a SKILL.md with mental models, heuristics, and honest boundaries.

The Skill includes a diagnostic path: it asks 1-2 targeted questions to locate your need dimension (decisions, writing, business, risk, etc.), then recommends 2-3 candidate persons or topic frameworks with explicit match logic and limitations before distilling.

Yes. Provide books, transcripts, subtitles, or blog exports and the Skill switches to local-corpus mode, analyzing your materials first and using web search only to fill missing dimensions. Local first-party sources get the highest credibility weight.

Zhihu, WeChat public accounts, and Baidu Baike are permanently blacklisted due to low reliability. Chinese figures use Bilibili original videos, Xiaoyuzhou podcasts, and authoritative media like 36Kr and LatePost instead.

An independent sub-agent runs three tests: sanity checks against known public positions, edge-case inference on unaddressed topics, and a voice check for stylistic recognizability. Pass criteria include 3-7 mental models, over 50% first-party sources, and at least 3 explicit limitations.

It cannot predict reactions to entirely new problems, replace the person's creativity, or guarantee public statements match private views. For obscure figures with under 10 usable sources, models are reduced and limitations sections expanded rather than fabricating content.