# 156 Build Karpathy's Second Brain With Obsidian + Claude Code

# 156 Build Karpathy's Second Brain With Obsidian + Claude Code Speaker 1: So Andres Karpathy put out a tweet a couple of days ago that went mega viral talking about how you can build your own second brain using Claude Code and Obsidian. And in the next 20 minutes or even less than that, you're going to be able to build your own. Nick Spisak is going to walk us through it. So Nick, from a high level, tell me what we're going to do today. What are we working with here? Speaker 2: Yeah, so basically what we have is this is the concept of the second brain. So essentially what this will give you the ability to do is to organize any of your data that you have in a way to make yourself more effective. That can either be for your personal needs, thinking through what are the items that you do on a day-to-day basis, or it can be more aligned to business practices. For us specifically, we're using Claude projects for a lot of the content that we're creating or the process of finding good content to put out on social media. This is a good practical use case to leverage this concept that Andre put out over the weekend, which is known as the second brain. Speaker 1: We're going to build it right here, right? Speaker 2: Yeah, we'll build it literally as part of this episode, start to finish, how to get this system up and running. Speaker 1: All right, let's do it. Speaker 2: So basically Andre was one of the founding researchers at OpenAI. He's one of the top most influential people in AI right now. So anytime he puts out a tweet, it goes super viral. So on April 2nd, he put out a tweet around how to build your own LLM knowledge base, and he breaks it down step by step. And as it went over over the last couple days, I saw his original ex post that was out there and built one as well that kind of talked through the fundamentals of how to build your second brain. Later that same day, he put out a follow up tweet that talked through this concept of an idea file. And this is really fundamental because today we're talking about second brain, but this concept can apply to anything. And the idea file was around creating your own LLM wiki, and he made it intentionally vague. He broke it down to describe what was the outcome or the core benefits that you're looking through based off of an idea, some loose-ended architecture of how you could build this type of system, Ultimately, what the operations of that system should look like, how you can go through to do indexing and logging to make it more useful. In the future, and then some optional command line tools that you can add to enhance the system to make it work based off of your needs. So what we did here is over the weekend, took that updated version and literally recreated the concept of the second brain, but this time as a full end to end workflow that's built as a Claude skill. So what we're going to do is we're going to literally walk through this process. Step-by-step and how to create a second brain utilizing the tools that Andre referenced loosely in his idea file. And some of the ones that we're going to use for today are going to be Obsidian, which you'll see really quickly here is a way to take any of your unstructured files and to put them into essentially a mind graph where you can see the connections and the patterns of how they're interrelated and how they work together. We're going to utilize an example of Obsidian Web Clipper, which is literally a Chrome extension that whatever page you are on, you can choose the location of where you want it to be stored in your Obsidian vault and click Add to Obsidian and it will grab or scrape not just the text, but also the images that are there. Optionally, you can install what is known as Summarize. This was put together as a open source library that Peter, the creator of OpenClaw has written. And what it does is it allows you to extract YouTube videos. You can get transcript files that are put into LLM friendly markdown to be able to use for the same purposes. So not only are we doing images and text, but we can actually grab YouTube transcripts from your favorite podcast creators and authority figures of the topics that they're looking for and have these registered into your vault. Speaker 1: And Nick, just before we keep going, so you that skill that you said that you built. What does the skill do? And I know we're giving it away to the folks who are in the audience here. So if you guys are, I'll let Nick kind of explain the skill. And if you guys want to grab it completely free, you can go in the description or in the show notes. And it's just like a one-click download. What does the skill do? And then we'll jump into the actual build. Speaker 2: Yep. So the skill itself is based off of Vercel's framework. So if you're familiar with Vercel, Vercel is essentially a framework that allows you to do web hosting. And they also put out a ton A ton of open source software. So we're using their framework to be able to build an end-to-end skill for all of the common AI harnesses that are in the industry right now. Claude Code, Codex, Gemini CLI, OpenCode, Pi, whichever your agent harness of choice is. The idea here is this is an end-to-end Skill that you can go ahead and add directly from the repository and it's going to give you the ability to set up a new vault through a guided wizard to ingest any new sources into your wiki. To query against it so you can ask a question, whether it's you want to do competitor analysis or you want to find out about a new topic or a new angle to your content. And then ultimately what's called Lint, which