Andrej Karpathy's Claude Code Method Just Changed AI

Andrej Kapafi, one of the co-founders of OpenAI, the guy who built Tesla's AI team, just shared something that most people scrolled past. And it's one of the most practical AI workflow ideas I've seen in a long time. Today, I'm going to walk you through exactly how to build what he's talking about using Cloud Code and Obsidian. Now, here's the short part.

Most people use AI like a smart search engine, right? Ask a question, get an answer close to that. The model forgets everything. Next session, you're basically starting from zero again every single time.

Now, Kapafi said something that hit me when I read it. He said a big chunk of how he uses AI has shifted, right? Less coding, more using LLMs to manipulate and to build knowledge stored as markdown files. And this whole system is about building something that gets smarter over time, right?

And compounds every week. And the tool sitting at the center of that is Cloud Code. Now, here's what makes this different from every other take a better note system you've ever seen. In those systems, you do all the work.

You organize, you tag, you structure it all. In this system, Cloud does most of the heavy lifting. You capture raw information, Cloud compiles it into a structured knowledge base, and you just read and ask questions. So let me walk you through exactly how this works step by step.

You start with Obsidian in three different folders, right? One called Raw, one called Wiki, and one called Reports. That's the whole architecture. Anything interesting you find, articles, transcripts, research, YouTube videos, gets dropped into Raw.

Now, if you're not sure about Obsidian, you can download it for free. It's super simple. And it's just like kind of Apple Notes or something like that, right? It's just a note-taking app.

But you can easily connect it to your AI agents like Claude, and also OpenCore and Hermes and whatever else you're using. Now, when you store this stuff into Raw, it's like a disorganized folder, right? So you don't sort it, you don't rename it, you just dump it in there. And the Obsidian web browser extension makes this automatic.

So in one single click, you can save any web page into your Raw folder. It's a clean markdown with the URL, the date, the title. It takes about two seconds, right? It's kind of like fancy bookmarking.

And that's important because if capturing is annoying, the system doesn't work from day one, right? You have to make this easy for yourself. And now here's where Claude Code comes in. So you set up what Capathe calls a compiler.

And Claude Code is absolutely perfect for this because it runs inside a real file system. It can read your files, it can write files, it can run scripts. It's an agent that works directly inside your local folders. So Claude reads everything inside your Raw folder, and then it builds pages inside your Wiki folder, right?

So each important concept gets its own markdown file, a short definition, key insights, and links back to the original sources, connections to related concepts. So if you've been bookmarking articles about AI agents for three weeks, well, Claude builds you a clean AI agents page that would pull from them, right? And learn from them and cross-reference them and link them together in different ways. You barely touch the Wiki folder manually.

Claude owns it. And inside Obsidian, you can see all of this visually. So the cool thing about this is there's a graph view that shows every note as a dot and every connection between notes as a line. So you can literally see clusters of knowledge forming.

You can see gaps. Isolated notes with no connections are topics you haven't explored yet. And this is where something clicks because you stop thinking of your notes as a list of things you've saved. You start to think of them as a map of what you know, right?

Now, if you want step-by-step coaching on how to set Claude Code up like this, plus a 30-day roadmap built specifically around using Claude Code for your business, we've got that inside the ARP offer boardroom. There are 2,800 business owners in there right now, a lot of them running Claude Code for content, for client work, and for lead generation. We've built the tutorials, the prompts, and the walkthroughs around exactly this kind of system. Link in the comments description or go to the AIProfitBoardroom.com to get access.

Now, let's talk about the part that most people skip here because building the wiki is only half of it. The real power is what you do once you have like 50 or 100 articles inside there. Instead of asking Claude a generic question and getting a generic answer, you ask Claude a question grounded in your own research. So you can say, for example, based on everything in my wiki about AI agents, what are the three most underrated use cases for small businesses or based on what you know about me?

And then Claude reads your files. It pulls from your research and writes the answer as a new file in your reports folder with links back to the sources and the pages it learned from. Every serious question becomes a permanent asset. Every answer files back into the knowledge base and makes it smarter.

And the key habit here is this. You stop answering in chat, right? You force Claude to answer in files, right? Report, slide decks, using the Marp plugin, which turns Markdown into full presentation slides.

