Andrej Karpathy recently shared how he uses markdown files to build personal knowledge bases for AI research. His approach is simple: collect source documents, convert them to markdown, organize them in a folder, and let an LLM work with the whole collection.

The result? An AI that actually knows what you know.

This isn’t a research-only technique. Anyone who uses Claude, ChatGPT, or any LLM daily can build the same thing --- and it takes minutes, not hours.

What Is an LLM Knowledge Base?

It’s a folder on your computer filled with markdown files on topics you care about. Articles you’ve read. YouTube videos you’ve watched. Documentation you reference. Tweets that shaped your thinking.

When you give an LLM access to this folder, it stops being a generic assistant. It becomes your assistant --- one that understands your domain, your sources, and your context.

Think of it as a second brain, but one your AI can actually read.

Step 1: Collect Your Sources With Save

This is where most people get stuck. Manually copying web content, cleaning up HTML, formatting it --- it’s tedious. That’s exactly what Save solves.

Save is a Chrome extension that converts any webpage into clean, structured markdown. One click.

What you can save:

Articles and blog posts
YouTube videos (full transcript + summary)
Twitter/X threads
Documentation and API references
Reddit discussions
Academic papers
LinkedIn posts
Paywalled content you have access to

Each save produces a clean .md file --- no HTML, no ads, no navigation menus. Just the content, properly structured with headings, lists, and code blocks.

Step 2: Organize Your Knowledge Folder

Create a folder structure that makes sense for your work. Keep it simple:

my-knowledge/
  marketing/
    distribution-strategies.md
    competitor-analysis.md
    pricing-frameworks.md
  product/
    user-research-findings.md
    feature-prioritization.md
  industry/
    market-trends-2026.md
    karpathy-llm-knowledge-bases.md
    ai-agent-workflows.md

Every time you read something valuable online, save it with Save and drop it in the right folder. Over time, this compounds into a comprehensive knowledge base that no generic LLM can match.

Step 3: Connect Your Knowledge Base to an LLM

Here’s where it gets powerful. There are several ways to give an LLM access to your knowledge folder:

Option 1: Claude Code (Most Powerful)

If you use Claude Code, open a terminal in your knowledge folder and start a session. Claude gets full access to every file and can search, read, cross-reference, and reason across your entire collection.

cd my-knowledge
claude

Claude Code can read hundreds of markdown files, find connections between them, answer complex questions, and even maintain index files --- exactly like Karpathy describes.

This is the most powerful option because Claude can actively browse your files, not just receive them as context.

Option 2: Claude.ai Projects

Create a Project on claude.ai and upload your markdown files to the project knowledge base. Every conversation in that project will have access to your documents.

Drag and drop files directly
Supports 200+ files per project
Knowledge persists across conversations
Works on any device with a browser

Best for: people who want a simple setup without using the terminal.

Option 3: Claude Desktop + MCP

If you use Claude Desktop, you can install the filesystem MCP server to give Claude direct access to a folder on your computer.

Add this to your Claude Desktop config:

{
  "mcpServers": {
    "knowledge-base": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/path/to/my-knowledge"
      ]
    }
  }
}

Claude Desktop can then read any file in your knowledge folder on demand --- no uploading needed.

Option 4: Copy-Paste (Simplest)

For quick questions, just drag one or a few markdown files into any Claude or ChatGPT conversation. No setup required.

This works for one-off queries but doesn’t scale. For a real knowledge base workflow, use one of the options above.

Why Markdown Makes This Work

You can’t feed raw HTML to an LLM and expect good results. HTML is bloated with navigation, ads, scripts, and styling --- noise that wastes tokens and confuses the model.

Markdown is the native language of LLMs. They were trained on millions of markdown files from GitHub, documentation sites, and developer blogs. When you give an LLM clean markdown:

10x fewer tokens compared to raw HTML
Better comprehension --- the model focuses on content, not parsing
Proper structure --- headings, lists, and code blocks are preserved
Universal compatibility --- works with Claude, ChatGPT, Gemini, or any LLM

This is why Save exists. It doesn’t just copy text --- it uses AI to extract the meaningful content and structure it as clean markdown that LLMs understand perfectly.

The Compound Effect

Here’s what happens when you build this habit:

Week 1: You have 10 markdown files. Claude can answer questions about those specific articles.

Month 1: You have 50+ files. Claude can cross-reference sources, find patterns, and give you insights you’d never spot on your own.

Month 3: You have 200+ files. You’ve built a personal research assistant that knows your industry, your competitors, your frameworks, and your thinking. Every question you ask gets an answer grounded in your curated knowledge, not generic training data.

As Karpathy noted, once your knowledge base is big enough, you can ask complex questions and the LLM will research the answers across your entire collection. No fancy RAG setup needed --- just organized markdown files.

Get Started in 60 Seconds

Install Save from the Chrome Web Store (free)
Create a my-knowledge folder on your computer
Visit an article you found valuable recently
Click Save --- the markdown file downloads instantly
Move it to your knowledge folder
Open the folder in Claude Code, upload to a Claude Project, or set up MCP

Every article you save from now on makes your AI smarter. It compounds.

Start building your LLM knowledge base today. Install Save --- it’s free to start.