Andrej Karpathy Just 10x’d Everyone’s Claude Code

[0:00]What you're looking at right here is 36 of my most recent YouTube videos organized into an actual knowledge system that makes sense. And in today's video, I'm going to show you how you can set this up in five minutes. It's super, super easy. You can see here how we have these different nodes and different patterns emerging, and as we zoom in, we can see what each of these little dots represents. So, for example, this is one of my videos, $10,000 agentic workflows. We can see it's got some tags, it's got the video link, it's got the raw file, and it gives an explanation of what this video is about and what the takeaways are. And the coolest part is, I can follow the back links to get where I want. There's a back link for the WAT framework, there's a back link for Claude Code. There's a back link for all these different tools I mentioned, like Perplexity, Visual Studio Code, Nano Banana, and N8N. It also has techniques like the WAT framework or bypass permissions mode or human review checkpoint. So as this continues to fill up, we can start to see patterns and relationships between every tool or every skill or every MCP server that I might have talked about in a YouTube video. And I can just query it in a really efficient way now that we have this actual system set up. And the crazy part is, I said, hey, Claude Code, go grab the transcripts from my recent videos and organize everything. I literally didn't have to do any manual relationship building here. It just figured it all out on its own. And right here I have a much smaller one, but this is more of my personal brain. So this is stuff going on in my personal life, this is stuff going on with, you know, up to AI or my YouTube channel or my different businesses and my employees and our quarter two initiatives and things like that. This is more of my own second brain. So I've got one second brain here, and then I've got one basically YouTube knowledge system. And I could combine these or I could keep them separate and I can just keep building more knowledge systems and plug them all into other AI agents that I need to have this context. It's just super cool. So Andre Karpathy just released this little post about LLM knowledge bases and explaining what he's been doing with them. And in just a matter of a few days it got a ton of traction on X. So let's do a quick breakdown and then I'm going to show you guys how you can get this set up in basically five minutes. It's way more simple than you may think. Something I'm finding very useful recently is using LLMs to build personal knowledge bases for various topics of research interest. So there's different stages. The first part is data ingest. He puts in basically source documents, so he basically takes a PDF and puts it into Claude code and then Claude code does the rest. He uses Obsidian as the IDE, so this is nothing really too game-changing. Obsidian just lets you visually see your markdown files. But for example, this Obsidian project right here with all this YouTube transcript stuff, that actually lives right here. This is the exact same thing. Here are the raw YouTube transcripts, and here's that Wiki that I showed you guys with the different, um, folders for what Claude Code did with my YouTube transcripts. And then there's a Q&A phase where you basically can ask questions about YouTube or about the research. And it can look through the entire Wiki in a much more efficient way and it can give you answers that are super intelligent. He said here, I thought that I had to reach for fancy rag, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this small scale. So right now he's doing about 100 articles and about half a million words. So there's a few other things it will cover later, but the TLDR is you give raw data to Claude Code, it compares it, it organizes it, and then it puts it into the right spots with relationships, and then you can query it about anything. And it can help you identify where there's gaps in that node or in that, you know, relationship, and it can go do research and fill in the gaps. All right, so why is this a big deal? Because normal AI chats (or tools like Notebook LM or basic rag) are ephemeral, the knowledge disappears after the conversation. But this method using Karpathy's LLM Wiki makes knowledge compound like interest in a bank. People on X are calling it a game-changer because it finally makes AI feel like a tireless colleague who remembers everything and stays organized. It's also super simple. It will take you five minutes to set up, I'll show you guys. You don't need a fancy vector database, embeddings, or complex infrastructure, it's literally just a folder with markdown files. That's it. You literally just have a vault up top. So in this example, it's called my Wiki. You've got a raw folder where you put all of the stuff, and then you've got a Wiki folder, which is what the LLM takes from your raw and puts it into the Wiki. So in here, you have all the Wiki pages, which it will create, but then you also have an index, and you have a log. So, for example, in my YouTube transcripts vault, here is the index. You can see that I have all these different tools, which I could obviously click on and it would take me right to that page. Or after that, I have all the different techniques, agent teams, sub agents, permission modes, the WAT framework. And then we've got different concepts, MCP servers, rag, vibe coding. We've got all these different sources, which are, you know, the YouTube videos. And then when I have people or when I have comparisons, they will be put in here in the index. And then we also have a log, which is the operation