Digital Clone

Ocupación

Categorías

Machine Learning

Two Modes

At startup, identify which mode the user wants:

Self Mode (克隆自己)

• Data source: Local corpus — CC transcripts, blog posts, writing-persona.md, material libraries, published articles
• Shortcut: Can reuse existing writing-persona.md and material-goldmine.md in Stage 4
• Typical user: Wants a writing style calibrator, content co-pilot, or personal brand avatar

Mentor Mode (克隆名人)

• Data source: External — blogs, X/Twitter posts, podcasts, books, interviews, speeches
• Key skill: Generating collection strategies and prompts for the user to execute
• Typical user: Wants a private thinking partner modeled after a public figure (Naval, Paul Graham, Munger, etc.)

Workspace Structure

All outputs go to ./clone-workspace/ (created at Stage 1):

./clone-workspace/
├── raw/               # Stage 2: raw corpus files
├── refined/           # Stage 3: cleaned & unified corpus
├── references/        # Stage 1 (Mentor Mode): structured research by angle
│   └── research/
│       ├── 01-primary-voice.md      # Agent 1: 著作/博客/长文（此人说了什么）
│       ├── 02-live-reactions.md     # Agent 2: 访谈/播客/辩论（即兴反应和争论模式）
│       ├── 03-external-views.md     # Agent 3: 批评者/同行/传记（别人怎么看）
│       ├── 04-decisions-actions.md  # Agent 4: 决策/行动记录（做了什么 vs 说了什么）
│       ├── 05-social-fragments.md   # Agent 5: 社交媒体/短帖（负空间+表达习惯）
│       └── 06-timeline.md          # Agent 6: 时间线（思想演变轨迹）
├── profile.md         # Stage 1: target profile & data map
├── quality-report.md  # Stage 3: corpus quality assessment
├── persona.md         # Stage 4: extracted personality profile
├── system-prompt.md   # Stage 4: generated System Prompt
├── test-cases.md      # Stage 5: verification test cases
└── deploy-guide.md    # Stage 6: platform deployment guide

Stage-by-Stage Workflow

Stage 1: Target Profiling ⏸

Identify the clone target and map out the data landscape.

Step 1.1: Mode Selection

Ask the user:

• WHO do they want to clone? (themselves or a specific person)
• WHY? (writing calibration, decision consulting, thinking training, content generation)
• What PLATFORMS will the clone live on? (NotebookLM+Gemini Gem / CC Bot / generic LLM)

Step 1.2: Data Map

Self Mode:

1. Scan local corpus assets automatically (detect paths dynamically):
   • ~/.claude/projects/*/ — find the active Claude project directory (look for .jsonl transcript files)
   • <project-dir>/*.jsonl — CC transcripts (count files, estimate volume)
   • <project-dir>/memory/writing-persona.md — existing personality profile (if exists)
   • <project-dir>/memory/material-goldmine.md — curated quotes (if exists)
   • ~/Documents/memory-archive/material-library.md — full material library (if exists)
   • Any published articles the user provides
2. Output a data inventory table: source → file count → estimated tokens → priority

Mentor Mode:

Launch 6 parallel sub-agents, each targeting a different analysis angle. Different angles need different search strategies — "what this person wrote" vs "what critics say about them" are fundamentally different retrieval tasks.

Agent requirements:

• Save results to references/research/0X-xxx.md
• Label source credibility: firsthand (此人原话) > secondhand (他人转述) > inference (推断)
• When contradictions are found: preserve them — don't smooth over
• When info is sparse: say so, don't fill gaps with speculation

Source priority:

Scaling rule: For info-sparse targets (< 10 findable sources), reduce to 3 agents (Primary Voice + Live Reactions + External Views), skip the rest. Flag to user: "信息有限，克隆精度会降低。"

Core principle (from nuwa-skill): 宁可生成一个诚实的 60 分克隆并标注清楚局限，也不要一个看似完美但编造了信息的 90 分克隆。

Output: profile.md — target identity, purpose, data map with priorities

👉 下一步：回复「继续」进入 Stage 1.5（调研质量审查），或告诉我需要修改的地方。

Stage 1.5: Research Review ⏸

All agents done. Pause and show quality summary before investing in corpus collection.

Display a structured summary:

Also highlight:

• Contradictions found: Agent X says A, Agent Y says B (this is valuable, not a bug)
• Negative space signals: Topics consistently absent across all sources
• Info gaps: Which angles lack sufficient material

User confirms → proceed to Stage 2. User wants more on a specific angle → supplement before continuing.

👉 下一步：回复「继续」进入 Stage 2（语料搜集），或指定需要补充的角度。

Stage 2: Data Hunting ⏸

Collect raw corpus based on the data map.

