Transcription
Using AI for personal knowledge management is so much simpler. Then I see other people are setting it up, and in this video, we'll show you proof. So you might have seen all these videos about using Claude with Obsidian, and I get why they're coming up because Obsidian has a following base and a lot of traction. And using this with AI now seems to be the right thing to do. And many of these videos are based on a post on X from Andrej Karpathy, also the inventor of the term vibe coding. Um, who came up with this LLM knowledge basis. And if you watched my previous videos, long before this post, I already showed you that we killed all the PKM tools out there that we've been using for years. This includes Obsidian, but also Heptabase, etc. All these knowledge tools have been replaced for my co-founder, Paco, multi-business owner, and myself, just using Claude code in a local folder.
And yet, even in this post, which is by the way, a great post summarizing exactly the things that I described, I wonder why there is still Obsidian in the equation. And we will break this down step by step in this video, and I will show you exactly how I'm doing it without Obsidian. This being said, there's no wrong or right. If you feel that Obsidian gives you the interface you want for your notes, that might be the reason. But you don't need to feel that you need to use Obsidian in order to make Claude work for you. And that's the misconception I see appearing over and over.
Alright, this post by Andrej got actually in the meantime, 17 million views, and all he shared here is a few things that you need to understand in order to use AI for your knowledge work. But if you watch my previous video, it's as simple as that. I even think there are things missing here. Something I'm finding very useful recently using LLMs to build personal knowledge bases. And we are saying personal knowledge management. Done. The era is over. The future is PKA, Personal Knowledge Assistance. And we will get to this point here too, because we consider personal knowledge management was always manual work. You had to make, create backlinks for yourself. You have to think about how things are connected, all these things. If you let AI do this work, it frees up so much mental power, so you can focus on the things that really matter and just retrieve the insights from your knowledge that really matter right now to you. That's what Andrej is describing here too. So he builds a wiki, he ingests data, like random articles that he finds, papers, reposts, data sets, anything that interests him, and he puts it into a raw directory.
If you look into my setup in VS Code, and now you can call me out, "Wow, but you're also using a tool." Come on, VS Code is really the basics of the basics that you can use. So I have my Claude code combined with my folder structure on the side. So if you are using already Obsidian, here is the folder structure on the side, and I have a team inbox here. Okay. That's what Andrej is calling raw, in fact. But I have a whole system. So if I move anything in here, so for example, here, if I scan anything with my paper scanner, any documents, they automatically end up in this folder, and my AI team will process this inbox and move it into my different areas in my knowledge system. PKM. There are the different things. So I don't need to have Obsidian for this to have this folder structure visually. But on the other hand, I also have Tom's inbox, which is the output of the AI. So whenever I work on something, I get it here for review.
Alright, he talks about a raw directory, which in fact to me is an inbox. That's the things where he's most, where he moves everything into it for LLM to processing later. So it compiles a wiki out of what it is in there. And then you have just a wiki. And the thing is, if you read through the whole thing, he's just referring to a specific knowledge base. So the wiki includes summaries of all the data, raw backlinks, and then categorizes to convert web articles into MD files, uses Obsidian web clip extension to get everything out of the web, including the images and all the things. Yeah, this is not tied to Obsidian. That's totally possible using just Claude, get it into a local folder instead of using Obsidian.
And we come to the interface now where he mentions the IDE here. He uses Obsidian as the IDE, which is the front, and it's the interface representing the data that you moved into this local folder where he can now view the raw data, the compiled wiki with the backlinks that you can click through this. And this is where people get hyped now that they are able to ask AI to create this inside Obsidian and then being able to click your way through. That's not something special. We look into this later in this video. The controversial thing here in this is that he says, "Important to note that the LLM writes and maintains all of the data of the Wiki. I rarely touch it directly." So what's the point of the interface then? An interface just representing the data would be more than enough. So if just for this reason alone, Obsidian is already overkill with all the input that you can do manually inside Obsidian. I've played with a few Obsidian plugins to render and view data in other ways, map for slides and so on. Again, this is where I would rather ask AI to build me something like this. And I show you in a moment my own setup instead of using something that is built by somebody else and modifies the local data in a way so it works with Obsidian only.