history. So in this case, in the YouTube project, the log isn't huge, because I only ran one huge batch of the initial 36 YouTube videos. But now every time I have one, I say, hey, can you go ahead and ingest the new YouTube video into the Wiki? And then we'll see every single time we update this. And then of course, you need your claw.md to explain how the project works and how to search through things and how to, you know, update things. It's also a big deal from a cost perspective, token efficiency plus long-term value. One X user turned 383 scattered files and 130 meeting transcripts into a compact wiki and dropped token usage by 95% when querying with Claude. And obviously, token management and efficiency is a huge conversation right now and will always be. The other thing that's really cool about this is there's not really like a GitHub repo you go copy or there's not a complicated setup. You literally just say, hey, Claude Code, read this idea from Andre Karpathy and implement it. And people on X are now talking about like this is how 2026 AI agentic software and products will be made. You just give it a high-level idea and it goes and builds it out. And Karpathy even said, hey, you know, I left this prompt vague so that you guys can customize it. And I'll show you the ways in my two different vaults right now that it changed things a little bit based on the context and understanding of what the project is actually for. Okay, so this was the original tweet I just showed you guys, and then he followed up and said, hey, this one went viral. So here is the idea in a gist format. So if you open this up, this is basically just another explanation of the core idea of how this works and why the architecture, indexing, all this kind of stuff. And by the way, this is the part where he says, hey, this is left vague so that you can hack it and customize it to your own project. So, we're going to come right back to this in a sec, but the first pre-wreck that we're going to do, it's not necessary, but I like to have a nice little front end to see the relationships. Is we're going to go to Obsidian and download it. So if you just go to obsidian.md, you can see this is the completely free tool, and you're going to go ahead and download it. So just for your operating system, download this, and then open up the wizard and open up the app. So when you open up the app, it'll look like this, and what we're going to do here is we're going to create a new vault. So down here, you can see I have Herk Brain, and I have YouTube Transcripts. I'll just make it a little bigger. I'm going to go to manage vaults. I'm going to create a new one, and now we just have to give this a name. So I'm just going to call this one demo vault, and you're going to choose a location where you want to put this. So I'm just throwing this on my desktop, and I'm going to go ahead and create this vault. Then what you're going to do is go to wherever you like to run Claude Code. So in this case, I'm doing it in VS Code, and I open up that folder. So demo vault, we get an Obsidian, and then we get a welcome.md. So I'm going to open up Claude, so I'm going to do it in my terminal. I'm going to run Claude. And lately I've been liking using my terminal better for Claude. I like to do it in inside of VS Code, but the reason is just because I like to see the status line, and I have, you know, a little bit more functionality. So anyways, now that we have Claude Code open, here's what we're going to do. We're going to go back over to the LLM Wiki thing that we got from Andre Karpathy. We're going to copy all of this, and we're going to go back into Claude Code. And then just paste it in there. So that is the prompt from Karpathy that's going to build out everything we need. And then before we send that off, we're dropping this in, which you guys can screenshot and then just throw into yours. But I'm saying, you are now my LLM Wiki agent. Implement this exact idea file as my complete second brain. Guide me step-by-step, create the Claude.md schema file with full rules, set up index.md and log.md, define folder conventions, and show me the first ingest example. From now on, every interaction follows the schema. So anyways, on the right, we have this Claude Code running, and on the left, we have our Obsidian vault. And you can see it just created those two folders. So it created the raw and it created the Wiki, as you can see. Now, by default, it threw in four folders, it threw in analyses, concepts, entities, and sources. Once we start to populate stuff, we can talk to it to see if that's actually the way we want to do it or not. Because it's interesting in my personal kind of second brain, the Wiki is literally just markdown files. There's no structure to it. And in some cases that's good. Karpathy actually said, sometimes I like to keep it really simple and really flat, which means like no subfolders and not a bunch of over organizing. But then you guys did see in my YouTube transcript one, there were different subfolders and I think that in this case, it actually makes more sense. But you can see that it went ahead and it created a claw.md, it created an index, and a log, and then a few different folders in our Wiki. But now it's saying, hey, let's go ahead and try it out. Drop in your first source into the raw folder and tell me to ingest it. Okay, so I'm at this website called AI 2027. If you guys haven't read this before, it's kind of an interesting read, so go check it out. And now, let's say I want to get this into my vault. What I could do is just copy the whole page, right? And it might just come through a little weird. Or we can just get an Obsidian extension, which lets us