30:59Nate Herk | AI Automation · Tutorial
Every Level of a Claude Second Brain Explained
A five-level framework for organizing knowledge so AI can actually find it — from a single CLAUDE.md to an always-on brain-OS.
June 17thThe argument in one line.
A markdown wiki that an AI agent incrementally organizes and cross-links turns scattered source material into a second brain that can answer questions no single document could answer alone.
Read if. Skip if.
READ IF YOU ARE…
- You already use Claude Code or a similar coding agent daily and want a low-effort way to make your past research, transcripts, or notes queryable.
- You're a solo creator or founder sitting on a large unstructured archive (videos, meeting recordings, docs) and want it to compound instead of rot in a folder.
- You're comfortable with markdown, local tools like Obsidian, and giving an AI agent free rein over a project folder.
SKIP IF…
- You want a no-code, point-and-click knowledge base tool - this requires running a coding agent yourself.
- You need multi-user or cloud-synced collaboration on the wiki, not a personal local vault.
The full version, fast.
The video demonstrates Andrej Karpathy's public LLM-Wiki idea: point a coding agent (here, Claude via Fable 5) at an Obsidian vault, give it a schema prompt, and let it ingest source documents into a growing set of cross-linked markdown pages with an index and a log. The creator ingests a PDF and a URL live on screen and the AI produces 20 new wiki pages that cross-reference each other, surfacing a genuinely useful catch (two AI labs' benchmarks aren't directly comparable). The setup takes about 5 minutes; the wiki's folder structure is emergent rather than fixed, staying flat for simple sources and growing subfolders for richer ones. The creator argues the model choice matters less for ingestion than for what you ask the assembled wiki to do afterward.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:13
01 · The LLM Wiki demo
Graph-view tour of the creator's existing YouTube-transcript wiki: backlinks between videos, tools, and concepts.
01:13 – 02:58
02 · What Fable does with the data
Contrasts a Fable-generated simple HTML explainer against a full day of manual iteration with Opus on the same underlying wiki data.
02:58 – 05:09
03 · Multiple wikis in the AIOS
Shows a second, flatter wiki (meeting transcripts) and a generated visual recap of the year using data pulled from across projects.
05:09 – 07:05
04 · Where this started + Obsidian setup
Credits Andrej Karpathy's public gist, installs Obsidian, creates a new empty vault.
07:05 – 08:12
05 · The setup prompt
Opens the vault in an editor with Claude Code, pastes Karpathy's gist, and instructs the agent to implement it as a full second-brain schema.
08:12 – 09:49
06 · Flat vs. structured wikis
Compares the flat Hercbrain (meeting transcripts) wiki to the subfoldered YouTube-transcript wiki, explaining the AI decides structure per source type.
09:49 – 12:38
07 · Ingesting two sources
Drops the Claude Fable 5 / Mythos 5 system-card PDF into raw/ and gives a URL, then lets the agent ingest both; result is 20 cross-linked wiki pages in ~10-12 minutes.
12:38 – 13:53
08 · Why it works: routing
Explains the index/log/CLAUDE.md routing rules that let the agent find the right wiki page efficiently instead of re-reading everything.
13:53 – 14:35
09 · Final thoughts
Reiterates that the wiki is just markdown with routing, so it isn't tied to Claude specifically, and points to a related second-brain video.
Lines worth screenshotting.
- A wiki built by an AI agent from raw documents is just markdown files with routing rules, so it isn't locked to any single coding agent or plugin.
- Letting a wiki's folder structure emerge from the content type, rather than forcing a fixed schema, keeps simple sources flat and lets complex sources grow their own taxonomy.
- Cross-linking two independently-ingested documents can surface a fact easy to miss reading them separately, such as two benchmarks not actually being comparable.
- The heavier-reasoning model isn't necessarily the right tool for bulk ingestion; the payoff shows up when you ask complex questions of the assembled wiki, not during ingestion itself.
- Feeding an AI agent emotionally specific instructions about the intended audience (e.g. 'a beginner who won't feel overwhelmed') can produce a more usable result than a technically thorough but generic prompt.
Cross-linking beats summarizing when sources reference each other.
WHAT TO LEARN
A wiki an AI agent maintains over time surfaces connections between documents that separate summaries never would, and the structure should be allowed to emerge rather than be forced.
- Letting an AI agent choose flat vs. nested organization per source type keeps simple archives simple and only adds structure where the content actually needs it.
- An index and a log of past ingests let an agent find the right page without re-reading the whole archive, which is what makes incremental growth practical.
- Cross-referencing two independently ingested documents can catch a mismatch (like comparing benchmarks run under different conditions) that reading each document alone would miss.
- Bulk ingestion work doesn't require your most expensive model; the higher-end model earns its cost when you ask complex reasoning questions of the finished wiki.
- Describing the intended reader's emotional experience (not overwhelmed, beginner-friendly) in a prompt can matter more than describing the technical spec of what to build.
Terms worth knowing.
- LLM Wiki
- A personal knowledge base where an AI model reads raw source documents and incrementally writes and cross-links a set of markdown wiki pages, based on an idea popularized by Andrej Karpathy.
- Obsidian vault
- A local folder of markdown files that the Obsidian app indexes and can visualize as a linked graph.
- CLAUDE.md schema
- A project-level instructions file that defines the rules a coding agent follows every time it works in that folder, here used to define how the wiki should be organized and updated.
Things they pointed at.
Lines you could clip.
11:00
“The connection that made this worth having as a wiki instead of two separate summaries: the two sources reference each other.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
00:00What you're looking at right over here are a bunch of my YouTube videos being ingested into an LLM Wiki. This LLM Wiki, as you can see if I zoom in, are different YouTube videos, and what's connecting them are different relations. So we're starting to see this actual kind of, like, second brain of all of my YouTube videos and and how they relate to each other, and all of this knowledge makes my AI OS so much smarter.
00:20And the coolest part about this is I didn't have to connect these concepts at all. I was able to just say, hey, Cloud Code, go grab my YouTube videos and then ingest them into this wiki, and this thing continuously grows and grows. If I zoom in a little bit, let's open up one of these videos.
00:33So right here, I've got nano banana two websites. When I open this up, we can see some information up here, but then as we scroll down, we can see, you know, summary key takeaways and other tools and things that are mentioned and other techniques that have been discussed. And I can follow all of these links around.
00:46Let's say I'm interested in GitHub. I can click on the GitHub. I can see what this is about, and then I can see other times that we've referenced GitHub.
00:52Here's some information that connects GitHub to Vercel. Why don't I click into that and learn some more about Vercel? Then Vercel can take me back to Cloud Code where I once again can follow all of these backlinks until I get to where I need to go.
01:02And so as this whole mind map of these YouTube videos starts to grow, we're able to see all of this come to life. And today, I'm gonna show you guys exactly how you can get up and running with something just like this in about five minutes. It's so much simpler than you may think.
01:13Now what's impressive about this isn't the fact that Fable was able to ingest all of it, it's what Fable can do once you've given it the power of all of this data. Because we all know that data is king, context is king. Here's a cool example.
01:24I asked Fable in one prompt. I said, hey. I want you to basically turn this messy blob of YouTube transcripts connections into something that people could actually look at and understand.
01:33I want this to be a simple resource that's not overwhelming, but shows my audience how these tools and techniques and ideas connect to each other. And now we have this super cool HTML, which I can click into, and I can see different ideas up top, agentic workflows, and what it connects to.
01:47It connects to routines. Routines connects to deterministic versus agentic automation, which connects back to n and n and Cloud Code, and all of this kind of stuff, and it's just amazing. In my mind, something like this is a much more user friendly interface than something like this.
02:01And what I think is awesome about this is that I was able to prompt it in an emotional way. I said things like, in a way that a beginner could understand and could click through and it wouldn't overwhelm them. And something like Opus 4.8 just doesn't understand what that means as well as Fable.
02:14To show you what I mean by that, this is something that I worked on with Opus for almost a full day. We went back and forth, we built this thing out, and I just didn't like it enough to share it with people because it felt overwhelming.
02:24It felt confusing. And the database on the back end that powers this is the exact same one. So anyways, what we're looking for here is the same thing.
02:31You can search through tools, techniques, videos. There's kind of like a layer framework that we discussed with the orchestrator, the models, the inputs, all this kind of stuff. And it has all the same data, and I can still click into these things, and I can follow the backlinks, and, you know, it's kind of the same idea.
02:45But once again, this version is just so much simpler to me, and I like it more. You can see as we click on a concept, we're able to see on this right hand side videos that it pulled this data from. We can read a little bit more about it, and we can see what else it's connected to.
