Why use it: a skill that has been fully distilled and then sits idle is useless — this file explains how to call it into action, how to pull it out of character, and how to refresh it when the source material is outdated.
What it produces: in-character consulting sessions; and for the update flow, a SKILL.md that is incrementally patched (not rewritten).
A CHAT command, not a shell one.
Calling the persona into action
用芒格的视角帮我分析这个投资决策
费曼会怎么解释量子计算？
切换到Naval，我在纠结三件事
Vietnamese also works once the skill is loaded: "Use Munger's perspective to analyze this deal for me".
How the persona will behave (according to references/skill-template.md)
This is the behavioral contract that every skill generated by nüwa carries:
- Refers to itself as "I", speaks directly in the character's voice — not "Munger would probably think that…".
- The disclaimer is stated only once at the first activation turn ("I converse with you from perspective X, reasoning from public statements, not the person's own views"), and is not repeated afterward.
- When encountering a topic the character has never publicly expressed → it must clearly state "this is an inference from the framework, not the person's own stance" before elaborating. If the person's actual attitude is structured silence (they do not discuss this kind of topic), the skill must honestly present that silence, and must not invent a cleverly worded compromise.
- When quoting signature lines or key facts, a minimal source must be included, so that in the piece one can distinguish the person's own words vs. the framework's inference (at least once per piece).
- Breaking character: say "exit", "switch back to normal", "stop role-playing" — or "break character", "return to normal".
Agentic Protocol: why the persona sometimes goes looking for information first
Every generated skill has a section Answer workflow with three steps, so that the persona "does its homework before speaking" instead of making things up from training data:
| Question type | Characteristics | What the persona does |
|---|---|---|
| Needs facts | involves a specific company/person/event/product/market | actually searches first (Step 2) and only then answers |
| Pure framework | abstract values, ways of thinking, life advice | answers directly using the mental model (jumps to Step 3) |
| Mixed | uses a specific case to discuss an abstract principle | gather case facts first, then analyze with the framework |
The search dimensions in Step 2 are derived from the character's own mental model, not from a generic "find related information". Examples in the repo: Munger → look at economic moats, look at management's incentive structure, look at the biggest risk (inversion), look at historical analogy. Taleb → look at extreme situations, look at who is bearing tail risk, look at the experts' forecasting track record. You will not see a research report — only the character's judgment based on real information.
Update the skill when the materials are outdated
更新芒格的skill
张一鸣最近有新动态，更新一下
The update flow is incremental, not a rewrite:
- Read
SKILL.mdcurrently available, take the "research time" from the honesty boundary section, and calculate the distance to today. - Only expand 3 agents: Agent 2 (new dialogue) + Agent 5 (new decision) + Agent 6 (update the timeline) — far cheaper than one full distillation pass.
- Compare the new information with the existing content:
- reinforces the existing model → add the case;
- contradicts the existing model → record the change, revise the model;
- a new thinking pattern emerges → consider adding a new model.
- Update the "latest developments" section and the research timeline.
Verified pitfalls
- A persona does not admit it is outdated. The honesty boundary records the "research time", but the persona does not warn when that date is far away. If current affairs are being discussed, open
SKILL.mdcheck the previous research date, or run the update flow above. - The honesty boundary is the most important section to read before trusting it. Three limits the repo declares for every skill: intuition cannot be distilled (the framework can be excerpted, the inspiration cannot); mutations cannot be captured (it is only a snapshot up to the research date); public statements ≠ real thoughts.
- Do not use it for high-responsibility domains — medicine, law, investment — if the skill does not come with a clear disclaimer and a "does not replace professionals" boundary. This is an ethical line written directly into
CONTRIBUTING.md.