is to health check your wiki. So it's a way to be able to prune or remove any outdated articles that you may have in there or to look for additional connections that you can use in your Obsidian vault. Speaker 1: Awesome. Yeah, that's super valuable. So again, show notes or description if you're watching on YouTube, go grab that skill for free. Now let's dive in and actually build this thing. Speaker 2: Okay, perfect. So for this demonstration, what we're going to do is we're going to fire up Claude Code. And any time after you've already done the install here, you will see, and the readme will walk you through the setup guides. This is literally the one-click install setup for doing second brain. And what that will allow you to do is you can literally just hit the command second brain, which will run the skill. It will take a little bit to do here. And it's going to ask you, okay, well, what do you want to call your second brain? So for this, we'll call it YouTube demo. And that's going to give the actual vault its name. It's going to ask us, where on your local computer do you want to have this stored? So for this example, we're going to store it on our desktop. All right, so it gives us a pathing of where our vault's going to be stored. And for this, we'll just call it AI research is the topic. So you can actually give the knowledge base itself a domain, just so it has some additional tags or data associated with it. So it gets into the ability of having a context around what this vault is used for. So now it's going to ask me if there's any other config files that I want to add. So for right now, I'm going to go ahead and see. It says I'm on Claude Code. Do you want to generate a Claude MD for this? Are there any other agents that you'd like to configure it for? So we're going to say Codex. And basically what this means is Claude Code has a Claude.md file for its configuration. The other agent harnesses that are available, they use what's known as an agent.md. So this is just going through and literally walking you through any of the additional setups you want to use. Now, the next step is it's going to say, is there any optional tools? So I already have Summarize, which will allow us to summarize and grab transcript links through the command line interface for YouTube. QMD is for searching over large knowledge bases. This was a tool that Toby, the CEO of Shopify, has created and it's really good. And Agent Browser is an alternative for web research and is made by Vercel. So I already have all three of these, so I'm just going to say already installed. And for the viewers, some folks will have other tools in their toolbox that they would prefer to use over these three. These are just ones that I have. This is fully customizable. You can take this exact same GitHub repository, put it into Claude Code, customize it to your own unique needs and have it build a system for your second brain that's more personalized to you. Speaker 1: Now, Nick, super quick question. So obviously the setup wizard makes it like brain dead simple to spin up as many second brains as you could ever want. I'm assuming you'd want to kind of specialize your second brains, right? Is it not very efficient to have just one master second brain for everything and anything you'd want to store information on? Anything from personal to fitness to AI to work. I assume that's inefficient, right? Speaker 2: Yeah, there's definitely a balance between how many brains you want to manage where you're just doing maintenance of all the different vaults. Making sure that the data that's in it has the right relationships and connections. So for right now, the spot where I'm kind of delineating this is personal and business at the moment, that may be subject to change. And the beauty of this is once you have it set up, it's just a matter of being able to follow the same process of taking your data, putting it into what is called the raw directory, right? This is part of that idea file that Andre laid out. And all you need to do is then run the ingest command, which we'll show here briefly, to be able to put it into the appropriate structure of the wiki. So right now, I'm using a lot of cloud projects, especially for our content strategies, going to be using this more and more for Obsidian. And the next tier really after these kind of like personalized second brain or knowledge bases is what is known as RAG. And that's kind of more of the advanced tier where you get into much larger data sets, more complexity that's involved. So for the users there, that's a good natural path as you go from projects or chatbot type knowledge bases into an obsidian vault. And from there, you can graduate into some of the more advanced capabilities when it's necessary with scale. Speaker 1: Got it. Okay. That makes sense. Speaker 2: All right. Now, so the installer has gone through and it's now just giving us a reference point here. This is the Chrome extension for the Obsidian web clipper. I've already installed this, but if you go to this link, what this will do is it's literally just going to take us out to the Chrome web store. You'll click Add to Chrome. For me, I already have it installed, so it says Remove from Chrome. And what this will do is you'll be able to grab this Chrome extension. I can't clip the existing page here, but I can go to the one that we're going to use for our demo. And you can right-click on it. And click on options. And the options are any of the customizable settings for Obsidian that you'd like to use. The main one that I've changed from the defaults is under, literally under default, there's a section of where do