So there's a Q&A that he's doing. That's where it gets interesting. He now loads in all the information and can ask questions about it. A hundred articles, about 400,000 words. And I wonder, okay, if this is all you want, you can use NotebookLM, which gives you a lot more out of the knowledge. That's what I would use if I have a specific use case loading in hundreds of articles to get something out of it. Load into something like NotebookLM. Also, I would say the setup that I have. K NotebookLM. That's the RAG system that he starts then talking about in a moment. But you see here, we have 118 sources in there, including links to YouTube videos, articles, and many things, uh, articles on websites and so on. And then now I can ask anything about these articles, but beyond that, I can also generate videos like here. You ask a question, it creates a slide deck, but also spoken text and many other things you can do with it. Audio overview, slide deck, video overview, mind map, reports, flashcards, quiz, infographics, start the table. Also for free.
If you argue now about privacy, that might be a reason to use now Obsidian. But if you compare it, just loading in articles and asking questions about these articles, the output might be much better in NotebookLM than doing it inside a wiki-style Obsidian. That's just I would need to prove because that's not something I have tested yet. But okay, let's assume it's privacy and you keep it locally. This is a very specific use case. It's just adding interesting things of interest and then asking questions about it. It's a very basic thing. And then he says, "I thought I had to reach for fancy RAG." But the LLM has been pretty good about auto maintaining. Obviously, if you use Opus 4.6 model and you have a 1 million context window, this now takes a lot of articles way beyond the 400,000 words to lose the context. But what is there? There's no structure at all. So it needs to always look at the whole thing to get the connections and understands what you actually want from it. So therefore, each time it's very likely that you get a different answer to the same question because it starts from scratch. That's where it makes no sense to just simply load in loads of articles and things like this to build the knowledge base. In my opinion, a knowledge base that you really leverage over time is the conclusion that you make, the output that you generated. That should be the things that build your knowledge base. So in my opinion, it's much more efficient using NotebookLM or using something locally in a separate folder. You do this research and then you compile all these conclusions and connections. Enrich your personal knowledge system with these conclusions because now you're generating a very personalized knowledge system that has no noise in there, but only signal. And what is signal? Well, it is this what matters to you specifically because another person might need other insights from an article than you right now need. So I would never fill up my system with all these things without indexing and giving it actual structure and weighing off the context and so on to leverage this on time. And that's where I think that this all works great, as he says here, on a small scale. But I'm pretty sure the moment you start scaling this, you get diverse results. And again, the way we built the systems locally on our local folder, I can switch at any time from my Claude to Gemini to Jet GPT or whatever is now the newest thing and the best brain. And it doesn't touch anything on my local knowledge base. And that's important to understand in my opinion.
So now we get to the output. Instead of getting answers in a text terminal, I like to have it random markdown files for me. Yeah, great. Mark that. Yeah, same for me. Or slideshows or, whatnot. And all these views in Obsidian, and you can imagine many other visual output formats, blah, blah, blah. Okay. Yeah. So he's just saying that he likes to get this visualized instead of just having it in a chat box. And that's completely right. And I think now is the moment where I show you the basic setup of any knowledge base that I build up in the basic structure. This is the folder. I call it AI Team Blueprint for now. Right? It doesn't matter. What we have here, there are inboxes and there is a team inbox. And there's an owner's inbox. And what happens now when Andrej says that he has a raw directory? Well, this is the team's inbox. So if I have anything I want to share with my AI team, I move it in there. And if there's anything that my AI team has generated and created for me, it moves into the owner's inbox for me to review. Which doesn't mean it stays there because once I approve whatever was generated, I will then say, well, digit or store it inside the knowledge. And here you see there's an index.md file. So in this knowledge folder, everything gets indexed, and that's how you can make a bottom-up approach. Building this, and then when you hit a ceiling, you go for databases and then you hit a ceiling and you go for a RAG system and so on. While these are things that maybe most of you don't even reach, and you will be fine just using markdown files. But in the end, this is just a folder structure. And if I struggle to visualize this, it's no point. I have an MD file, I can open it up and read it this way. I can have markdown visualizers. And that's why I use VS Code. It's free and it's all you need to visualize things in a way you need. And you are much more flexible than using Obsidian because the difference to Obsidian is VS Code won't change your files. It just gives you access to these files, and you can work on these files. But Obsidian changes these files to get a database view and all these things. You need to make it exactly like Obsidian needs it. So you can only use Obsidian to visualize it.
