01What is Andrej Karpathy's Claude Code Method?

Andrej Karpathy's Claude Code method is a workflow where you use Claude Code to turn your raw notes into a growing, connected knowledge base. Instead of using AI like a smart search engine that forgets everything after each chat, this system keeps your information and makes it smarter over time.

Karpathy, a co-founder of OpenAI and former head of Tesla's AI team, shared this idea. Most people missed it. But it is one of the most practical AI workflows.

You capture random articles, transcripts, or YouTube videos. Then Claude Code reads those files and builds a Wiki of organized concept pages. You do not have to sort or tag anything.

Claude Code does the heavy lifting.

02Main Features of Karpathy's Knowledge-Building System

This system has three main parts. First, the Obsidian vault with three folders: Raw, Wiki, and Reports. Raw is for unorganized captures.

Wiki is for structured concept pages built by Claude Code. Reports is for answers Claude writes based on your research.

  • An Obsidian web browser extension lets you save any web page into Raw in one click. It keeps the URL, date, and title as clean markdown.
  • Claude Code acts as a compiler. It reads everything in Raw and creates a Wiki page for each important concept, with definitions, insights, and links to original sources.
  • Obsidian's graph view shows every note as a dot and connections as lines. You can see clusters of knowledge and spot isolated notes — topics you have not explored yet.

The real power comes from the MCP server for Obsidian. It has over 3,300 stars on GitHub. It lets Claude read and write into your vault in real time.

Claude can browse folders, open notes, search topics, and insert content under specific headings, all from one chat window.

03How to Set Up Karpathy's System with Obsidian and Claude Code

  1. Install Obsidian and create three folders inside it: Raw, Wiki, and Reports. Obsidian is free and works like Apple Notes.
  2. Install the Obsidian web browser extension so you can save any web page into Raw with one click. This makes capturing fast and automatic.
  3. Set up the MCP server for Obsidian: Go to Obsidian settings, then community plugins, search for local REST API, install and enable it, and copy the API key from the plugin settings.
  4. Configure Claude Code: Open your Claude desktop config file (on Mac it is in Library/Application Support/Claude). Add a config block with the API key from step 3. Restart Claude.
  5. Test the connection: Ask Claude to search a note in Obsidian or to do something with Obsidian. Claude will access your vault immediately. If Claude cannot find the UVX command, open your terminal, type which UVX, copy the full path, and paste it into the config file.

Remember: Obsidian needs to be open while you use Claude with it.

04How to Use the System to Capture and Compile Knowledge

Start by capturing any interesting content — articles, transcripts, research, YouTube videos — into your Raw folder. Do not sort or rename anything. Just dump it in.

The Obsidian browser extension makes this automatic. In two seconds, you save a web page as a clean markdown file with URL, date, and title.

  • Set up Claude Code as a compiler: Write a skill that tells Claude to read your Raw folder and build concept pages in your Wiki. Each Wiki page gets a short definition, key insights, and links back to the original sources.
  • Ask grounded questions: Once you have 50–100 pages, ask Claude questions based on your own research. For example, "Based on everything in my wiki about AI agents, what are the three most underrated use cases for small businesses?"
  • Force answers into files: Have Claude write answers as new files in your Reports folder. This turns every serious question into a permanent asset. Use the Marp plugin to turn Markdown into full presentation slides.

You stop answering in chat. You force Claude to answer in files. Everything stays reusable and grows your knowledge base.

05Results and Comparisons Shown in the Episode

Karpathy's system is stateful, not stateless. Most AI workflows are stateless — one conversation ends, the model forgets you. You come back tomorrow and explain everything again from scratch.

This system keeps state. Today's question becomes tomorrow's context. Your terminology becomes a canonical page in your Wiki.

Six months of research becomes a private knowledge base that Claude can read in a single context window. Questions compound into better questions. Answers compound into better answers.

If you are a freelancer researching a niche and tracking client strategies, after six months you have a private intelligence system your competitors do not have. If you are an agency owner tracking AI tools or case studies, you get documentation, onboarding materials, and a research archive all in one place. The gap between people building this and those who do not will widen fast.

06Tips, Limits, and Mistakes to Avoid

Make capturing easy. If capturing is annoying, the system fails on day one. Use the Obsidian web browser extension to save pages in one click.

Do not sort or rename anything at first. Just capture.

  • Do not touch the Wiki folder manually. Claude owns it. Let it compile and organize everything.
  • Force answers into files. Never answer in chat. Have Claude write answers as new files in your Reports folder. This makes every answer reusable and permanent.
  • Watch for isolated notes in the graph view. Notes with no connections are topics you have not explored. They show gaps in your knowledge.
  • If Claude cannot find the UVX command, fix it by typing which UVX in the terminal and copying the full path into the config file.

One limit: Obsidian must be open while you use Claude with the MCP server. For privacy, you can run the whole thing locally with Ollama. That way nothing leaves your computer.

Claude Code can still layer on top for tasks where you are comfortable using cloud processing.

07Who This Method Is For and Why It Matters

This method is for anyone who wants to use AI as a real business asset, not just a search engine. If you are a freelancer, an agency owner, or a business owner, you will benefit. You build a private knowledge base that gets smarter every week.

Your competitors will not have this.

Karpathy mentions one more long-term play: once your Wiki gets big enough, you can use it as a fine-tuning corpus. This means you bake your knowledge directly into a model's weights. For a team, that turns internal docs and processes into a custom system.

It starts building now. If you start this weekend, in six or twelve months you can fine-tune an AI local model with your own research.

The system stops being a note-taking exercise. It becomes a system that works for you. You capture raw information, Claude compiles it, and you just read and ask questions.