you want in your Obsidian vault to store. So this normally says web clippings on it. I switched it to raw to align to the idea That Andre's laid out of these three tier structures of your raw items, your wiki, and then what he calls your outputs. Your outputs are essentially reports or things that you've generated that are essentially decision files based off of the content that came in from raw and then the wiki itself that came in as part of what you're searching or querying against. Speaker 1: And that makes sense because the raw is basically like your brain dump. It's like that's just where you dump anything and everything that you think you want in your second brain in some capacity. I'm assuming the wiki is when the AI goes in and like organizes it and kind of trims it and makes it digestible and then outputs is anything that where I guess you're querying your second brain and it's going and looking through the wiki and saying, well, hey, based on what's in the wiki, here's what you're asking for. Am I kind of understanding that correctly? Speaker 2: Yep, that's perfect. All right. So here, now that we have Obsidian, I already have it installed. You'll see a similar type experience when you load it for the first time. So normally you'll do a create vault, but given that we've just used the wizard to do that creation process for us and lay out the whole scaffolding of the setup, all we need to do is open a new vault or open a vault. So we saved ours to the desktop. So let me go to desktop and there should be a YouTube demo folder here. So we'll click open and what this will do It will create a new vault file for us. So you're going to see here, right, we have our outputs folder that's empty. We have our raw with assets that are currently empty, and we have our wiki, which lays out the scaffolding, the concepts, the entities, the sources, so all the metadata that you need. And then we have our agent.md file and our Claude. And this is to make it agent agnostic. So we can use any of our AI agent harnesses to be able to manage the vault that's there. Now, what you're going to see, and we're going to use this in real time, is there's this concept of a graph view. And this, you can see that these files right now are currently all, there's no groupings with them, right? They're all at this point, there's no relationships. So they're categorized as orphans, which you're just seeing that these are the four files. They exist in the system, but there's no, there's nothing that you can do really with them yet. They just exist there. There's no interconnectivity between the data that they're providing. So the first thing we need to do is we need to put some data sources in. So for this example, I'm literally just going to go to the Wikipedia entry for Andre that talks about him. And we're going to grab the web clipper. All right, and then we're going to go ahead here and we're going to click Add to Obsidian. And it's going to grab the article. So you can see now on this side, if I close the side panel here, it literally grabbed the Wikipedia entry, all of the content that's associated with it, and any of the sources that were there. So now we have an artifact. That we can use and inside that artifact, it's stored in the raw like we talked about and it's sitting right in there that we can utilize now as any of our brain dump thoughts. So the next step in the process is we literally just say second brain ingest, which is a separate skill that's in that repository. And what it's going to do is it's going to go through here and it's going to first check the directory. So we're in YouTube demo raw. It's going to read the files and find it. You can see that it actually identified that Wikipedia entry as a markdown file. It summarizes all of the key points that it learned in terms of its takeaways. And then I'm just going to say ingest. Speaker 1: Okay, I see. So it's like you can just go and basically, quote unquote, brain dump, meaning just Obsidian, Web Clipper, as many files as you want into the raw folder. And then what? Once a day, once a week, or just whenever you feel like doing it, then you go back into Claude Code and you run ingest. And that's when it's going to take everything in raw. And essentially read it or analyze it and then dump it into the wiki. Is that correct in terms of order of operations? Speaker 2: Yeah, that's exactly it. So basically what it's going to do, anytime you want to rerun it, now you could even get more sophisticated with this, right? And think of this from the overall ecosystem of AI agent harnesses. If you're on Claude Code, there's a concept of loop. Right? So loop allows you to do it on an interval and a prompt. Speaker 1: That's what I was going to say. I was like, couldn't you just set up it to ingest on a cron like every, you know, well, it depends on how often you're using it, but 24 hours, 48 hours or weekly ingest automatically. Speaker 2: Yeah. So you could literally have just a session running and you can see here on the right as this is going, it's building the graph. So before where we saw those orphan files that are there, now you're starting to see based off of the Wikipedia asset that we gave it, that now there is these interconnections of anytime we're talking about Andre, it's now going to say, okay, well, what is he known for? Being a founder over at OpenAI, his deep learning, where did he go to school at Stanford? So all of these connections are being built in this graph view. And this graph is essentially the interconnectivity of all the different raw data sources that we have. So the next tier of this as it finishes, Being able to index