A 6-stage semi-automated workflow to create high-fidelity digital clones (mentors/personas) from corpus data.
Corpus-driven, platform-agnostic, human-in-the-loop.
用语料驱动的方式，构建任何人的数字分身。支持克隆自己，也支持克隆名人导师。
| Stage | Name | What Happens |
|---|---|---|
| 1 | Target Profiling | Identify clone target, map out data sources |
| 2 | Data Hunting | Collect raw corpus (auto for self, guided for public figures) |
| 3 | Data Refining | Clean, deduplicate, assess corpus quality |
| 4 | Soul Forging | Extract personality + generate System Prompt |
| 5 | Verification | Test with trap questions, iterate until pass |
| 6 | Deployment | Deploy to NotebookLM + Gemini Gem / CC Bot / any LLM |
- Self Mode: Clone yourself using your own writing, transcripts, and published content
- Mentor Mode: Clone a public figure (Naval, Paul Graham, Munger, etc.) by collecting their public corpus
Copy SKILL.md to your Claude Code skills directory:
mkdir -p ~/.claude/skills/digital-clone
cp SKILL.md ~/.claude/skills/digital-clone/Start a new Claude Code session and say:
# Clone yourself
帮我克隆自己，用我的文章和对话记录
# Clone a public figure
帮我克隆纳瓦尔做数字导师
# Clone with existing corpus
我已经收集好了 Paul Graham 的语料，在 ~/pg-corpus/ 里
The skill will guide you through each stage with confirmation checkpoints.
- First-hand corpus > second-hand summaries: Raw voice, original writing, real Q&A
- Data determines the ceiling: More high-quality corpus = better clone
- Platform-agnostic output: Produces System Prompt + cleaned corpus, deployable anywhere
- Context-aware tone: The clone knows when to be casual and when to be serious
| Person | Domain | Corpus Richness |
|---|---|---|
| Naval Ravikant | Wealth + Happiness | Rich (blog, X, podcasts) |
| Paul Graham | Startups + Thinking | Rich (essays, X, HN) |
| Charlie Munger | Multi-disciplinary Thinking | Good (speeches, books) |
| Nassim Taleb | Anti-Fragility + Risk | Rich (books, X) |
| Ray Dalio | Principles-Based Decisions | Good (books, LinkedIn) |
作者的开发环境，仅供参考，你可以用自己喜欢的工具替代
| Item | Setup |
|---|---|
| Hardware | MacBook Air M4, 16GB RAM |
| Models | Claude Opus 4.6 (primary), Gemini Pro 3 (secondary), MiniMax M2.5 (scheduled tasks) |
| Runtime | Bun, Docker |
| API | OpenClaw subscription |
Author's setup — yours may differ.
这些项目配合使用效果更好
| Project | What It Does |
|---|---|
| content-alchemy | 5-stage content pipeline (v5.0) — use this clone's persona in Stage 5 for personalized writing |
| content-publisher | Image generation, layout formatting, and WeChat API publishing |
| openclaw-worker | Task API + Worker for bot-driven automation |
| openclaw-cli-bridge | Discord → CC/Codex/Gemini bridge |
| local-memory | Hybrid search over AI conversation transcripts |
| cc-shell | Lightweight Claude Code chat UI |
| telegram-ai-bridge | 3 Telegram bots for Claude / Codex / Gemini via SDK |
| telegram-cli-bridge | Telegram CLI bridge (Gemini CLI path) |
Inspired by @MinLiBuilds's zero-code Naval digital clone tutorial. The method of using Deep Research for data mapping, NotebookLM as a RAG engine, and Gemini Gem for personality injection was pioneered in that thread.
This skill adapts and extends that methodology into a reusable, semi-automated Claude Code workflow with Self/Mentor dual modes, corpus quality assessment, and structured verification testing.
Built by 小试AI (@AliceLJY) · WeChat: 我的AI小木屋
医学出身，文化口工作，AI 野路子。公众号六大板块：AI实操手账 · AI踩坑实录 · AI照见众生 · AI冷眼旁观 · AI胡思乱想 · AI视觉笔记
Six content pillars: Hands-on AI · AI Pitfall Diaries · AI & Humanity · AI Cold Eye · AI Musings · AI Visual Notes
MIT