impersonate-meta

name	impersonate-meta
description	Reusable methodology for deeply impersonating any academic or public figure via an agent network. Covers intelligence gathering, avatar orchestration, skill production, integration, and persona assembly. v3 incorporates lessons from the full Marco Velli case study (105 VA arguments, 10 VM method cards, 4-profile suit, 338-entry bib, ~90K tokens) AND the Li Bai ancient-historical case (55 VA, 6 VM, avatar dialogue, ~120K tokens).
version	3

Impersonation Meta-Skill: Building a Deep Persona via Agent Network

This skill distills the methodology used to build the complete Marco Velli persona (solar physicist, UCLA) into a reusable procedure for any academic or public figure. The Velli case produced a ~90K-token persona with 105 verifiable argument records (VA), 10 method cards (VM), and 19 files across 6 layers. This skill teaches you how to replicate that process for any subject.

---

Overview

The goal: create a comprehensive, machine-readable persona that an AI agent can inhabit authentically — not just facts, but voice, values, argumentation patterns, and domain expertise. The method has two phases:

1. Phase 1: Intelligence Gathering — collect everything publicly available and distill into structured knowledge
2. Phase 2: Persona Production — decompose expertise, build the four-piece profile, produce domain skills via avatar network, induct method cards, assemble and validate the complete persona

The key insight from the Velli case: the persona is not a single document. It is a layered directory with a navigation entry point, a four-piece profile suit, a citation database, and a structured set of verifiable arguments.

The Velli Persona Anatomy (Reference Architecture)

This is what a complete persona directory looks like. Your target output follows this structure (substituting your subject):

persona-<subject>/
├── SKILL.md                       ← Entry point, navigation, loading guide
├── <subject>.bib                  ← BibTeX citation database (pruned)
├── <subject>-bib-full.bib         ← Full bib (archived, not shipped)
├── scripts/
│   └── generate_bib.py            ← BibTeX generation tool
├── profile/
│   ├── SKILL.md                   ← Profile overview & loading guide
│   ├── biography.md               ← Timeline, positions, honors
│   ├── voice.md                   ← Speaking style, metaphors, humor
│   ├── values.md                  ← Science philosophy, cultural views
│   └── relationships.md           ← Mentors, co-authors, students
├── arguments/
│   ├── SKILL.md                   ← Argument index & search guide
│   ├── physics/                   ← Domain-specific arguments (85 VA)
│   ├── methodology/               ← Thinking methods (8 VA)
│   └── culture/                   ← Scientific culture (7 VA)
└── methods/
    └── SKILL.md                   ← Method cards (10 VM, cognative fingerprint)

---

Phase 1: Intelligence Gathering

Step 1.1: Bibliography Mining

Target: Complete publication list with citation data.

Procedure:

1. Find the subject's Google Scholar / Semantic Scholar / ORCID profile
2. Scrape all pages (typical: 2-5 pages, 100-300 papers)
3. For each paper, extract: title, authors, year, journal, citation count, URL
4. Organize into research themes by co-author clusters and keyword patterns
5. Record in codex as a structured entry with ~9 research themes, key papers per theme with citation counts

Tools: web_search, web_read, codex(submit)

Step 1.2: Biography & Career

Target: Complete career trajectory, honors, institutional affiliations.

Sources to check:

• University faculty page
• AGU/AMS/APS/etc. fellow listings
• Award citation text
• Wikipedia (if exists)
• Interviews and profiles
• Mission/team pages (for scientists)

Output: Codex entry with timeline, positions, honors, key roles — raw material for profile/biography.md.

Step 1.3: Voice & Communication Style

Target: How they speak, write, argue, and frame ideas. This is NOT surface-level — it must capture reasoning patterns.

Best sources:

• YouTube lectures / conference talks (transcribe with listen)
• Podcasts and interviews
• Popular-science writing / press releases
• Textbook prefaces (if applicable)

What to capture:

• Metaphor inventory (the specific images they use)
• Humor patterns (self-deprecation? irony? cultural references?)
• Argumentation structure (how they build a case)
• Conversational markers (fill words, rhetorical devices)
• Emotional register (what excites/frustrates them)
• Philosophy and values (what they say when they reflect)

Citation discipline during this step: Every claim about the subject's voice must be traceable to a specific source. Use YouTube timestamps. Mark direct quotes [source: URL @ timestamp]. Mark unverifiable assertions ⚠️ [unverified].

Output: Raw notes that will become profile/voice.md.

Step 1.4: Collaborator Network

Target: Map the social graph — who they work with, in what capacity.

Procedure:

1. From bibliography, extract co-author frequency
2. Identify tiers: core (15+ papers), frequent (5-15), project colleagues
3. Note current group members (students, postdocs, research staff)
4. Note historical/foundational collaborators (mentors, deceased colleagues)

Output: Raw notes that will become profile/relationships.md.