And here you see now I just opened up this AI Team Blueprint folder that I just showed you here in Finder. Okay. We have in Finder here. There's the Claude or them, the Claude file. There's some commands in there. Uh, we have a Claude or MD how it works. This is just an empty blueprint that I made here. And if you want to get access to it, if you're free to join. But all you need to have is here VS Code. And now I'm in an empty folder. I could also open just an empty folder and get started this way. So I could now say, open folder, and I go to my desktop. I right-click, I say, new folder, test folder. Create. Open. Boom. Okay. Here's a test folder. And I showed you this already in previous videos. If you want to see how I use VS Code with Claude, you can watch this later on. I just want to make my point here that this is how you access your folder as easily. And now I see these videos where people use VS Code and Claude together with Obsidian, and that's the point where I really don't understand. Why would you do this either or, right? Because here I can now open Claude, work inside this folder. So whatever I do here, create a PKM folder, just an example, okay? So it knows it is right. Working inside this folder, you will see that here it is. Okay. It created this PKM folder. If we open this in the Finder again, is this PKM folder? See, we are in the, we are in the test folder on desktop, and it created this PKM folder. And that's how you can work with Claude and build whatever you like with full transparency and no additional metadata that is added to your markdown files and so on. So let's close this and show you my real setup. Okay? This is my per, that's my personal knowledge assistant system here, private. You see this goes way beyond just asking AI about my wiki knowledge because I want to get things done. I want to have, I have workflows that I need to get done regularly, and therefore I have a team of several agents that are all specialized in different things they do. I have SOPs. Okay. Standard operating procedures for the team. And you see this is not, has nothing to do with Claude, but it was generated by Claude. The way I say do it. So for example, YouTube extraction, okay, here it understands. Well, it's in German. I can show you in a moment our AI team working inside our business. Is it, which is set up exactly the same way, and you will understand this goes right way beyond just having a local wiki and asking questions. So I'm not saying that this is a bad thing to do, okay? I'm just saying it's overly hyped for what it is. And it is just scratching the surface while overcomplicating things by using Obsidian.
And now you could argue, yeah, but how do you visualize your information now? Yes, of course you can use Obsidian to do this and get stuck to it. Or you simply build a visualization of this for yourself using AI. And in fact, when we keep reading, he mentions that he does health checks of the wiki to see if it is the correct. And then he says, "Extra tools. I find myself developing additional tools to process the data. I vibe coded a small and a search engine over the wiki, which I both use directly in a web UI, but more often I want to hand it off to an LLM via Claude as a tool for larger queries." Here again, I wonder, why do you need Obsidian? Is it the knowledge graph? You can vibe code it. Is it, I don't know, because I don't like the visuals of Obsidian at all. And if I compare this to what I vibe coded in my thing here, where you can be very simple, just having an HTML file showing things in your browser, or you can even have an actual app created, which is very lightweight. Oh, shit. In comparison. Here's my interface that I build based on my wiki, and this is not just a wiki. There's a daily journaling application. There are all my PDF files, like invoices, contracts, all the things that I scanned in. I can access via here. Here are the different agents, um, that I have running. Here's a database explorer. There we go. See, this is a proper database where I can visualize all my data this way. Do I need a knowledge graph? I don't need it, but I could create it in minutes if I want one. Right? Uh, then here, journal. Look at this beautiful journal where I have everything greatly with Seneca, you know, giving me feedback on stewards about what I added there. Here is backlinking. I can click on Paco, it opens up Paco Canero. Here all the entries about Paco cross-linked in a beautiful way that I think is beautiful, and you can do it the way you think it would be beautiful for you, including metadata and all the things that you can have. And, uh, this is something Obsidian will never provide to me. I cannot have a mood track or things like that without making a very complicated plug-in setup and things like this that either need to rely on external developers. Here I can really build everything myself. And this, my friends, is a real personal knowledge base that I can leverage over time versus having a wiki where I just dive in and extract some information out of it. And then going it here is my life is in there. Uh, everything related to my life, finances, family, legal. And do I share anything with except with Claude? No, it's a local folder and it's backed up on Dropbox and all of benefits that you also know from what Andrej is sharing here.
So here is already vibe coding. I just think the layer of Obsidian is complete nonsense. And I see people riding the wave now because people on the cover, it seems like it makes all the sense. But in the end, if you want to really build a proper AI system where you start growing knowledge over time, persistent memory, and all this, you need to have a proper AI system set up and not just building a wiki to ask anything that you were able to do with NotebookLM or by the way, using something like projects inside your desktop version of Claude or in Cowork project. Okay? It's nothing different. You can upload in projects. I can create a new project, start from scratch, and I could now load in all the articles and so on that I'm interested in. This is a lot more friction. I agree. That's why it's great to have this. If you say that you're using Obsidian for the extension to extract the data from the website and so on, maybe that's worth it. I never needed it. I have Perplexity connected. I extract everything this way. And the point is, as I said in the beginning of the video, is it really worth having everything inside my system or just a signal that I need to make my own conclusions and build my own knowledge base in a proper way?
And he says, "Even as the repo grows." And that's where people get confused. "Whoa, why? He said, why do you say repo? We have a vault in Obsidian right now, man." In the end, it's a repo. He says, "A correctly right." And that's where you have VS Code again, where you have much better control over your repo. And that's, uh, something I show you in another video where you can now connect this knowledge base to GitHub and make it a synchronized repo. And you have access from anywhere, even from mobile using your Claude mobile application. So here is, for example, the team. This is what I'm using in our business every day. And if you go to the owner's inbox, you see there's a lot of things going on. And if I go to the archive, you see how long I'm using this already. So I started in February to build this up, and this is now automatically archiving. Whenever I finished a session, I say, "Close chat," and it reorganizes everything automatically. So this is really something we need to dive deeper into because here I can always go here. Boom. There are the different Q&As I ask for. Uh, here is the PDF, the MD files. I don't have need to show this because I can also visualize it nicely. Um, also, where is it? Maybe I find one. Here's an example of social images and, uh, slides that are generated, not randomly, but based on our design identity, the design language we have in the company. The whole team is trained on this, and I have specific agents who are designing the slide. We have QA persons reviewing the things and so on. And this is not overly complex. This is how you build a proper AI team working for you with a knowledge base. Because here is the business knowledge management. Here are, for example, references, reference assets. So for example, here, diagram references, negative and positive. You see here, that's something I didn't want to get. And here are examples of something that worked well. So this is how I train the team over time to create me better output over time. And this is the thing where the persistent memory keeps growing. They use this as an example. Or here, daily reports that we generate that get sent out via PDF files and so on. I don't need Obsidian for this. Look, I can visualize the images. I have full access to the files that are relevant to this. All the things. And this is a very specific setup here for our business. And I showed you previously that for my personal thing, I have a complete different setup. But here, for example, the business AI team is not only accessing this local information locally, right in a folder. It's also connected to our actual databases, to our membership platform. We have an AI assistant working inside the membership platform, answering questions because it has a knowledge about our nearly thousands of videos and articles and podcasts and coaching sessions that we had of with our clients that this knowledge, our AI is leveraging to give the best answer possible and create new articles, new videos, and things like this out of it. And this is exactly what Andrej is referring here to, that he might hit a wall when he gets bigger with the RAG system. And that's the case for us because our words are go more than millions and millions of words that we have there, obviously. And this is where you need a RAG system for AI to understand what you're looking for because also the moment you start having different entities, what about here? Maybe he's focused on one specific topic that he's researching about, but what about if you have 20 different projects running and you need to understand the different knowledges? I'm