02:57So that's just one very small example. If you guys have been following me for a while, you know that in my AIOS, I have a few different LLM Wikis. This is my YouTube transcript one.
03:04I've also got like my HercBrain one, which is pretty much where I put all of my meeting recordings. So all of my meetings, whether they're internal or external, I store them here, and that's how I'm able to see how the different concepts that I'm talking about with people have evolved and how they are going to continue to evolve.
03:18And when I'm scripting community posts, LinkedIn posts, writing emails, it takes all of this stuff into account because it knows everything about me and my business. So much so that right before this video, I said, hey, mister Fable. I want you to go ahead and just tell me a story about the past six months.
03:32So, you know, we're halfway through 2026. Build me a visual journey of what we've done so far in 2026. And this is what it gave me in one shot.
03:39It was able to pull this picture of me. It pulled our logo, and you can see that this thing is even feel and you can see that the branding of this even feels like AIS.
03:47It's kind of dark mode, blue graph colors, and this is what it gave me. It pulled actual stats like how many subscribers I gained. Um, it has our highest revenue month, which I'm gonna blur out, but it was able to look at all this data and just pull it for me.
03:59This was a big pivot I made this year. I went from pretty much doing only editing content to doing a lot of Cloud Code content, and you can see how this was able to pay off if we look at my average views and our revenue and how the business has grown since I made that pivot. Then we look at some other things like how our churn has changed, how our conversion has changed, other things about our revenue, but look at this.
04:17This is pretty funny. It pulled this different picture of me. If you guys remember the one up at the front was a smiling one.
04:23This is one of me thinking. And so it's able to just crawl through so much of the data and the resources that it has available inside of my perk two project. It shows the whole funnel of the business, which proves that it understands how people enter our ecosystem and all the decisions they can make inside of our funnel and where we try to push them to.
04:40Anyways, the point I'm trying to make there is the more data you give your projects, the better. But specifically, making sure that you route them in the right way, and that's what the LLM Wiki is really good at. This is my Herc two project.
04:51You guys know that this is my AIOS, and we have so much information in here. We have different Wikis, different projects, everything that I've worked on, and that's what you guys are trying to build.
04:59By the way, if you want to go through a full free course where I show you how to do that, in my free school community, link is in the description. I've got a full build your own AIOS course in there completely free. So links in the description for that.
05:09Okay. So here's where all of this started. Andre Kaparthi said, LLM knowledge bases.
05:13Something I'm finding very useful lately is using LLMs to build personal knowledge bases for various topics of research interest. He indexed the sources, which I'm gonna show you how to do, and then he uses something like Obsidian as the front end, which is what you guys just saw. So the first thing you wanna do is go to obsidian.md and then install this for whatever operating system you're on.
05:30So in my case, I installed this for Windows, run the wizard, get it set up, and open up the app. When you open up the app, it will look like this except for you won't have this stuff, and then you're gonna go down here and click on manage vaults. It might just pop up like this, and then you're gonna go ahead and create a new vault.
05:43So this one, I'm just going to call AI test, and then you're gonna choose a location for this, so it could be on your desktop. Or what I typically like to do is I put my vaults inside of my Herc two project.
05:55So my Herc two is able to look at a ton of these different little LM wikis that I have split up and separated by topic essentially. But for this example, I'm just gonna put this one on my desktop. So I'm gonna go ahead and create that Wiki.
06:07You can see here or sorry, not Wiki, Vault. We're gonna turn this into a Wiki. So right now, this is what we have, and then what I'm gonna want you to do is you're gonna go into Claude code or wherever you use Claude code.
06:16So in this case, I'm using Versus code, and you're gonna open up that vault in something like this. Alright.
06:22So I've just opened up the vault. You can see that we have a dot Obsidian folder. We have a welcome dot m d.
06:27That's just gonna be, you know, the default when you open up an Obsidian vault. And then what we're gonna do is open up Claude code. So however you like to use it, in Versus code, I like to use it in the terminal.
06:35So I'm gonna go ahead and run Claude to open this up, and then we're gonna go ahead and get started. Now one thing to call out here is we see that it says, okay, Fable, you only have it until July 7 on your limit. Otherwise, it will be usage credits.
06:47I did, however, see this tweet from Thorick that said, yeah. That's true, but we do plan to bring it back to part of your subscription as soon as possible. As mentioned in our original blog post, which if you see at the bottom of this blog post, it does actually mention that.
06:59So it doesn't say when, but hopefully they will be able to extend the window and bring it back as a standard part of your subscription plan. Anyways, what you're gonna do after that is you're gonna go to this page. I will have this linked in the description.
07:10It is Carpathi's LLM Wiki gist, and I'm literally just gonna copy this entire thing. If you want to stop and read it, feel free, but I'm gonna copy this entire thing. And what we're gonna do is take that back into our Claude and paste that in there.
07:22And then what you're gonna do is just go ahead and take a screenshot of this so you can paste it in. I said, you are now my LLM Wiki agent. Implement this exact idea file as my complete second brain.
07:31Guide me step by step, create the Cloud. M d schema with my full rules, set up the index, the log, define folder conventions, and show me the first ingest example. From now on, every interaction follows the schema.
07:42So I'm gonna go ahead and send that off. We're using Fable here. Like I said, you probably don't need Fable.
07:47Fable's probably overkill for this. It's about what Fable does after you have all that data in there. So if you wanna switch this back to Opus, run the ingest, and ingest future documents with Opus, that's probably a better call, honestly.
07:58Just gonna be showing you Fable in this video. Now what's really cool is as you start to put different stuff in here, it's going to sort of dynamically change the structure. So let me show you what I mean by that.
08:07If I open up this wiki, you can see that in the wiki, have comparisons. I have concepts. I have sources because this one is about my YouTube videos.
08:15So it's read through them, and it's analyzed that. We've got techniques, and we've got tools. But if I switch into something like my Hercbrain, which is the one that's more so around, like, my meeting transcripts, you can see that this is pretty much a flat structure.
08:26It basically just has all of my meeting recordings right in here, and it didn't wanna organize them yet. And maybe at some point, if we run some, you know, sweeps through, it will find some different folders to organize them in. But sometimes keeping this flat is actually better.
08:39And by flat, what I mean is basically just having everything in the wiki rather than having it drill down to even more folders. The reason being, you want to make sure that your AI can easily search through all the stuff.
08:49We have the raw, which is where you put stuff. Then the AI will read everything in the raw and ingest it into the wiki, and that's where it might take one source and split it up into, like, five or six or maybe even 10 little Wiki pages. Then we also have the index, is like a table of context.
09:03We have the log, and then the dot m d files are all of the other Wiki files. And this exact structure is how this Hercbrain one is set up. As you can see, it's very flat.
09:12But if we go back into my YouTube transcript one, this one's not flat. Right? This one has all of these other subfolders like we just talked about.
09:19And to show you what that looks like in this example, here is my index. You can see all the tools. You can see all of the techniques, everything that has been mapped out here with all the backlinks.
09:28And then the log, you can see that I did a few batch ingests here. I've done one there. And then every time that I've ingested another YouTube video or ingested other data sources, it will show a log of that there.
09:38The whole point of this is that the AI can incrementally build and maintain this Wiki. So it needs to be able to look at things like the index and the logs and all the backlinks to actually crawl around and find the data that you're looking for. Alright.
09:49So now you can see that this is done. Our project is set up. We have our index, which is blank.
09:53We have our log, which is pretty much blank, and then we have our raw folder and our wiki folder. So what you can see in here already is that in the raw, it processed this LLM Wiki idea.
10:04This file is basically the gist that Carpathi wrote up, so it decided to ingest that. And then in the wiki, it's already planned out to have concepts, entities, and sources.
10:14So that is what we have to start with. Now what I wanna do is we're going to ingest two different things, and I'm gonna show you different ways you can do it. So the first thing that we're gonna ingest is the Claude Fable five and Mythos five system card.
10:26So I'm gonna go ahead and download this as a PDF. What I'm gonna do is take this PDF, and I'm gonna drag it into the raw. So now that PDF lives right there in the raw.
10:33And then what I'm gonna do is we're going to take this OpenAI previewing GPT 5.6 soul, and we're gonna just do this as a URL instead. So what I'm gonna do is I'm gonna paste in the URL, and I'm gonna say, hey, Claude.
10:45Read this article and then ingest that into our wiki here. And then also, I dropped in a PDF in the raw called Claude Fable five, and I want you to also ingest that.
10:56And that is all that I'm going to tell this model. Once again, it should be understanding how this is set up.
11:02We should see a new record in the index as well as a new record in the log and some new sources in the wiki. So whether Fable decides to turn this PDF into one or five or maybe even 50 wiki pages because of how big that PDF was, same thing with the URL, I will let you guys know when this finishes up.
11:23Okay. So that finished up. It took about ten to twelve minutes, and you can see here that out of those two sources, it created 20 Wiki pages, and they're fully cross linked.
11:30Now look at this. The connection that made this worth having as a Wiki instead of two separate summaries, the two sources reference each other, and frontier model cybersecurity is where that lives. OpenAI benchmarked GPT 5.6 SOL against Mythos Preview, and I flagged the thing easy to miss reading them separately.
11:46OpenAI compared to the April predecessor not to Mythos five, and the two labs use different harnesses, so the numbers don't line up directly. So anyways, let's go ahead and pull open this wiki full screen and take a look. Okay.
11:57So this is what it looks like. You can see we've got OpenAI down here, and these are the ones that it relates to. Like the Claude Code said, it referenced Claude Mythos five in the article, so that's pretty cool.
12:07And we can see sort of the distribution here. We can see how much we've got things like government coordinated model releases. We've got layered safeguards, competitive use safeguards.
12:15And if we go over here to the wiki, we can see that we have concepts, we have entities, we have sources, and then we have topics. So the entities is cool because in here we have models.
12:25We've got Fable, Mythos, Mythos Preview, Opus 4.8, GBT 5.6. We've got Anthropic and OpenAI. And then if we go to the log, you can see this was the initial setup, and then we had the OpenAI article and the Claude Fable five system card.
12:38So the lesson here is that we now have this system where we have Claude code that looks at a bunch of our data sources. Right? It looks at the wiki, and it looks through potentially multiple wikis.
12:49And inside the wiki, what happens is there are routing rules set up so that our agents are able to figure out where to look for what specific thing, like the data that it's looking for. Because it has to crawl through basically all of this in an efficient way so it's not wasting our time and our tokens to find the right answer.
13:04And that theory is basically what Cloud Code is. It's basically figuring out how can my Cloud. MD work as a router to be able to look through my past projects, to be able to look through my business context, and find the right spot.
13:15And so from here, once you've already started to get the structure figured out, you're just gonna start adding more data sources, and you're gonna watch how this evolves, and you're gonna constantly check and see if it all makes sense. Because if you do a batch ingest and you don't like the way that it organized some of these folders and files, then maybe you go ahead and change that up a little bit.
13:32You know, you start to open up these pages like competitive use safeguards, read about it, click through, and see if it all still makes sense. And if you don't like how things are happening, then update the rules in the way that you ingest. Like I said, every LLM Wiki that I've set up, they have different structures and a little bit different rules because of the type of data that's in there, whether that's meeting transcripts or, you know, personal data or, you know, whatever it is that you're ingesting here.
13:55Make it make sense, Not only to the AI, but make it make sense to you. The whole point is that you could also go through this and follow the chain and find what you're looking for. And the greatest part about all this is once you realize, oh, look.
14:06Everything in this wiki, it's just a markdown file. It's just markdown files with routing. That means you're not locked down to using this only in Cloud Code.
14:12You can connect your Hermes agent to this. You can connect Codex to this. You can connect whatever you want to this because it's just markdown files.
14:18And if you guys wanna learn more about the whole idea of, like, building a second brain, then check out this video right here where I go over every level of building a second brain and how you know if you actually need to, like, move up a little bit or move down a little bit and figure out what's right for you and your system.
14:30So, anyways, thanks for making it end the video, and I'll see you guys in the next one. Thanks, guys.
The Hook
The bait, then the rug-pull.
Nate Herk opens on a graph of dots slowly resolving into a dense web of connections - his own YouTube catalog, reorganized by an AI agent into a wiki that links every video to the tools, techniques, and people it mentions. The pitch: this took five minutes to set up, and he's about to prove it live.
Named ideas worth stealing.
09:05concept
LLM Wiki folder convention
- raw/ (source dump, read-only)
- wiki/ (AI-written pages)
- index.md (table of contents)
- log.md (ingest history)
- CLAUDE.md (schema + rules)
The five-part folder/file convention the agent maintains so it can incrementally build and navigate the wiki without re-scanning everything each time.
Steal forany personal or team knowledge base built with a coding agent instead of a dedicated wiki product
How they asked for the click.
VERBAL ASK
04:55product
“if you want to go through a full free course where I show you how to do that... link is in the description”
soft mid-video mention of a free Skool community course, reinforced by the paid agency-playbook optin link in the description
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