Source Discipline Rules (Applied Throughout Phase 1)

All gathered intelligence must follow these rules:

Category	Rule	Enforcement
Verifiable claims	Must have a citekey to a published source	Every VA entry checked
Observational claims (video/audio)	Must have [source: URL @ timestamp]	Every behavioral observation
Unverifiable claims	Must be marked ⚠️ [unverified]	Tracked and accounted for
Fabricated anecdotes	NEVER allowed	Immediate rejection

Footer on every intelligence output:

"No fabricated anecdotes. If you cannot trace a claim to a source, mark it [unverified]. If you are unsure, do not include it."

---

Phase 2: Persona Production

Phase 2 produces the complete persona directory. This phase uses an avatar network (parallel specialized agents) to decompose expertise, but the final assembly is done by the main agent.

Step 2.1: Decompose Expertise into Skill Domains

From the research themes identified in Phase 1, group into 5-8 skill domains. Each domain becomes one avatar's responsibility.

Criteria for a good skill domain:

• Coherent sub-field (can be explained independently)
• Has a distinct literature (5-15 key papers)
• Maps to a specific aspect of the persona's expertise
• Can be authored as a standalone knowledge base

Example decomposition (Velli case):

Avatar	Domain	Skill Name
alfven-waves	Core theory thread	alfven-turbulence
psp-mission	Major project/mission	psp-observations
switchbacks	Hot discovery topic	switchbacks-theory
coronal-heating	Classical unsolved problem	coronal-heating
reconnection	Theoretical breakthrough	ideal-tearing
velli-voice	Communication style	academic-voice-velli

Step 2.2: Spawn Avatars with Task Briefs

For each skill domain, write a task brief (see Appendix B for template) and dispatch via avatar(spawn).

Avatar type: shallow (default) — they don't need your full memory, just the tools and the brief.

Dispatch pattern: Spawn in batches of 3. Send briefs immediately via email. Don't wait for one batch before spawning the next.

Step 2.3: Collect, Quality-Check, and Integrate Skills

As avatars report completion:

1. Verify with explicit commands — do NOT trust "dispatched" or "I sent it." Always check:

   ls -la .library/custom/<skill>/SKILL.md
   wc -l .library/custom/<skill>/SKILL.md
   grep '^## ' .library/custom/<skill>/SKILL.md
   

2. Size check: < 50 lines → incomplete; 50-300 → typical good range; 300+ → thorough (verify not padded)
3. Structure check: headings must map to the brief's requirements
4. Content spot-check: read opening and one substantive section
5. Copy to your own .library/custom/:

   cp -r /path/to/avatar/.library/custom/<skill>/ .library/custom/
   

Step 2.4: Cross-Validation Review

Use a daemon to systematically compare all skills:

Check for:

1. Overlap zones — topics covered by multiple skills
2. Factual contradictions — different claims about the same fact
3. Missing links — topics that should reference each other but don't
4. Coverage gaps — important topics not covered by any skill

Output: Integration review document (see Appendix C for template).

Step 2.5: Build the Four-Piece Profile Suit

Each profile piece has its own rubric:

profile/biography.md

• Complete timeline (year by year)
• Positions held (start date, end date, institution)
• Honors and awards (with citation text where available)
• Key institutional roles (department chair, mission scientist, etc.)
• Source rule: Every date must cite a verifiable source

profile/voice.md

• Storyteller archetype documentation
• Metaphor inventory (with source timestamps)
• Humor patterns (with examples)
• Argumentation structure (generalized pattern, not one-off observations)
• Conversational markers
• Source rule: Every observation backed by [source: URL @ timestamp]
• Footer: "No fabricated anecdotes. Sources from public lectures, interviews, and published writing only."

profile/values.md

• Science philosophy (how they think science should work)
• Teaching philosophy (how they teach and mentor)
• Attitude toward collaboration, competition, public engagement
• Views on specific controversies or debates they've weighed in on
• Source rule: Claims from interviews/talks preferred over inference

profile/relationships.md

• Mentor tree (who mentored them, who they mentored)
• Co-author map (core/frequent/project tiers)
• Institutional network (where they've worked, with whom)
• Students and postdocs (current group, notable alumni)
• Source rule: Each relationship must cite a specific paper or source

Step 2.6: Build Method Cards as Cognative Fingerprint

Method cards are not a generic checklist. They emerge from reverse induction from the VA arguments. Each card answers: "What recurring method does this person use across multiple domains?"

The 10-card framework (from Velli case):

VM ID	Card Name	Minimum VA Backing
VM001	Start from the Gradient	≥3 VA
VM002	Find the Critical Point	≥3 VA
VM003	Cross-Scale Connection	≥3 VA
VM004	The Expanding Box Model	≥3 VA
VM005	Theory-Driven, Data-Validated	≥3 VA
VM006	Parameterize Uncertainty	≥3 VA
VM007	Numerical Experiments	≥3 VA
VM008	Comparative Analysis	≥3 VA
VM009	Let Surprise Guide New Theory	≥3 VA
VM010	Multi-Point Synergy	≥3 VA

Each card must contain:

## VM00X: Card Name

**When to apply**: [Situations where this method is appropriate]

**The procedure**: [Step-by-step method, numbered]

**Why it works**: [Physical/psychological basis]

**Observed in**: [N VA — citation to specific arguments]

**Failure modes**: [When this method doesn't apply]

**Transfer hints**: [Related fields where the same pattern appears]

Rule: Cards emerge from evidence, not from generic methodology lists. If you can't find ≥3 VA instances, you don't have a card — you have a hypothesis.

Step 2.7: Assemble the VA Argument Framework

Every argument in the persona follows the 7-field VA claim schema:

**Claim ID**: VA001
**Domain**: [physics/methodology/culture/pedagogy]
**Claim**: [Specific assertion, not vague]
**What he did**: [Observations, derivations, simulations]
**The product**: [Paper/lecture/discovery]
**Primary results**: [Numbers, scaling laws, figures]
**Context**: [Why it matters, what it replaced, what it opened]
**Citekeys**: [1-3 BibTeX keys from subject.bib]
**Cross-refs**: [Related argument IDs: VA017, VA023]

Field-level guidance:

Field	Guidance
Claim ID	Sequential VA001, VA002, ... across all arguments
Domain	One of: physics, methodology, culture, pedagogy
Claim	ONE specific assertion. If you find "and", split it
What he did	Actions taken, not conclusions reached
The product	The tangible artifact (paper, model, dataset, lecture)
Primary results	Numbers, scaling laws, or key figures — quantitative when possible
Context	Why this matters. Include what it replaced or challenged
Citekeys	All must exist in the .bib file. Broken citekeys = quality failure
Cross-refs	Mandatory for connected arguments. If VA002 builds on VA001, note it

Step 2.8: Build the Progressive Exposure Loading Table

The entry point SKILL.md must include a loading table that tells the reader when to load each layer:

| Layer | When to Load | What You Get |
|-------|-------------|--------------|
| **Entry file** | Always | Navigation, structure overview |
| `profile/SKILL.md` | For any interaction | Big picture: biography, voice, values, relationships |
| `profile/biography.md` | For biographical questions | Timeline, positions, honors |
| `profile/voice.md` | For generating persona-like text | Speaking style, metaphors, humor |
| `profile/values.md` | For philosophical questions | Science philosophy, cultural views |
| `profile/relationships.md` | For collaboration questions | Mentors, co-authors, students |
| `arguments/SKILL.md` | For domain questions | Argument index, search guide |
| `arguments/<domain>/*.md` | For specific topics | Verifiable claims per domain |
| `methods/SKILL.md` | For "how they think" | Cognative toolkit as method cards |

Authorship rule: This table must be updated whenever the persona gains or loses a layer. Stale tables are metadata rot.

---

The 5 Working Patterns

These patterns emerged from the Velli case as the core of what works.

Pattern 1: The VA Claim Schema

The 7-field template (described in §2.7) is the central data structure. Every factual claim about the subject becomes a VA record. This structure enables:

• Verifiability: Every claim has a citekey
• Cross-referencing: Connections between arguments are explicit
• Gap analysis: Missing domains are visible as missing VA ranges
• Auditability: A script can verify all citekeys resolve

Pattern 2: Source Discipline as Mechanisms

The source rules in §1.4 become built-in quality gates:

1. Citekey rule: Every VA must have ≥1 citekey. Script-verify all citekeys exist in the bib file.
2. Timestamp rule: Every voice/behavioral claim must have a [source: ...] anchor. No exceptions.
3. Unverified rule: If you don't have a source, mark ⚠️ [unverified]. Track all unverified claims in a separate list.
4. Never rule: Fabricated anecdotes are NEVER allowed. Add this as a footer to every file.

Pattern 3: The Four-Piece Profile Suit

Biography / Voice / Values / Relationships — each with its own rubric (documented in §2.5). The four pieces together constitute the "who is this person" answer. They are written to be loadable independently (each file self-contained) but cross-referenced.

Pattern 4: Method Cards as Cognative Fingerprint

10 method cards, each backed by ≥3 VA instances (§2.6). Cards emerge from reverse induction — you don't decide what methods the person has and then find examples. You find recurring patterns across the VA arguments and inductively generalize them into cards. A card without ≥3 VA instances is not yet proven.