pretty sure AI gets confused very quickly if you don't, from the ground up, build the system properly to make AI understand where it finds the different things. And that's why there's no mention of an AI team here, and I think it's essential to have this from the get-go. Because if you go to myico.com and you go to team, you see exactly the team that I'm working with every day. It's me. It's our co-founder, Paco, and nobody else in the business. This is the crazy thing. AI completely replaced human beings except the two of us in this business. That doesn't mean Paco has four other businesses with over 70 people working. But for our particular use case here, this is just the two of us and the team. And you see here, this is Larry the orchestrator. And you know, I made this up with the animal faces and so on. This is something visual, but I like it too, to visualize my agents this way. And Larry is just an orchestrator. So this looks fancy here, but what it really looks like is this. I launch Claude, I hit slash Larry. And what this does, it loads in the instructions about who is Larry. Okay? Larry is the orchestrator. He's not allowed to do any work. He just needs to understand what I ask for, and then he delegates the work to the team members who are all in there. So therefore, there's an index. So Larry has the full understanding of all the capabilities the team has. And if there is no specialist to work on the request that I handed in, he will reach out to Nolan, who is our HR person, to hire a new person, uh, in order to do the work best way possible. This allows me now for the different workflows to optimize it to the maximum and make very specialized team members. So he has checks and he has, you know, he has his own PKM inbox, he has his own inbox, he has commands that looks very complex. Now, this is how you end up when you really go deep into this. That's not something I say that you need, but even if you would use Obsidian, you might end up leveraging a team like this and have it running in Obsidian. That's not the point here. The point is that people still don't get the point how to manage AI agents and get the most out of your knowledge base over time.
And as you can see, here's the Claude file and here are all these dedicated Claude files. However, everything is separate. That this is just a local folder with the team things in there. So this means I can now bring in Gemini and call Larry and it will work with just a Gemini brain instead of a Claude brain. And the rest remains the same. I can ask Larry now. And I say, "What's the latest comment in my iCore?" And here, you see, he's rooting to checks. He is the community agent. Okay? He knows that he's the guy who has all the understanding about the community. And if we go to the team, here is Jack. Okay? The community manager. And if we go back to the website, here is Jack. That's him. That's how we visualize him. He's 24/7 community manager and he has all who he is, why he joined the team, and all the things, the identity that they gave them themselves. And if you watch my previous video, you know that there's even an Easter egg here, pub night. And there you see they went out in London having a party together, and that's where I just told them, go out, have fun and come back with images from your party of this evening where they had a pub crawl from one pub to the other. And this is the fun thing behind this where I really think this is where it shows that a system is working because they really identify and everything is consistent over time. And now we're going back here. You see he's going now to Supabase, which is the backend of our community. Jack is now fetching this, and we see now here one hour ago. Here's the full comment. Here's the link. I can click the link and it opens up the community and scrolls to this comment. And here it is, two hours ago, there's this comment that we see. And you see I access the community and the comment, which is a whole knowledge base on a server elsewhere. Localized, accessible through my team on my local machine to work on this. Guys, this is what you can do when you really leverage AI in the best way possible for knowledge work that goes way beyond just asking a few wiki articles and creating some articles out of it. That's what is useful, no doubt, right. But when it really comes to saving money and being efficient and more productive, and then people say, "I'm not more productive with AI," and this is all they do, then you don't get the point behind AI. And that's what we are here for, to teach this. And obviously, no offense against Andrej or what he shared here. I think it's important that he brought this up and what is possible nowadays, but it goes so much more beyond without adding another layer of complexity, in my opinion, when it comes to using Obsidian on top of it.
I also saw this comment from Ola, "Why would I use Obsidian when I can't just use Claude code for a knowledge base? What's the advantage?" And I, I want to dive into this quickly too, because he asks absolutely the right question. But the comments below this is surprising because it seems people don't get the point. I see here, "Why do we use cars when we can walk?" I tell you, maybe it's better to walk for some people so they get to the destination instead of getting in a car and have no clue what the signs on the street means, and you build a crash. And then you can not drive a car anymore. So get your driver's license first before you get into a car. That's what I would say. And you rather walk to the destination. You understood the whole path towards this. So next time you say, "Man, but I would be faster using a car." And then you get a car. And I'm pretty sure this car won't be named Obsidian. But if you go here, you know there are so many people saying, "Yeah, perfect example. No, because he gets it right." This statement makes no sense or connection. "Why not use MD files in a directory?" Exactly. Why not? You can use it in a directory so that the folder becomes the car, not Obsidian. Obsidian is not the car. Obsidian is just the brand of the car, I would say. So you can use different brands of a car, but in the end, it's just getting from point A to point B. And you're not faster at all. This way, the compute, the comparison is just nonsense in my opinion. We see here, "Claude code is not a knowledge base. Claude code would be using Obsidian as a knowledge base, not you, so you can look at it and change stuff." No, I agree. Claude code is not a knowledge base. That's what something many people get wrong. They have a conversation with a chatbot and they think that's now their PKM system. That's nonsense because you don't want to search through your chats to find the information later on. However, Obsidian is not your knowledge base either. What is your knowledge base? It's the local folder that you've created. And as I have shown, you can open this folder in VS Code, and this is natively. Or you open it up in Obsidian, and it will alter your to the way so it can be shown inside Obsidian. And I think that's a big distinction here. But he's right. Obsidian is nothing more than the reader. All this has nothing really to do with Obsidian, but more about organization. And I agree with this. If you don't want to create an interface, you know, or let Claude create an interface for you and you want to have something, then Obsidian is the thing that gets closest to use. I just want to make you aware. You need to set up the files in a way so Obsidian can actually use it properly. "Obsidian is just permanent storage that sinks everywhere. Claude code loses context when you lose the tab. But yeah, if you're just dumping stuff in for a single session, Claude code works fine." What the heck? No, it is the local folder. Right. And what we need to understand here, that's correct. He refers here to Claude Chat. If you use Claude code, it always will generate a Claude.md file and things like this. So that's not even true. But you could say ChatGPT Chat, for example. And even there is persistent memory. This being said, I am using only Claude code and VS Code to visualize. And a local folder. I could even use only the local folder as I showed you in a one-hour video how I set this up, just using a folder and Claude code to build up this AI team from scratch. And the three agents that you really need to run the whole thing. And I'm not going deeper in this here, but if you really want to see proof that this works in a much more simple way, watch the video next. Then let me know in this other video if it is more or less complicated. What we see in here, you will see I literally just use a local folder and Claude code in a terminal, and this is it. And we have in the end of the video an interface showing you the knowledge in a beautiful way, in a way that you personally like it. And if you're interested to really master all these things, you can join us inside my iCorp, where we have monthly workshops and weekly coaching sessions with Paco and me, where we help you to set this up for your own use case and your businesses that go way beyond then just asking your knowledge base about specific questions. And actually start leveraging your knowledge to take actions on these things and actually make actual workflows in your business easier. So it's not about only knowledge management. It's, as I said, mentioned in this video, scanning documents, organizing these documents, but then also track record of, did I pay this invoice or not? How much expenses did I have from these invoices? It's accounting, man. All these things is something that probably most people are not aware yet is possible. And it's all possible on a local folder. That's the crazy thing that happened in the beginning of this year.
Alright guys, I hope this was insightful and let me know in the comments below what you think about all this. I'll catch you up in the next one.