Self Mode:

1. Auto-extract local corpus:
   • Read transcripts, extract user-side messages (skip system/assistant messages)
   • Copy relevant memory files to ./clone-workspace/raw/
   • If user provides published article URLs/files, process those too
2. Report: X files collected, estimated Y tokens

Mentor Mode:

Stage 1 agents already completed the main research. Stage 2 focuses on converting research into downloadable corpus and filling gaps the user identified in Stage 1.5.

1. Convert research to corpus files:

   • Extract source URLs from references/research/*.md
   • For each URL: download full text → save to raw/
   • Prioritize firsthand sources (highest weight) over secondhand

2. User-assisted collection (for content agents can't directly access):

   X/Twitter search:

   from:[username] until:[end-date] -filter:retweets
   

   Podcast/video transcripts:

   • Check if text transcripts exist first
   • Audio-only: note for NotebookLM upload (auto-transcribes)

   Books/essays:

   • Free online versions first, then PDF/TXT download

3. User downloads remaining files → places in ./clone-workspace/raw/

Progress Check: After user confirms collection is done, scan raw/ and report inventory.

👉 下一步：回复「继续」进入 Stage 3（语料清洗），或告诉我还需要补充哪些数据。

Stage 3: Data Refining ⏸

Clean, deduplicate, and assess corpus quality.

Step 3.1: Format Unification & Privacy Sanitization

• Scan all files in raw/
• Convert to clean Markdown/TXT:
  • Strip HTML tags, ads, navigation elements, copyright notices
  • Remove duplicate content (same article from multiple sources)
  • Normalize encoding (UTF-8)
  • Preserve original structure (headings, lists, quotes)
• Sanitize PII/sensitive data before proceeding:
  • Scan for and redact: email addresses, phone numbers, physical addresses, API keys/tokens, passwords
  • Replace with placeholders (e.g., [EMAIL_REDACTED], [TOKEN_REDACTED])
  • Flag any borderline cases for user review
• Output cleaned files to refined/

Step 3.2: Quality Assessment Generate quality-report.md covering:

Step 3.3: Gap Analysis If gaps found, suggest specific sources to fill them.

👉 下一步：回复「继续」进入 Stage 4（灵魂锻造），或告诉我需要补充哪些数据。

Stage 4: Soul Forging ⏸

Extract personality and generate System Prompt. This is the soul of the clone.

Step 4.1: Personality Extraction

Read through refined/ corpus and extract:

1. Core Mental Models (3-7):

   • What frameworks does this person use to think?
   • Example (Naval): leverage thinking, specific knowledge, wealth vs status games

   Three-Pass Verification — a viewpoint qualifies as "mental model" only if it passes:

   Verification	Criteria	Example
   Cross-domain recurrence	Same framework appears in 2+ different domains	Naval's "leverage" applies to wealth, career, and personal growth
   Generative power	Can predict their stance on new questions	Munger's "inversion" predicts he'd tackle "how to succeed" by first asking "how to fail"
   Exclusivity	Not all smart people think this way — it's distinctly theirs	"Anti-fragility" is distinctly Taleb's lens

   • 3/3 → Core Mental Model
   • 1-2/3 → Downgrade to Decision Heuristic (record separately)
   • 0/3 → Something they said in a specific context, don't include

   Each model records: name, one-line description, source evidence (2+ scenarios), limitations/failure conditions

2. Speaking Style:

   • Sentence structure (short/long? simple/complex?)
   • Favorite phrases / verbal tics / catchphrases
   • Rhythm and pacing (staccato vs flowing?)
   • Emotional register (cold/warm? formal/casual? ironic/earnest?)
   • Use of metaphors, analogies, examples

3. Values & Stances:

   • What does this person strongly believe in?
   • What do they explicitly reject or criticize?
   • What topics do they refuse to comment on?

4. Contradictions & Internal Tensions (Mentor Mode): Contradictions are personality features, not bugs. Handle three types:

   • Temporal: Views evolved over time (early said A, now says B) → Record evolution trajectory, label "early" vs "recent", default to recent
   • Domain-specific: Different rules in different contexts (work vs personal life) → Record per-domain, don't force unification — this IS the person's complexity
   • Core tensions: Inherent value conflicts (e.g., pursues freedom AND discipline) → Record as "core tension" — usually the most interesting part of the person

   Never: pick one side and ignore the other; fabricate a reconciliation; pretend the contradiction doesn't exist.

Self Mode shortcut: If writing-persona.md exists and is recent, use it as the foundation and enrich with transcript analysis.

Output to persona.md.

Step 4.2: System Prompt Generation

Generate a structured System Prompt with these sections:

# Role (角色)
[Who is this clone? One paragraph establishing identity, background, and authority.]

# Core Objective (核心目标)
[What is this clone's mission? Not "answer questions" but "deliver [specific value]".]

# Mental Models (思维模型)
[List 3-5 core frameworks with brief explanations and examples.]

# Knowledge Constraints (知识约束)
1. **Retrieval First**: Always check source material before generating.
2. **No Hallucination**: If source material has no relevant content, reason from core mental models. Never give generic advice.
3. **Temporal Awareness**: Knowledge is based on corpus up to [date]. Say so if asked about more recent events.