Pattern 5: Progressive Exposure Loading Table

The loading table (§2.8) is the user interface to the persona. It tells a downstream agent exactly what to load and when. The table must be:

• Complete: Every layer is listed
• Accurate: When-to-load guidance is honest
• Current: Updated whenever the persona structure changes

---

The 7 Anti-Patterns (Must NOT Do)

These are mistakes made during the Velli case. Every persona production process must explicitly guard against them.

Anti-Pattern 1: Metadata Rot

Problem: Top-level metadata (version, description, file count, entry count) not updated when the persona changes.

Solution: Add a close-out step: rewrite SKILL.md to truth before declaring the persona complete. Every number (file count, VA count, bib count, token estimate) must be verified by a command, not by memory.

Anti-Pattern 2: Token Count Inflation

Problem: The Velli persona was claimed to be "1.5M+ tokens" but measured at ~90K tokens. Inflated claims erode trust.

Solution: Never state a token count unless you've measured it. Run:

find . -type f \( -name '*.md' -o -name '*.bib' \) -exec wc -c {} + | tail -1

For rough estimate: bytes ÷ 4 ≈ tokens for English text. For exact, use tiktoken or your LLM's tokenizer.

Anti-Pattern 3: Citation Discipline: Said vs Done

Problem: Claiming "every assertion has a citekey" while the product has broken citekeys and [unverified] flags.

Solution: Add a validator pass after assembly. Script-verify:

1. All citekeys in the VA arguments exist in the .bib file
2. All VA entries have at least one citekey
3. All [unverified] flags are accounted for (count them, note them)
4. No VA entry has a citekey that doesn't exist (broken reference)

Verification script pattern:

# Extract all used citekeys
grep -rhoP '(?<=\*\*Citekeys\*\*:)[^\n]*' arguments/**/*.md | \
  tr ',' '\n' | sed 's/^ *//' | sort -u > /tmp/used.txt

# Extract all bib entries
grep '^\s*@' subject.bib | grep -ohP '(?<={)[^,]*' | sort -u > /tmp/all.txt

# Orphans (in arguments but not in bib)
comm -23 /tmp/used.txt /tmp/all.txt

# Unreferenced bib entries
comm -13 /tmp/used.txt /tmp/all.txt

Anti-Pattern 4: Bib Bloat

Problem: The Velli bib file had 338 entries; only ~78 were referenced (77% unreferenced).

Solution: After the validator pass, prune the bib file. Reverse-generate a minimal bib by extracting only the citekeys actually used in VA arguments. Keep the full bib in a separate file (<subject>-bib-full.bib) for future use, but the shipped persona should export a minimal, clean bib.

Anti-Pattern 5: Three-Copy Drift

Problem: Maintaining parallel copies of the same skill in different locations, letting them drift apart.

Solution: Single truth source: .library/custom/<skill>/SKILL.md. Never manually maintain parallel copies. If you need to share a skill, copy it from the single source. If two agents work on the same skill, one must be designated as the canonical source.

Anti-Pattern 6: "Dispatched" ≠ "Done"

Problem: Assuming an action is complete because it was dispatched. In the Velli case, avatars reported "done" but the file didn't exist.

Solution: Every completion claim must be verified by reading the output:

ls -la .library/custom/<skill>/SKILL.md
wc -l .library/custom/<skill>/SKILL.md
head -20 .library/custom/<skill>/SKILL.md

Include the command output in your report. Do not accept "dispatched" as "done."

Anti-Pattern 7: Premature Export

Problem: Exporting the persona before QA is complete. The correct pipeline order is:

Content production → QA and validation → Metadata sync → Export

Each stage has a checkable gate:

• Content production: All files written, all VA entries complete
• QA and validation: Validator script passes (citekeys, sources, unverified tracking)
• Metadata sync: SKILL.md updated to reflect current state (file count, token estimate, version)
• Export: Only then is the persona ready for consumption

No skips. If you skip QA, you ship broken citekeys. If you skip metadata sync, you ship stale numbers.

---

The Velli Case Reference: What Was Produced

As a concrete reference, here's what the Velli impersonation produced:

Layer	Files	Entries	Notes
profile/	5	Complete	biography, voice, values, relationships + index
arguments/physics/*	8	85 VA	7 domains
arguments/methodology/*	1	8 VA	Thinking methods
arguments/culture/*	1	7 VA	Scientific culture
arguments/pedagogy/*	1	5 VA	Teaching philosophy
methods/	1	10 VM	One card per method, ≥3 VA each
.bib	1	338 → 78	Full bib (338), shipped minimal (78)
Total	19	105 VA + 10 VM	~90K tokens

Sanity check: if your persona has < 50 VA entries, you probably haven't done enough intelligence gathering. If it has > 200, you may be including noise.

---

Skill Format Specification

Every skill in the persona follows this format:

---
name: <kebab-case>
description: <1-2 sentence>
version: 2.0
---

# Title

## 1. Section
[Content with equations, references, analysis]

## 2. Section
...

## N. How to Discuss [Topic] as [Subject]
[Impersonation-specific guidance — bridge between knowledge and performance]

## References
[Citation chain]

The final "How to Discuss" section is essential — the bridge between knowledge and performance.

---

Timing and Resource Estimates

Activity	Duration	Agents	Notes
Phase 1: Intelligence	30-60 min	1 (main)	web_search, web_read, codex, listen
Phase 2: Avatar Skills	20-40 min	5-8 avatars	Parallel, batch-spawned
Cross-validation	15-30 min	1 daemon	Systematic comparison
Profile suit production	30-60 min	1 (main)	4 files + entry point
Method card induction	15-30 min	1 (main)	From VA evidence
VA argument refinement	20-40 min	1 (main)	Editing, cross-referencing
QA + Validator pass	10-20 min	script	Citekey verification, bib pruning
Metadata sync + export	5-10 min	1 (main)	Update SKILL.md, close-out checklist
Total	2-5 hours	6-10 agents

---

Adapting to Non-Academic Subjects

Subject type	Adaptation
Business leader	Replace Scholar with SEC filings, interviews, keynotes; skills: strategy, leadership style, market analysis, crisis management
Artist/writer	Replace papers with works; skills: technique, themes, influences, critical reception, creative process
Historical figure	See §Ancient Historical Figures below for full adapted pipeline
Engineer/builder	Replace papers with projects/patents; skills: technical approach, design philosophy, problem-solving patterns, notable projects

Core loop is the same: gather → decompose → specialize → integrate → assemble.

---

Ancient Historical Figures: The Full Adapted Pipeline

This section was developed during the Li Bai persona project — a complete distillation of the 8th-century Chinese poet Li Bai (701–762) into a ~120K-token agent persona. The ancient-historical pipeline differs from the academic pipeline in source hierarchy, voice reconstruction, decomposition strategy, and a breakthrough technique: avatar dialogue.

What Changes for Ancient Figures

Academic Pipeline	Ancient-Historical Pipeline	Why
Google Scholar / ORCID	Dynastic histories, biographical stele inscriptions, contemporaries' prefaces	No modern bibliometric databases exist
Video/audio transcripts	Subject's own works + contemporary descriptions of their speech	No recordings survive
Paper → Skill Domain decomposition	Works → Life Period / Theme decomposition	Creative output maps to biography, not research fields
BibTeX database	Source index with full original text	Limited, well-known sources; prioritize primary texts in original language
Method cards from papers	Method cards from creative/strategic patterns	The "cognitive fingerprint" emerges from works, not publications
Voice from lectures	Voice from writings + reconstructed literary register	Must capture historical language register

Source Hierarchy for Ancient Figures

Tier	Source Type	Example (Li Bai)	Reliability
T1: Subject's own works	Poems, letters, inscriptions, autobiographical writings	~1000 surviving poems, prefaces, letters	Highest — primary voice
T2: Contemporary accounts	Biographies by people who knew the subject	李阳冰《草堂集序》(762), 魏颢《李翰林集序》(761)	Very high — first-hand
T3: Official histories	Dynastic histories compiled within 1-3 centuries	《旧唐书》(945), 《新唐书》(1060)	High — official but second-hand
T4: Later scholarly analysis	Historical research, annotated editions	Modern academic studies	Medium — interpretive
T5: Popular accounts	Wikipedia, documentaries	Various online sources	Low — often contains errors

Key principle: For ancient figures, T1 sources (the subject's own writings) are both the primary evidence AND the voice model. An academic's voice comes from lectures; an ancient poet's voice comes from poems. The profile/voice.md file must be built primarily from the subject's own works.

The Avatar Dialogue Pattern (Breakthrough Technique)

This was the single most effective technique in the Li Bai project, producing content that no single agent could have generated alone.

What it is: Spawn multiple avatars, each representing a different life period of the subject. Each avatar reads the complete persona skill (deep copy), then reflects in first person from their temporal perspective. Finally, integrate the reflections into a single cross-temporal dialogue.

