Karpathy's recent tweet about his AI-powered personal knowledge base went viral, garnering 44K saves and millions of views in just two days. The key insight? He leverages AI to organize scattered notes and articles into a searchable system — no Notion or Obsidian needed, just a simple folder structure and a text file. Here's how to build your own, with practical steps you can implement today.
The 3-Folder Structure: Super Simple, Fully Flat
This system can be set up in two minutes, with each folder serving a distinct purpose:
- raw/: A drop zone for all your unorganized raw materials — articles, screenshots, notes, etc. Just dump everything here without worrying about organization.
- wiki/: The AI-curated knowledge base. This is where AI writes organized content after processing raw materials. Never edit this folder manually.
- outputs/: Stores AI-generated answers, reports, and analysis results.
mkdir my-knowledge-base
cd my-knowledge-base
mkdir raw wiki outputsThe Schema File: Your AI's Instruction Manual
A schema file is crucial to guide AI on how to structure and manage your knowledge base. Create a CLAUDE.md file in the root directory with rules like:
# What This Is
A personal knowledge base about [your topic].
# How to Organize
- raw/: Raw materials, never modify.
- wiki/: AI-maintained organized content.
- outputs/: Generated reports and answers.
# Wiki Rules
- One .md file per topic.
- Link related topics using [[topic-name]] format.
- Maintain INDEX.md to list all main topics.Karpathy emphasizes that this schema is "super simple" — just a text file telling AI the rules. It directly impacts the quality of the AI-curated wiki, so don't skip this step.
Launch AI Processing with One Command
With the folders and schema ready, open Claude Code and issue this prompt:
"Read all content in raw/. Follow the rules in CLAUDE.md to compile a wiki in wiki/. First create INDEX.md, then create a .md file for each main topic. Link related topics. Summarize each section."
Then let AI run — no need to monitor it. Once complete, you'll get a wiki folder with topic-organized articles and an INDEX.md for easy navigation.
Accelerate Content Collection with agent-browser
To supercharge raw material collection, use agent-browser — a tool that lets AI control a real Chrome browser to scrape web content. It handles JavaScript-rendered pages, login-required content, and infinitely scrolling pages, while saving 82% more tokens than tools like Playwright MCP.
npm install -g @vercel/agent-browser
agent-browser install
# Example usage
agent-browser open https://some-article.com
agent-browser get text "article" # Saves text directly to raw/Start Using Your Knowledge Base Today
Karpathy's knowledge base has 44K saves, but the real value comes from actually building and using it. With three folders, one schema file, and an AI tool, you can have your own system up and running in a weekend. Pick a topic you're eager to master, dump your existing articles and notes into raw/, and let AI do the rest.
常见问题
Do I need Claude Code specifically, or can I use any AI tool?
The system works with any AI coding agent that can read files and write to disk — Claude Code, Codex, Cursor, or even ChatGPT with file access. Claude Code is recommended because its agent loop naturally handles the multi-step process of reading raw files, organizing content, and writing wiki articles without manual intervention. The CLAUDE.md schema file is named for Claude Code but the concept works universally — just adapt the filename to your tool's convention (AGENTS.md for Codex, .cursorrules for Cursor).
How is this different from using Notion or Obsidian?
Karpathy's system is intentionally simpler: plain Markdown files in folders, no proprietary formats, no vendor lock-in. You own your data as files on disk. The AI does the organizing work that you'd normally do manually in Notion — tagging, linking, summarizing. It's a build system for knowledge: dump raw materials in, get organized wiki out. Notion and Obsidian are great tools, but they require you to do the organizing. This system lets AI handle that.
What happens if the AI makes mistakes in the wiki?
The wiki/ folder is AI-maintained — never edit it manually. If you find errors, fix the source material in raw/ (or add corrective notes) and re-run the AI processing. The system is designed to be regenerative: you can delete the entire wiki/ folder and rebuild it from raw/ at any time. Think of raw/ as your source of truth and wiki/ as a compiled output.