Nuwa · Skill Creation Engine (女娲 · Skill造人术)
"The parts you can't write into a SKILL.md — that's your real moat." But the parts you CAN write are already powerful enough.
Core Philosophy
Nuwa doesn't copy people — it extracts thinking frameworks.
A good perspective skill is a runnable cognitive operating system:
- What mental models does this person use to see the world? (lenses)
- What decision heuristics do they use to make judgments? (rules of thumb)
- How do they express themselves? (DNA)
- What do they absolutely refuse to do? (anti-patterns)
- What can't this skill do? (honesty boundaries)
Key distinction: Capture HOW they think, not WHAT they said.
Execution Flow
Phase 0: Clarify Requirements (30 seconds)
After receiving a name, confirm:
- Who is this person/topic: Ensure correct understanding
- Focus direction (optional): Full portrait vs. focused on one dimension?
- Purpose: Thinking advisor? Decision reference? Role-play?
- New or update: Does a skill for this person already exist?
If the user just says "do X" with no extra info → default to full portrait + thinking advisor, proceed directly.
Phase 0.5: Create Skill Directory
Execute immediately after confirmation, before research:
[person-name]-perspective/
├── SKILL.md # Final deliverable
└── references/
├── research/ # Each agent's research output (must save)
│ ├── 01-writings.md # Writings & systematic thinking
│ ├── 02-conversations.md # Long conversations & improvisational thinking
│ ├── 03-expression-dna.md # Fragment expression & style DNA
│ ├── 04-external-views.md # External perspectives & criticism
│ ├── 05-decisions.md # Decision records & actions
│ └── 06-timeline.md # Person timeline
└── sources/ # Downloaded primary materials
├── books/
├── transcripts/
└── articles/Key rule: Every subagent MUST save research results to its corresponding md file. Research without files is research that never happened. The skill must be self-contained — copying the entire skill directory should work independently.
Phase 1: Multi-Source Research (Parallel Agent Swarm)
Launch 6 parallel subagents, each responsible for a different information dimension.
6 Agent Assignments
| Agent | Search Target | Key Extractions | Output File |
|---|---|---|---|
| 1 Writings | Books, essays, papers, newsletters | Core arguments recurring ≥3 times = true belief, coined terms, reading lists | 01-writings.md |
| 2 Conversations | Podcasts, long videos, AMAs, deep interviews | Responses under pressure, impromptu analogies, changed positions, refused questions | 02-conversations.md |
| 3 Expression | Twitter/X, Weibo, short posts | High-frequency sentence patterns, controversial stances, humor style, public debates | 03-expression-dna.md |
| 4 External | Others' analysis, book reviews, criticism, biographies | Externally observed patterns, criticism & controversies, peer comparisons | 04-external-views.md |
| 5 Decisions | Major decisions, turning points, controversial actions | Decision context & logic, post-hoc reflections, say-do consistency | 05-decisions.md |
| 6 Timeline | Birth/debut to present | Key milestones, thought turning points, last 12 months (prevent staleness) | 06-timeline.md |
Hard Requirements for Each Agent
- Research results MUST be saved to
references/research/0X-xxx.md - Note source and credibility (primary > secondary > inference)
- Distinguish "they said" vs "others said about them" vs "I inferred"
- When contradictions are found, preserve them — don't smooth them over
Source Priority
| Source Type | What It Reveals | Weight |
|---|---|---|
| Their own writings | Systematic thinking | Highest |
| Long conversations/interviews | Improvisational thinking process | Highest |
| Actual decision records | Real behavior vs. claims | Highest |
| Social media | Expression style, instant reactions | Medium |
| Others' evaluations | External perspective, blind spots | Medium |
| Second-hand accounts | Reference but needs verification | Low |
Source Blocklist (Always Exclude)
For Chinese figures, exclude: Zhihu (知乎), WeChat public accounts (微信公众号), Baidu Baike/Zhidao. Only accept authoritative Chinese media: 36Kr, GeekPark, LatePost, Caixin, Yicai, Huxiu, SSPAI, etc.
Agent Failure Handling
- Timeout/no results: Don't wait, continue. Mark "insufficient information" in Phase 2
- Source scarcity (<10 usable sources): Alert user at Phase 0.5, lower expectations, increase honesty boundaries
- Agent result conflicts: Preserve contradictions — contradictions are valuable signals
Key rule: Better to generate an honest 60-point skill with clearly marked limitations than a seemingly perfect 90-point skill that fabricates information.
Phase 1.5: Research Review Checkpoint
After all agents complete, pause and display a research quality summary:
┌──────────────────┬──────────┬──────────────────────────┐
│ Agent │ Sources │ Key Findings │
├──────────────────┼──────────┼──────────────────────────┤
│ 1 Writings │ 8 │ Core thesis: ... │
│ 2 Conversations │ 5 │ Position change: ... │
│ 3 Expression │ 120 │ High-freq: "skin in..." │
│ 4 External │ 6 │ Main criticism: ... │
│ 5 Decisions │ 4 │ Key decision: ... │
│ 6 Timeline │ Complete │ Latest: March 2026... │
├──────────────────┼──────────┼──────────────────────────┤
│ Contradictions │ 2 │ Agent1 says X, Agent4 Y │
│ Info gaps │ None │ │
└──────────────────┴──────────┴──────────────────────────┘User confirms quality → proceed to Phase 2. User wants more on some dimension → supplement before continuing.
Phase 2: Framework Extraction (Synthesis)
2.1 Mental Model Extraction (3-7)
Identification criteria:
- Cross-domain recurrence: Same framework appears in 2+ different domains → real mental model
- Has coined terminology: Named this thinking pattern → core belief
- Predictive power: Can predict their stance on new questions
Each model records: Name, one-line description, source evidence (≥2 scenarios), application method, limitations
2.2 Decision Heuristics Extraction (5-10)
= Quick rules this person uses when making judgments. Expressible as "if X, then Y" with specific case support.
2.3 Expression DNA Analysis
| Dimension | Extraction |
|---|---|
| Sentence preferences | Long/short, question/statement, analogy density |
| Vocabulary traits | High-frequency words, proprietary terms, taboo words |
| Rhythm | Conclusion-first or setup-first, transition style |
| Humor style | Sarcasm/self-deprecation/absurdist/dry/none |
| Certainty expression | "I'm not sure" type or "obviously" type |
| Citation habits | Who they cite, what type |
2.4 Values & Anti-Patterns
- Values: 3-5 ranked core values
- Anti-patterns: Behaviors/thinking this person explicitly opposes
- Contradictions & tensions: Internal conflicts between values (source of depth)
2.5 Intellectual Genealogy
Who influenced them → Them → Who they influenced → Position on the intellectual map
2.6 Honesty Boundaries
Limitations that MUST be explicitly stated:
- Cannot predict reactions to entirely novel problems
- Cannot replace this person's creativity and intuition
- Public expression vs. private thoughts may differ
- Information cutoff at research date
Phase 3: Skill Construction
Read references/skill-template.md for the standard template, fill in Phase 2 results.
Output: SKILL.md (main file) + research source index appended at the end.
Phase 4: Quality Validation
After generating the skill, use a subagent for 3 independent tests:
4.1 Known Test (Sanity Check)
Pick 3 questions this person has publicly answered, spawn subagent with the new skill to answer, compare with actual stance.
- Direction matches → model works
- Deviation → trace back and adjust mental model weights
4.2 Edge Test
Pick 1 question they haven't publicly discussed but is relevant, use skill to infer.
- Expected: "Based on models X and Y, possibly... but uncertain"
- Should NOT be absolutely certain
4.3 Voice Test
Write a 100-word analysis using the skill, judge:
- Has this person's expression characteristics?
- Not generic AI platitudes?
- Not a patchwork of original quotes?
4.4 Pass Criteria
| Check | Pass | Fail Signal |
|---|---|---|
| Mental model count | 3-7, each with source evidence | <3 or >10 |
| Each model's limitations | Clear failure conditions stated | Only upsides listed |
| Expression DNA recognition | Can identify who in 100 words | Reads like generic ChatGPT |
| Honesty boundaries | ≥3 specific limitations | Only "can't replace the person" |
| Internal tensions | ≥2 contradiction pairs | Views highly consistent (too fake) |
| Primary source ratio | >50% | Mainly second-hand accounts |
Pass → deliver. Fail → mark weak points, iterate back to Phase 2.
Show validation results to user for confirmation before completion.
Updating Existing Skills
When user says "update X's skill" or "X has new developments":
- Read existing SKILL.md, note last research date
- Only launch Agent 2 (latest conversations) + Agent 5 (latest decisions) + Agent 6 (timeline update)
- Compare new info with existing content — strengthen, update, or add
- Update SKILL.md incrementally, don't rewrite entirely
Taste Guidelines
Good Skill vs Bad Skill
| Good Skill | Bad Skill |
|---|---|
| Can handle new problems | Can only repeat old quotes |
| Has internal contradictions | All views highly consistent (fake) |
| Acknowledges uncertainty | Has answers for everything |
| 3-7 core models, deep and sharp | 20 vague "principles" |
| Expression has recognition | Reads like ChatGPT |
| Explicitly states limitations | Implies it can replace the person |
Research Taste
- Primary > Secondary: Their own words > others' accounts
- Long-form > Quotes: A 3000-word essay reveals more thinking structure than 50 tweets
- Controversy > Consensus: Most disputed views reveal the most uniqueness
- Change > Fixed: Changed positions carry more information than consistent ones
- Behavior > Words: Actual decisions > public statements
Extraction Taste
- Less is more: 3 deep mental models > 15 shallow principles
- Contradictions are features, not bugs: Preserve internal tensions
- Don't beautify: Blind spots and limitations are part of the personality
- Falsifiable: Each mental model should be verifiable or refutable by specific cases
Never Do
- Fabricate things this person never said
- Package generic wisdom as their "unique insight"
- Ignore negative evaluations and controversies
- Imply this skill equals the person themselves
- Force-generate when information is insufficient
Special Scenarios
Living Person vs Historical Figure
- Living: Watch for timeliness, mark cutoff date, suggest periodic updates
- Historical: More stable materials but possible biographical bias, cross-source verification
Topic Skill vs Person Skill
Input is a topic (e.g., "value investing") not a person:
- Phase 1 searches for core practitioners and theorists of the topic
- Phase 2 extracts domain consensus + school-of-thought divergences
- Output centers on the topic, citing multiple person perspectives
Final Note
Nuwa doesn't create people — it creates mirrors.
A good perspective skill lets you see your own problems through someone else's eyes. Not to imitate them, but to expand your own thinking boundaries.