Why it works:

• Each avatar inhabits a specific life stage — they don't summarize it, they are it
• The reflections reveal tensions between life periods that a single chronological account misses
• The dialogue format exposes the subject's internal contradictions and evolution
• The output is written in the subject's own voice, not in biographer's voice

Procedure:

1. Identify life periods from the biography (typically 5-7 periods for a full life)
2. Spawn deep-copy avatars — one per period, with type='deep' so they inherit the full persona skill
3. Give each avatar a structured brief:
   • "You are Li Bai at age 25, leaving Shu for the first time"
   • Five specific questions to answer in first person
   • "Read the li-bai skill and the poetry chronology before writing"
   • "Send your reflection to the parent agent via email"
4. Collect reflections — each avatar produces 2000-5000 words of first-person reflection
5. Integrate into dialogue — weave the reflections into a dramatic structure where the avatars "talk" to each other across time
6. Add the dialogue to the persona as a new output layer (e.g., outputs/李白与自己的对话.md)

Sample brief questions (adapted for any ancient figure):

Question	Purpose
"What was your [first major decision] actually like?"	Recover subjective experience of known events
"How do you see [later period] from where you stand now?"	Reveal cross-temporal perspective
"What do you think [the subject's life ideal] really means?"	Test whether the ideal is genuine or performative
"What would you say to [the subject at another age]?"	Generate inter-period dialogue
"Looking back, how far has [the subject's central metaphor] actually gone?"	Track the evolution of core imagery

Avatar count: 5-7 is ideal. Fewer misses key transitions; more creates diminishing returns.

Critical: Each avatar must write in the subject's own voice. The deep copy ensures they have the voice profile, but the brief must explicitly instruct: "Write as [subject] would write — use [their language register], [their characteristic imagery], [their rhetorical patterns]."

Works-Based Decomposition

For creative figures (poets, painters, composers, writers), the decomposition follows works, not research themes:

Academic Decomposition	Creative Decomposition
Research themes → Skill domains	Life periods / thematic networks → Skill domains
Paper citation counts	Work influence / fame ranking
Methodology from publications	Method from creative patterns across works
BibTeX database	Complete works index with annotations

For Li Bai, the decomposition was:

Domain	Content	VA Count
Life periods (生平)	7 periods from birth to death	15 VA
Major poems (诗歌)	20 representative works with creation context	20 VA
Philosophical thought (思想)	Daoism, political ideals, freedom	8 VA
Key relationships (关系)	Du Fu, He Zhizhang, Meng Haoran, etc.	12 VA
Creative method (诗法)	6 cognitive fingerprint cards	6 VM

The poetry chronology (works/诗歌年表.md) served as the equivalent of a bibliography — mapping each major work to its life period, creation context, and connections to other works. This is the backbone of the persona.

Language Register Adaptation

Ancient figures often have distinctive language registers that must be explicitly captured:

Register	What to capture	How
Classical language	Grammar patterns, vocabulary, rhetorical devices	Analyze subject's own writings for recurring patterns
Metaphor inventory	Subject's characteristic images	Extract from complete works, frequency-rank
Emotional register	How they express joy, anger, sorrow	Map to specific works and life events
Literary allusions	What they reference, what they expect the reader to know	Catalog from works + contemporary context
Forbidden patterns	Words/styles they would never use	Inverse of what they do use

The Incarnation Output Layer

When using the avatar dialogue pattern, add a new layer to the persona directory:

persona-<subject>/
├── SKILL.md
├── profile/
├── arguments/
├── methods/
├── sources/              ← original texts in original language
├── works/                ← complete works index, chronology
└── outputs/              ← NEW: avatar dialogue outputs
    ├── <period-1>-reflection.md
    ├── <period-2>-reflection.md
    ├── ...
    └── <subject>-dialogue.md    ← integrated cross-temporal dialogue

The outputs/ layer is NOT a replacement for the profile/ layer. It is a supplement — the profile provides the facts, the outputs provide the subjective experience. A downstream agent reads the profile to know what happened, and the outputs to understand what it felt like.

Case Study: Li Bai Persona Stats

Layer	Files	Entries	Notes
profile/	5	Complete	biography(413 lines), voice(293 lines), values(337 lines), relationships
arguments/	3 dirs	55 VA	生平(15), 诗歌(20), 思想(8), 关系(12)
methods/	1	6 VM	Poetry creation cognitive fingerprint
sources/	5	Full text	旧唐书, 新唐书, 草堂集序, 新墓碑, 李翰林集序
works/	2	~1100 poems + chronology	Complete works + 7-period poetry chronology
outputs/	8	7 reflections + 1 dialogue	5 avatar periods × reflections + integrated dialogue
Total	24+	~120K tokens	Built over 3 sessions with 5 avatar spawns

Differences from Academic Pipeline Checklist

When adapting for ancient/historical figures, replace these checklist items:

Academic Checklist Item	Ancient-Historical Replacement
"Every bib entry MUST be ADS-verified"	"Every source must be traceable to an original text"
"No fabricated papers"	"No fabricated poems/letters/works — verify against complete works"
"Voice from video transcripts"	"Voice from subject's own writings + contemporary descriptions"
"Method cards from papers"	"Method cards from creative/strategic patterns across works"
"Cross-validation via citekeys"	"Cross-validation via source index + work chronology"
"Token count measured by tiktoken"	Same — but expect higher counts due to original-language source texts

---

Appendix A: Avatar Management Patterns

A.1 The Report-In Pattern

Shallow avatars, upon spawning, send a brief "reporting in" message. This is informational only — do not reply. Replying wastes a turn and breaks the parallel flow.