across the data that it found, we can actually now run the next stage of this, which is to actually try to make something useful, which is to answer a question around the vault or the second brain that we have. So we can literally say like, where did Karpathy go to school? And if it can find it in its knowledge base of that wiki, we'll get answers related to the content that's there. Speaker 1: And so to kind of bring this down to ground level, like from a business perspective, so let's just say that I'm a business and I have a team and I'm doing, you know, three, four, five, six Zoom calls a week myself. And then my team's doing a bunch of calls as well. Probably what I should do is set up my second brain to have all the transcripts from every call dumped into my second brain. So that way, if at any time, any place, If I'm like, oh, what did me and Karen talk about last week? I could just go into my second brain and query it and it would tell me exactly what was discussed without hallucinating, without me having to go dig through transcripts or even ask Karen. And that's also the most basic use case. I mean, this is, I feel like a company asset that compounds significantly over time. I mean, that's one thing that he talks about, right? Is that like, Day zero, this thing is kind of useful, but day 30, day 60, day 90, this becomes a hugely valuable company asset that compounds over time that nobody else is going to have except you. Speaker 2: Yep, that's exactly it. This is definitely the, when you set it up the first time, it's the dumbest it's ever going to be. Obsidian also offers the capability on their paid tier to sync vaults across devices. So let's say you're out on a walk and you're on your mobile device and you get an idea and it just needs to be flushed out later. You could utilize your mobile phone to quickly record a note of what you were thinking about, drop that into your raw, and then if you have Back on your desktop, you have a loop running, right? Maybe it's running every few hours, you come back and it's already been indexed into your vault and is ready to use. So this is just something that you incorporate into your day-to-day practices where it makes sense. And having those, maybe it makes sense to have them one for personal, one for business, since the topics are different, but each person's going to be a little bit different in how they set them up. But hopefully now we've done in under 20 minutes an end-to-end setup. This kind of gives you an idea of where you could use this capability and gets the ideas kind of flowing in terms of how it can help you in your personal and in your business. Speaker 1: Now in 30 seconds or less go through how you prune it because you mentioned that is one of the features because obviously over time data gets stale or maybe you have. We're going to talk a little bit about what's built in the wiki entries that contradict each other. So give us like the 30 second overview of how do we trim this thing and make sure it's always relevant and never stale. Speaker 2: Yeah. So literally it's a it's just another skill, right? Everything's built as a skill. So in this case, you're going to run the second brain Lint and Lint is just going to review what's already in the wiki and see if there's anything that conflicts. One article right now and clearly we just did it so it's gonna have it's gonna be up-to-date at that point But it'll do some research and ask you a little bit more about what it found if it's anything that's there So in this case, it didn't find any errors that were must error fixes It's just giving some different warnings that we can look at and say, okay Well, does this make sense? Is there any gaps that are there and that can give you the hint of like? Oh Well, that's great. Well, there's some missing items here around University of Toronto. I could literally just go out to the University of Toronto wiki page, use the web clipper, drop it into raw, re-index it, and now I potentially am closing some of those gaps. Unknown Speaker: That's insane. Speaker 2: You have that domain, right? You just pick a domain that you want to work in and you spend some time refining that. It's a great way to build ultimately what is the most important thing, which is how do you build your own private knowledge base or your own data set? And that's really the final boss of how to create a moat here in the AI age. Speaker 1: And that data set that you're building is your moat for any agents that you build, any skills that you build, the entire company you're building if you're an entrepreneur, etc. So just to kind of put a bow on it, I think this is one of the most significant things to come out of AI in the last few months, which is saying a lot because so much happened so quickly. Every company is going to have a second brain like this eventually. I think second brain as a service is a It's going to explode in popularity in terms of people offering this as a service to other businesses, like a $2,000 to $5,000 setup fee, and then a few hundred dollar a month retainer to maintain it for businesses. If there's not people out there doing it right now, they will be within the next week, I guarantee it. So this is crazy. Nick, this is awesome. For those in the audience, again, reminder, if you want the skill that Nick just showed us that literally sets all of this up for you, Get it in the description if you're watching on YouTube. If you're listening on the podcast, go in the show notes and download it there. It's completely free. Nick, thank you for your time and for everybody else. We'll be back in a few days with another episode. My pleasure.