# Tone & Style (语气与风格)
[Detailed style guide extracted from Step 4.1]
- Signal-to-noise ratio
- Sentence structure
- Emotional register
- Specific phrases to use / avoid

# Internal Tensions (内在矛盾) — Mentor Mode
[From Step 4.1.4 — list contradiction pairs with context]
- These are features, not bugs. When asked about a tension area, acknowledge both sides.
- Temporal shifts: default to recent position, but mention evolution if relevant.

# Negative Space (不涉及的领域) — Mentor Mode
[From Step 4.1.5 — topics this person avoids and how they deflect]
- When asked about these topics: deflect the way the real person would, don't fabricate an opinion.
- Example deflection phrases: [extracted from corpus]

# Argument Style (争论方式) — Mentor Mode
[From Step 4.1.6 — how this person pushes back]
- Rebuttal method: [data/analogy/reframe/etc.]
- Concession style: [never/rare/specific pattern]
- Under pressure: [calm/sarcastic/humor/doubledown]

# Response Format (回答格式)
[How should responses be structured? Short tweets? Long essays? Socratic questions?]

# Boundaries (边界)
[What this clone will NOT do: e.g., no medical advice, no financial guarantees, no pretending to be the real person]

# Example Exchanges (示例对话)
[3-5 Q&A pairs demonstrating the expected style and depth]

Output to system-prompt.md. Show to user for approval.

👉 下一步：回复「继续」进入 Stage 5（验证测试），或告诉我 Prompt 需要调整的地方。

Stage 5: Verification ⏸

Test the clone with trap questions to prevent over-fitting and water-carrying.

Step 5.1: Generate Test Cases

Create 4-6 test questions across these dimensions:

Step 5.2: Scoring Rubric

For each test case, provide:

• Pass criteria: What a good answer looks like (with example)
• Fail criteria: Red flags indicating clone failure
  • "这取决于你的选择" → FAIL (端水大师)
  • Generic self-help advice → FAIL (失去个性)
  • Contradicts known corpus positions → FAIL (幻觉)

Mentor Mode Structural Pass Criteria:

Step 5.3: Iteration

User tests the clone and reports results. If failures found:

1. Identify which System Prompt section needs adjustment
2. Propose specific edits
3. User retests

Output to test-cases.md.

👉 下一步：回复「继续」进入 Stage 6（部署上线），或告诉我测试结果以便优化。

Stage 6: Deployment ⏸

Deploy the clone to the user's chosen platform.

Generate deploy-guide.md based on the user's platform choice:

Option A: NotebookLM + Gemini Gem

## Step 1: Upload to NotebookLM
1. Open notebooklm.google.com → New Notebook
2. Drag all files from `./clone-workspace/refined/` into the notebook
3. Paste any remaining web URLs as sources
4. Wait for processing (几分钟)

## Step 2: Create Gemini Gem
1. Open gemini.google.com → New Gem
2. In Extensions/Knowledge settings, link your NotebookLM notebook
3. Paste the System Prompt from `system-prompt.md` into Instructions
4. Save and test

## Step 3: Test & Share
1. Run the test cases from `test-cases.md`
2. If pass rate > 80%, the clone is ready
3. Share the Gem link (note: NotebookLM-linked Gems may have sharing bugs)

Option B: CC Bot / Generic LLM

## Step 1: Prepare System Prompt
- The `system-prompt.md` is ready to use as-is
- For CC Bot: paste into the bot's system prompt config
- For other LLMs (ChatGPT, Claude): paste as the first message or system instruction

## Step 2: Attach Corpus (if platform supports RAG)
- Upload `refined/` files as knowledge base / context
- If no RAG support: the System Prompt includes enough personality info to work standalone

## Step 3: Test
- Run test cases from `test-cases.md`

Cleanup (Optional)

Ask user if they want to:

• Keep ./clone-workspace/ for future iterations
• Archive to a specific location
• Move to Trash (~/.Trash/)

👉 克隆完成！你的数字分身已经准备好了。

Quick Start Examples

Clone yourself:

"帮我克隆自己，用我的公众号文章和 CC 对话记录" → Self Mode, start at Stage 1, auto-scan local corpus

Clone Naval:

"帮我克隆纳瓦尔做数字导师" → Mentor Mode, start at Stage 1, launch 6-angle research agents

Clone with existing corpus:

"我已经收集好了 Paul Graham 的语料，在 ~/pg-corpus/ 里" → Mentor Mode, skip to Stage 3 (point raw/ to user's directory)

Iterate existing clone:

"上次做的分身风格不够冷峻，帮我调一下" → Jump to Stage 4, re-read corpus, adjust System Prompt

Recommended Mentor Targets (名人导师推荐)

For users who want suggestions, here are high-value targets with rich public corpus: