34:20Nate Herk | AI Automation · Tutorial
I Turned Claude Fable Into The Ultimate Second Brain
A 34-minute live walkthrough of one creator's AI operating system, built on the four Cs: Context, Connections, Capabilities, and Cadence.
June 10thThe argument in one line.
The structure of your second brain should be decided by the question you need to answer, not by the sophistication of the technology — and most people who jump to vector databases or knowledge graphs would get better results from a well-routed folder of markdown files.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code, Codex, or another AI coding agent and keep having to re-explain your setup at the start of every session.
- You have more than 30 notes or project files and your agent keeps missing context you know exists somewhere.
- You're debating whether to add RAG, a vector database, or a knowledge graph to your workflow and want a clear framework for when each actually pays off.
- You want to make your AI setup tool-agnostic so it works across Claude Code, Codex, and local agents without rebuilding everything.
SKIP IF…
- You're looking for a specific implementation tutorial — this is a framework video, not a step-by-step build guide.
- You already run a knowledge graph with LightRAG or similar and want to go deeper on that layer specifically.
The full version, fast.
Building an AI second brain is really a routing problem: can your agent find the right information, in the right format, without wasting your time or tokens? The answer is a five-level progression — from a CLAUDE.md with folder routing (Level 1) to a curated wiki (Level 2), semantic vector search (Level 3), a knowledge graph for relationship chains (Level 4), and an always-on brain-OS that syncs autonomously (Level 5). The key insight is that complexity does not equal better — most builders find that Level 2 handles 90% of their real pain, and jumping to Level 4 or 5 adds infrastructure they won't use. The correct level is the simplest one that eliminates your current pain point.
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 – 01:08
01 · Visual hook — three levels of graph density
Three side-by-side knowledge graph visualizations (sparse nodes → clusters → dense relationship web) establish the core idea before any verbal explanation. The presenter's real HERC2 project is shown as the source data.
01:08 – 02:31
02 · What is the actual job of a second brain?
Defined as: save info in a way that the agent (and you) can find it again. The key question is not 'can I store it?' but 'can it be retrieved?' Introduces the reverse-engineer-from-the-question mindset: design storage based on how data will be queried.
02:31 – 04:03
03 · Start with the question, not the tech
The basketball-hoop analogy: the shape of the ball must fit the hoop. Decide how you will query data before deciding how to store it. The slide reads: 'The answer you need → The smallest thing that gives it.' Pain is the correct prompt for climbing levels.
04:03 – 05:12
04 · The five levels mapped
Overview slide: Level 1 = find by exact word/name. Level 2 = pull everything on a topic together. Level 3 = semantic search (searched different words than I wrote). Level 4 = trace relationship chains. Level 5 = autonomous/always-on. 'Complexity climbs as you go up, not cost. Most people land at 1–3.'
05:12 – 09:29
05 · Level 1 — The Folder + CLAUDE.md
The foundation every second brain starts with. CLAUDE.md acts as a router: tells the AI which folder holds personal context, which holds projects, which holds decisions. When routing is correct, Claude stops asking for re-explanations and just knows where to look. Wall: grows too big and starts being ignored; only finds exact words. Move: one folder + a 20-line CLAUDE.md about you.
09:29 – 12:35
06 · Level 2 — The Curated Wiki
Adds an LLM Wiki layer: AI ingests transcripts/notes and auto-generates an index with summaries and cross-links. Shown via Obsidian visualization of YouTube transcript wiki (concepts, comparisons, techniques, sources). Also introduces memory.md and Claude Code's auto-memory toggle. Tool-agnostic tip: duplicate routing as agents.md for Codex. Presenter admits he personally runs at Level 2. Wall: must keep feeding it; a wrong summary loads as if it's true.
12:35 – 13:40
07 · My wiki has links — isn't that a knowledge graph?
Clarifies the distinction: wiki backlinks are 'see also' references; knowledge graphs have typed edges with explicit meaning (endorsed by, competes with). You can read the entire wiki page to follow a trail; a knowledge graph lets you traverse relationships without reading full pages.
13:40 – 19:33
08 · Level 3 — Semantic / Vector Search
Explains vector databases and RAG. Searching for 'feedback' in regular search finds exact matches; semantic search finds 'live test results' and 'evaluations' because they mean feedback. Diagram shows document → chunks → embeddings model → vector space by meaning (Company, Finances, Marketing clusters). Critical caveat: chunking breaks full-context queries — asking for a meeting summary may return the wrong five chunks. Markdown is better when you need the whole document. Best use case: looking up one specific rule from a thousand.
19:33 – 23:02
09 · Level 4 — Knowledge Graph
Slide title: 'Knowledge Graph: usually the skip rung.' Real questions aren't relationship chains for most people. Typed edges carry meaning (Jordan works at Acme; Acme endorsed by PostPilot; PostPilot competes with Cadently). Shows LightRAG visualization of presenter's real second brain — blurred due to sensitive business data. When IS it your level: VC tracking, recruiter CRMs, recurring cast of businesses/clients where relationship chains matter. Cost: free open-source software, but real work to build and maintain.
23:02 – 26:31
10 · Filling the knowledge graph — the Grill Me skill
The bottleneck for knowledge graphs is not the system — it is generating enough structured data. The 'grill me' skill (originally from Matt Pocock, available in the free AIS community) interviews the user relentlessly on a topic until the AI knows everything, then writes it to a brainstorm file. Aside: privacy warning inserted in post-production — data sent to Claude goes to Anthropic; consider open-source/local models for sensitive client data.
26:31 – 28:51
11 · Level 5 — Always-on Brain-OS (gbrain)
The only level that works while you sleep. Continuously syncs memories, runs 'dream cycle' to enrich notes. Based on Gary Tan's gbrain paired with g-stack. Requires a 24/7 server — real infrastructure burden. Windows users: use WSL2 or Postgres. For Q&A, Level 3 covers most of it. Only warranted when running agents offline vs 5,000+ page brains. Presenter currently experiments with gbrain only on a Hermes agent, not Claude Code.
28:51 – 30:00
12 · Find your level — stop at the first yes
Decision tree: re-explaining setup → Level 1; 30+ notes, forgetting what's in them → Level 2; agent whiffs on notes you know exist → Level 3; relationship chains across a recurring cast → Level 4; running agents offline vs 5,000+ pages → Level 5. First move: L1 folder + CLAUDE.md → L2 ask Claude for an index → L3 Smart Connections sidebar/memory. Climb only for a pain you felt this week.
30:00 – 30:59
13 · Team second brains + CTA
The team problem is 90% change management: getting people to update docs and pull from the brain instead of pinging colleagues. Recommends getting your personal setup working first before solving the team problem. CTA: free AIS school community, link in description, seven-day AI OS challenge.
Lines worth screenshotting.
- The job of a second brain is not storage — it is retrieval. If your agent cannot find the file, the file might as well not exist.
- Design the ball for the hoop: decide how you will query your data before you decide how to store it.
- Most builders who think they need a vector database actually need better routing rules in their CLAUDE.md.
- The CLAUDE.md file is not a system prompt — it is a router. It tells the AI which folder to look in for which type of question.
- Auto-memory in Claude Code writes your memory.md on its own — turning it on means your second brain grows passively between sessions.
- A wiki with backlinks is not a knowledge graph. The difference is whether connections carry meaning (endorsed by, competes with) or are just 'see also' pointers.
- Vector search finds things that are similar in meaning; markdown routing finds things by exact word. Choose based on the question you will actually ask.
- The meeting-summary problem: chunking a transcript into vectors means your agent may summarize the wrong five chunks and miss the key decision entirely — a single markdown file is more reliable.
- Level 5 (always-on brain-OS) is the only level that works while you sleep, but it requires a 24/7 server, real infrastructure burden, and is overkill for 99% of solo builders.
- Your second brain should only ingest evergreen data — ask 'will this memory be useful in a year?' If no, don't store it; give the agent access to pull it live instead.
- The same project can contain different levels in different folders — your meeting transcripts might be a vector index while your quarterly goals stay as plain markdown.
- Running a 'grill me' skill — where the AI interviews you relentlessly about a topic until it knows everything — is one of the fastest ways to generate structured knowledge-graph seed data.
- Tool-agnostic means one file per agent: CLAUDE.md for Claude Code, agents.md for Codex. Duplicate the routing file so the brain works across harnesses without rebuilding.
- The team second-brain problem is 90% change management and 10% technology — the bottleneck is getting people to update their docs, not which software to use.
Build the simplest brain that answers your real question.
WHAT TO LEARN
The level of your second brain is determined by the question you cannot currently answer — not by what sounds most impressive to build.
02What is the actual job of a second brain?
- A second brain's value is not in storage but in retrieval — the test is whether the AI can find the file when you need it, not whether the file exists.
- The reverse-engineering mindset: decide how you will query data before deciding how to organize it.
03Start with the question, not the tech
- The shape of your storage must match the shape of your queries — building a sophisticated system before you know what questions it needs to answer is the most common second-brain mistake.
04The five levels mapped
- Complexity climbs as you ascend the levels, but cost does not — Level 4 (knowledge graph) can be free open-source software while Level 1 (CLAUDE.md) just costs a few tokens per session.
- Most builders find their real pain is solved at Level 2 or 3; jumping to Level 4 or 5 adds infrastructure that sits unused.
05Level 1 — The Folder + CLAUDE.md
- A CLAUDE.md with routing rules is the foundation of every level — without it, the AI will not know which folder to search, so even well-organized files become inaccessible.
- When routing is correct, you stop having to re-explain your setup at the start of every session.
06Level 2 — The Curated Wiki
- The most common upgrade path is Level 1 to Level 2: once you have more than 30 notes, ask the AI to generate an index page with summaries and cross-links so it can navigate by topic instead of exact filename.
- Claude Code's auto-memory toggle lets the AI write and update memory.md on its own — passive context accumulation with no manual effort.
- Tool-agnostic design means duplicating the routing file (CLAUDE.md for Claude Code, agents.md for Codex) so the same knowledge base works across multiple agent harnesses.
07Level 3 — Semantic / Vector Search
- Vector databases solve a specific problem — finding semantically similar content across a large corpus — but they break full-context queries like meeting summaries, where the whole document matters more than the most-similar chunks.
- One folder does not have to pick one level: meeting transcripts might benefit from semantic search while quarterly decisions stay as plain markdown.
08Level 4 — Knowledge Graph
- Knowledge graphs are the right tool when your questions trace relationship chains across a recurring cast: investors, clients, competitors. For project-based solo work, they typically add complexity without solving a real pain.
- The bottleneck for building a knowledge graph is not the software — it is generating enough structured data. An interview-style 'grill me' session is one of the fastest ways to produce that seed data.
09Level 5 — Always-on Brain-OS
- The only level that works while you sleep, but it requires a 24/7 server — a real infrastructure burden that only makes sense when running agents offline against a corpus of 5,000+ pages.
- Autonomous ingestion risks polluting the brain with stale or low-signal data; deliberate, controlled ingestion keeps accuracy high.
10Find your level — stop at the first yes
- The evergreen filter is the most underrated habit: before ingesting anything, ask whether you would still want that memory in a year.
- Climb only for a pain you felt this week — if there is no current friction, adding infrastructure creates work without solving anything.
Terms worth knowing.
- CLAUDE.md
- A markdown file at the root of a Claude Code project that loads automatically every session. Functions as both a system prompt and a routing table — it tells the AI who you are, how you work, and which folder to search for which type of information.
- agents.md
- The Codex equivalent of CLAUDE.md. An identical or near-identical copy of the routing file renamed so Codex's agent harness can read it. Keeping both files lets the same knowledge base work across multiple AI tools.
- LLM Wiki
- A folder structure of interlinked markdown files where an AI has ingested source content (transcripts, notes) and auto-generated an index page with summaries and cross-links. Related to but distinct from a knowledge graph — connections are 'see also' references, not typed relationships.
- Vector database / RAG
- A system that converts text chunks into numerical embeddings (points in high-dimensional space) so that queries can retrieve chunks by semantic similarity rather than exact keyword match. