Andrej Karpathy's three-folder knowledge system

I read a lot and watch a lot, but when I actually need it, I’ve forgotten most of it.

I just watched an explanation of how Andrej Karpathy manages what he knows. He’s one of the founders of OpenAI and the former Director of AI at Tesla.

What surprised me: he doesn’t use Notion. He doesn’t use Obsidian. Just ordinary folders and text files.

That way, every time he learns something new, the old stuff doesn’t get buried. It connects.

Three folders

The system is just three folders:

  • raw/: where you dump all the raw material
  • wiki/: a tidy archive organized by AI
  • outputs/: analyses and reports

I picture it like a cabinet. The first drawer is full of messy papers. The second drawer has been tidied up by an assistant. The third drawer holds the work made from all that material.

Simple. But that’s exactly what makes it work.

The part that changed how I think the most is the raw/ folder. This is where everything gets dumped: articles, screenshots, meeting notes, PDFs, study material, anything.

And the rule? Don’t tidy it. Don’t think about file names.

I used to get stuck a lot because I wanted to “get organized first, then start.” It turns out that’s what kept me from ever starting. Just dump it in. Let the AI do the cooking later.

Letting the AI tidy up

So who tidies it up? The AI.

You write one instruction file, a kind of SOP for your assistant. It covers:

  • What topics you’re focusing on
  • How the wiki is structured
  • How to write summaries
  • Your interests and priorities

Write it once, use it over and over. I think of it like a brief to a designer: if the brief is clear, the result doesn’t drift.

Once the material is in and the instructions are ready, you just tell the AI:

“Read everything in raw/. Organize it into a wiki. Make a table of contents. Link the topics that connect.”

The AI reads, groups, summarizes and connects everything. Our role? Read and review.

Honestly, this is the part that made me think, “Why didn’t I do this sooner?”

Knowledge that compounds

But the most exciting part for me is how it builds on itself. Once the wiki has a decent amount in it, you can start asking:

  • “What’s my biggest knowledge gap on topic X?”
  • “Source A vs. B, where do they differ?”
  • “Write a briefing from everything I’ve collected.”

Every question adds to it, so the next answer is richer. This is what’s meant by compound knowledge: learning that actually builds up instead of evaporating.

One thing I almost skipped but turned out to matter is a health check. If wrong AI output gets saved too, over time that error becomes a “fact.” So every month, ask the AI to review:

  • Are there any contradictions?
  • Are there topics that get mentioned but never explained?
  • Are there claims without a source?

I think of it as an audit, but for what’s in your head.

The habit matters more than the tool

What I like about this system: no expensive app, no complicated plugins, no new tool to learn. All you need is:

  • 3 folders
  • 1 instruction file
  • 1 AI you use consistently

Karpathy himself just uses folders and text files. The tool is only the surface. What actually makes it work is the habit of looping with AI.

If you want to try it, AI tools that can read local folders like this include Claude Code (runs in the terminal) or Cursor (a code editor with AI built in). What matters is that the tool can access the files on your laptop, not a regular online chatbot where all you can do is copy and paste.

So in short, the whole system is:

  • Dump all your material in one place, and don’t tidy it
  • Write clear instructions for the AI, once
  • Let the AI organize it, while you review and ask questions

I’m still early in trying it myself. But the logic really clicks for me. I’ll stop looking for the perfect tool and just start.

The full explanation is in Nick Spisak’s thread on X about how to build your second brain. Really worth it.