basically take articles right from the web and just put it right into our vault, super easy. So search for this extension called Obsidian Web Clipper, you would go ahead and add this to Chrome. So then when you're in the article that you want, you basically just click on your extensions, you open up Obsidian Web Clipper, and then you can just chuck it into your vault. And then right here, you're going to want to set this to raw, because this is the actual folder that it's going to put it in. So you can go ahead and click add to Obsidian, open Obsidian, and then now you can see in my raw section, we have this AI 2027 source with the title, the source, and it's not super, super populated yet, because the LLM in Claude Code is going to do that. So here is our file. I'm going to open up Claude Code and say, awesome. I just threw in an article called AI 2027 into the raw. Can you please go ahead and ingest that? It might ask you some questions. It might also be helpful to before you start ingesting stuff, say, hey, by the way, this project is specifically for my second brain. So personal things, business things, whatever. Or this is just a research project. This is where I'm going to chuck you all of the articles and all the things that I want to learn about and all the things that I know. So there's different ways that you can set up the project, as you saw with mine, one for YouTube, one for just personal second brain. So now what it's doing is it's going to read through this article and then it's going to figure out where should I chuck everything into the Wiki? It's not just going to create one MD file for this. It might create five, or it might create 10, and there's going to be relationships between each of the different sections that it creates. So it's kind of doing its own method of chunking. Now, one thing I want to call out real quick is with this extension, if you go here and you open up the options for it, you can see that you can actually change where by default the folders are dropped. Which is in the location section, by default, it'll be going to a place called clippings, but just go ahead and change that to raw. Okay, so here it came back with all these questions, right? It said, here are my key takeaways from this article, blah, blah, blah. And now it'll ask you, what do you want to emphasize from this article? What's your focus? How granular do you want to be? What's your plan? So I'm just going to say, I want this to be extremely thorough. This is my passion, looking at where AI is going to go. This whole project, by the way, that you're setting up in this vault is basically just going to be my place to dump in research about AI. Help me keep all that organized so that I can query it and keep my thoughts related. So that's just a quick example of what it might look like for you to give it some more context to continuously build your project. So I'm going to switch over over here to the graph view, because I think it'll be interesting to see as it is starting to go through and create those different Wiki files. It's going to go ahead and it's going to create all those relationships, and we'll be able to watch it in real time. All right, so it's creating all of the Wiki pages now, and you can see that it said it's going to make about 25, because there's so much stuff going on in the original AI 2027 article. Okay, so our first one just popped in here, and there a second one just came through. And now you can understand, you're starting to see where do you have hubs or where do you just have little individual nodes. So this is a major hub. Someone named Eli, someone named Thomas, Daniel, and you can see all the different relationships here with things like AI governance, with things like Open Brain, Superhuman Coder. Okay, so that ingest took about 10 minutes. So sometimes you have to be a little patient with, you know, it reading through everything and organizing everything, but it does a lot of heavy lifting, of course. When I uploaded the 36 YouTube transcripts in batch, it took about 14 minutes, so it kind of just depends. But it created 23 Wiki pages. We have the source, we have six people, five organizations, and one AI systems page. Different concepts, so technical, alignment, and geopolitical, and then an analysis, and then it asks some questions about it so that it can help make the relationships and make the structure even better. Now, let's just open this one up a little bit deeper and see what it actually did in here with this stuff. So, we have, this is the source with all the main relationships. So as we start to add other articles, we will see other big kind of like nodes and maybe in some cases, we'll have relationships between like compute scaling with different articles that we upload as well. So let's just see. If I click into the main source, we can see the tags that it got, we can see the authors, and we can click around. So here's a link to OpenAI. Okay, what's OpenAI? Here's references in AI 2027. Here's some other connections with OpenAI, like model spec. Okay, we're in model spec, let's take a look. We can see other things about model spec, and we could also go to how the LLM psychology model works. So this is just super, super cool, all the relationships that we get. And once again, all of this stuff that we're looking at was derived from one article. And automatically organized and related. The question now is like, what do we do from here? Do we query it inside of this environment? Do we query it from somewhere else? And that's completely up to the way that you want to use this. So, for example, with my YouTube project, I'm probably just going to keep this here. And whenever I want to ask questions about YouTube or if I want to turn this into like a website, I can just do that from here. Or if I need to, I can point a different project at this folder, since everything's here. And it can crawl through the Wiki, it can read the index, and it knows how this stuff works, because you can give it the claw.md, so it understands the project as well. So, for example, in this one, which is just my second brain, where we have all of the different things about like I drop in my meeting recordings, I drop in, you know, click up channels, summaries, and things like that. This is something that I want to use in my executive assistant. So what I did in my executive assistant here, called Herk 2, if I go to this claw.md, you can see that we have a Wiki path. So whenever you need to read things about me and my business that you don't have already, you would basically go to my Herk Brain vault. You would go to that directory, and then you would read through the Wiki. You can read the hot cache, which I'll explain in just a sec, you can read the index, you can read the domain sub-index, and then you can also just search through everything here. And I said, don't read from the Wiki unless you actually need it. Here are some things that you might do that you don't need to go read the Wiki for. And all of this is my business knowledge. Now, if you guys remember, if you watch my video on setting up an executive assistant, I used to do this with context files inside of this project. And when I changed over to this method, I actually saw a reduction in tokens that I was actually calling in this project. So the thing about the hot cash, right? I didn't actually have this in my YouTube one. So if I go to YouTube, you can see there's no hot cash. But if I go to the Herk Brain in the Wiki, you can see there's a hot.md right here. And this is basically just a cache of like 500 words or 500 characters that it saves, which is like, what is the most recent thing that Nate just gave me or that we talked about? In the context of my executive assistant, this is really helpful. You know, it might save me from having to crawl different Wiki pages. But in something like the YouTube transcript project, I don't really need a hot cache. So another thing that I alluded to, but didn't really cover was the idea of linting. So Karpathy says that he runs some LLM health checks over the Wiki to find inconsistent data, impute missing data with web searchers, find interesting connections for new article candidates, etc. So it basically helps you run a lint, you know, every day, every week, whenever you want, which helps make sure that everything is scalable and structured in the right way. And it might even come back and say, hey, I don't fully understand this. Can you give me some more info or can you grab some more articles that might help me out here? So now the final question about this that I wanted to cover is like, does this kill semantic search rag? And the answer is no, but kind of yes. And it all depends on the goal of the project and the goal of the context, how much context you have. So here's a really quick chart that I had my Claude Code make. I was in my Herk Brain, where I dumped in a bunch of information about Karpathy's LLM knowledge, and I just said, hey, can you please explain Karpathy knowledge as simple as possible? Keep it super concise, and compare it to typical semantic search rag. So it found Karpathy's idea and said instead of a database, you just give the LLM well-organized markdown files with clear structure, indexes, and links between pages. And it compares it here to the actual semantic search rag. So actually, I might as well just read it off from here. So it finds it by reading indexes and follows links rather than using similarity search. So we're getting a deeper understanding of relationships, because they're links, rather than just saying, hey, these chunks seem similar. As far as infrastructure, it is literally just markdown. So like I said, you don't even need the Obsidian, you just need these markdown files. Whereas with semantic search, you need an embedding model, you need a vector database, and a chunking pipeline. The cost over here is basically free, your only cost is going to be tokens, whereas over here, you might have ongoing compute and storage. And for maintenance, you just run a lint, you clean up things, you add more articles, you give it more context, rather than having to re-embed when things change. But right now, the weakness, of course, with the LLM knowledge Wiki is that it doesn't scale huge across enterprises, right? Because it's just a bunch of files. Um, and that is where the cost will probably get more and more expensive than going to something like standard semantic search or knowledge graph or light rag or whatever other tool is out there for that. So here you can see if you have hundreds of pages with good indexes, you're fine with Wiki graph. But if you were getting up to the millions of documents, then you're going to want to actually do more of a traditional rack pipeline. At least for now, with how the current models are and everything we know right now in April 2026. So, that is going to do it for today. I hope you guys learned something new or enjoyed the video. And if you did, please give it a like, it helps me out a ton. Now, after this video, if you're interested in learning how you can create your own sort of executive assistant and then plug it into this Obsidian vault, then definitely check out this video up here, where I go over how I built my executive assistant and the way that you should be thinking about it. So, hopefully, I'll see you guys over there, but if not, I'll see you in the next one.