So diagrams are saved as images and everything goes into the vault and everything is reusable as well. Now, let me show you the layer that most people don't know about. There's an MCP server for Obsidian, has over 3,300 stars on GitHub that lets Claude read and write directly into your vault in real time. So Claude can browse your folder structure, open any note, search for any topic, insert content under specific headings, and then create new files and append to existing ones, right?

All from one chat window. So instead of copying and pasting things back and forth, Claude is literally working inside your notes. You can say, find all of my notes that mention lead generation, give me a summary of each one. And Claude does it, right?

You can say, take my last meeting notes and write a follow-up email draft in a new file. You could say, for example, add this idea to my research and note the strategy section. And it puts it exactly there. So the setup is three simple steps.

In Obsidian, you go to settings, then community plugins, search for local REST API, install and enable it, and then copy the API key from the plugin settings. And then you open your Claude desktop config file. On Mac, it's in library, then application support, then Claude, and you add a short config block with your API key. You restart Claude and you're done.

Or you can just ask Claude to set up and you give it your API key and it will fix it directly, right? But Obsidian needs to be open whilst you're using it. And the easiest way to start is to begin with your first prompt with use Obsidian 2, right? Or do this with Obsidian or search Obsidian about this.

So based on the notes inside Obsidian, go and search and give me some answers back on this, right? So Claude knows to access the vault immediately, right? Now, one thing worth knowing is that if Claude can't find the UVX command, you can open up your terminal, type which UVX, copy the full path and paste that into the config file instead of just typing UVX, right? And that fixes it.

But this is a little technical tweak that you probably don't need to worry about. Now, here's what makes this system genuinely different from everything else out there. Every other AI workflow is stateless, right? So you get one conversation, it ends, the model forgets you, you come back tomorrow and explain everything again from scratch.

This is stateful, right? Basically, you don't have enough context to get really useful answers out of the AI. But today's question now becomes tomorrow's context, right? 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. And your questions compound into better questions, right? Your answers compound into better answers. And here's what this means practically.

So for example, if you're a freelancer using this system to research a niche and track client strategies, after six months, a private intelligence system is available that your competitors don't have, right? This is a competitive advantage. Also an agency owner tracking AI tools or case studies and client workflows has documentation, onboarding materials, and a research archive all in one place. And the gap between the people building this and the people who aren't is going to widen fast, right?

There's also a privacy option worth knowing too. So if you don't want your research going to the cloud, you can run the whole thing locally, right? Obsidian is already local first, but you can pair it with Olama, which runs local language models on your machine and nothing leaves your computer. So Claude Code can still layer on top four tasks where you're comfortable using cloud processing, or you can set up Olama directly with Claude Code, and then everything stays local and private, right?

And Karpathy mentions one more thing that's actually worth thinking about. So once your wiki gets big enough, you can use it as a fine tuning corpus, right? Meaning you can bake your knowledge directly into a model's weights. So for a team, that means internal docs, processes, and institutional knowledge becoming a customer system.

That's a longer term play, but it starts building the system right now, right? So if you start working on this now, in six months, 12 months, you've got something that you could actually fine tune an AI local model with, which is amazing. So here's where to start this weekend, right? Day one, install Obsidian, create your raw wiki and reports folders, install the extension for Obsidian, and then pick one topic you care about, right?

It could be AI tools, could be SEO, could be your industry, and just gather some articles and some bookmarks together. You don't need to organize anything, just capture. And then day two, inside Claude Code, you can build a simple skill that reads your raw folder and compiles it into concept pages in your wiki, right? You run it, you open Obsidian, and you look at the graph view.

Then you ask one real question about your topic and have Claude write the answer into your reports folder. And if you do that seriously for like two days, you already feel the difference, right? And this stops being a note-taking exercise, it starts being a system that actually works for you. Now, inside the AR Profit Boarding, we've built a 30-day roadmap around Claude Code, including exactly how to set up the Obsidian integration, how to build the skills and prompts to power the system, and how to use it to generate leads and deliver better client work.

We've got four coaching calls every week, step-by-step videos, tutorials, and 2,800 members, a lot of them already running Claude Code inside their business and happy to share what's working. So if you want to build this system properly and use it to actually grow your business, link in the comments description or go to thearprofitboarding.com to check it out. Now, this is what separates people who use AI as just a simple answer machine or a search engine from people who use it as a real business asset, who get a competitive advantage from it. Kapafi built his system over time, yours starts this weekend.