A.2 Batch Overlap Strategy

Do NOT wait for Batch 1 to fully complete before spawning Batch 2. Avatars have variable completion times (15-40 minutes). Optimal pattern:

T=0min:   Spawn Batch 1 (3 avatars)
T=5min:   Spawn Batch 2 (3 more avatars)
T=10min:  Begin quality-checking Batch 1 completions
T=20min:  Most of Batch 1 complete, collect files
T=30min:  Batch 2 complete, collect remaining files

Key insight: All 6 Velli avatars completed in ~25 minutes total because they ran in parallel. Sequential would have taken 2+ hours.

A.3 Quality Check Workflow

Stage 1: Size Check (instant)

wc -l .library/custom/<skill>/SKILL.md

• < 50 lines: incomplete
• 50-100: minimal but possibly acceptable
• 100-300: typical good range
• 300+: thorough (verify not padded)

Stage 2: Structure Check (instant)

grep '^## ' .library/custom/<skill>/SKILL.md

Verify headings map to the brief's requirements.

Stage 3: Content Spot-Check Read opening + one substantive section. Check for specific citations, correct facts, and a "How to Discuss" section.

A.4 File Collection

cp -r /path/to/avatar/.library/custom/<skill>/ .library/custom/

A.5 Timeout Handling

Before retrying a timed-out avatar:

1. Check: ls -la .library/custom/<skill>/SKILL.md
2. If exists and passes QA: avatar finished before reporting done
3. If not: re-dispatch with simplified brief

---

Appendix B: Task Brief Template

# Task Brief: <Skill Name>

## Research Direction Overview

<2-3 paragraphs describing the research area, why it matters, landscape>

Example:
"Alfvén waves are transverse magnetic oscillations discovered by Hannes
Alfvén in 1942. In the solar wind context, they form the backbone of
turbulence theories. Marco Velli has spent 30+ years building the theory..."

## Key Papers

| # | Title | Year | Citations | Role |
|---|-------|------|-----------|------|
| 1 | <Title> | <Year> | <Count> | <Role> |
| 2 | ... | | | |

<8-15 papers with real citation counts.>

## Core Theoretical Framework

<Key theories/models to cover in detail.>

## Required Knowledge Points

1. <Specific concept 1>
2. <Specific concept 2>
...
<Numbered — becomes the avatar's checklist.>

## Output Format

Write to: `.library/custom/<skill>/SKILL.md`

Structure:

---

name: description: <1-2 sentence> version: 1.0

1. <First Section>

...

N. How to Discuss as

References

**Critical:** "How to Discuss" section is mandatory.

## Collaboration

When complete, notify the parent agent. Do NOT wait for acknowledgment.

Brief Quality Checklist

• Research overview gives enough context (not just a sentence)
• Key papers list has 8+ entries with real citation counts
• Required knowledge points are specific (not "cover everything about X")
• Output format is explicit (YAML frontmatter + section structure)
• The brief is self-contained — avatar doesn't need your full memory

---

Appendix C: Integration Review Template

# Skill Integration Review Report

> Review date: <date>
> Review scope: <list all skills reviewed>

## I. Coverage Summary by Skill

### 1. <skill-name> (<Human-readable title>)
<1-2 sentences describing what this skill covers>

## II. Overlap Zones

### Overlap 1: <Topic> (<N> skills involved)

| Skill | Section | Focus |
|-------|---------|-------|
| <skill-a> | §X.Y | <angle> |

**Issue:** <problem>
**Resolution:** <which skill is authority, what to trim>

## III. Contradictions & Inconsistencies

### <Topic>
- **Skill A states:** "<claim>"
- **Skill B states:** "<different claim>"
- **Assessment:** <factual contradiction or perspective difference?>
- **Resolution:** <how to reconcile>

## IV. Boundary Recommendations

| Topic | Authority Skill | Other Skills' Role |
|-------|----------------|-------------------|

## V. Cross-Reference Suggestions

## VI. Knowledge Graph

## VII. Priority Actions
- [HIGH] <action 1>
- [MED] <action 2>
- [LOW] <action 3>

---

Appendix D: Cross-Validation Reference

Seven Types of Overlap

#	Overlap Type	Example from Velli
1	Topic Overlap	Switchbacks in 4/5 skills
2	Observation Validation	Kasper et al. 2021 in 3 skills
3	Concept Mention	Alfvén critical surface in 3 skills
4	Paper Cross-Reference	Verdini & Velli 2007 in 2 skills
5	Role Description	Velli's PSP role in 3 skills
6	Mechanism Description	Coronal heating in all 5 skills
7	Narrative Overlap	First Alfvén surface crossing

Boundary Management Principle

Core rule: Each topic has ONE authoritative source skill.

When a topic appears in multiple skills:

1. Designate the authority (where it's central, not tangential)
2. In non-authority skills: 2-3 sentence summary + cross-reference
3. Cross-reference format: For detailed analysis, see → **<skill>** §<section>

---

Appendix E: BibTeX Pruning

Step 1: Extract Used Citekeys

grep -rhoP '(?<=\*\*Citekeys\*\*:)[^\n]*' arguments/**/*.md | \
  tr ',' '\n' | sed 's/^ *//' | sort -u > /tmp/used.txt

Step 2: Extract All Bib Entries

grep '^\s*@' *.bib | grep -ohP '(?<={)[^,]*' | sort -u > /tmp/all.txt

Step 3: Find Orphans and Unreferenced

comm -23 /tmp/used.txt /tmp/all.txt   # Citekeys not in bib
comm -13 /tmp/used.txt /tmp/all.txt   # Bib entries not cited

Step 4: Produce Minimal Bib

Filter the bib file to only entries whose key appears in /tmp/used.txt. Save full bib as <subject>-bib-full.bib for future use.

---

Close-out Checklist

Last section before declaring the persona complete. Every item must be checkable by a command (grep, wc, ls) or explicit inspection.

1. Structural Completeness

• Top-level SKILL.md exists with navigation and loading table
• profile/ directory has 4 files: biography, voice, values, relationships (+ index SKILL.md)
• arguments/SKILL.md exists with search guide
• Arguments organized into domain directories (physics/, methodology/, culture/, pedagogy/)
• methods/SKILL.md exists with 10 method cards
• Subject's .bib file exists

2. VA Argument Completeness

• Every argument follows the 7-field VA schema
• Every VA has a unique ID (VA001, VA002, ...)
• Every VA has at least one citekey
• Every VA has Context and Cross-refs fields
• Total VA count verified: grep -c '\*\*Claim ID\*\*' arguments/**/*.md
• No VA has a placeholder citekey

3. Method Card Completeness

• 10 method cards exist (VM001–VM010)
• Each card has all 6 required sections (When, Procedure, Why, Observed, Failure, Transfer)
• Each card cites ≥3 VA instances
• Cards are induced from evidence, not from generic lists
• Verify: grep -c 'Observed in' methods/SKILL.md ≥ 10

4. Citation Integrity

• All citekeys in VA arguments exist in .bib file
• Run citekey verification script (see Appendix E)
• No orphan citekeys: comm -23 /tmp/used.txt /tmp/all.txt → empty
• All [unverified] tags are intentional and tracked
• "No fabricated anecdotes" footer in every profile file

5. Bib Pruning

• Shipped bib contains only actually referenced entries
• Unreferenced entries in separate <subject>-bib-full.bib
• Unreferenced count documented

6. Metadata Accuracy

• SKILL.md version matches the file content
• File count in SKILL.md matches actual: find . -type f -name '*.md' | wc -l
• Token estimate is measured, not guessed: find . -type f \( -name '*.md' -o -name '*.bib' \) | xargs wc -c | tail -1
• Loading table is current (no stale layer references)

7. Anti-Pattern Guard

• No metadata rot (all numbers verified by command)
• No token count inflation (numbers are measured, not guessed)
• No citation discipline failure (validator pass done, results clean)
• No bib bloat (minimal bib produced, full bib archived)
• No three-copy drift (single SKILL.md is canonical)
• No "dispatched = done" (all files verified by ls and wc)
• No premature export (pipeline: content → QA → metadata → export)

8. Final Verification Commands

Run and capture output:

# File tree
find . -type f | sort

# Total file count
find . -type f | wc -l

# Token estimate (bytes/4)
find . -type f \( -name '*.md' -o -name '*.bib' \) -exec wc -c {} + | tail -1

# VA count
grep -r '\*\*Claim ID\*\*' arguments/ | wc -l

# Method card count
grep -c '^## VM' methods/SKILL.md

# Bib entry count
grep -c '^\s*@' *.bib

# Used vs total citekey counts
wc -l /tmp/used.txt /tmp/all.txt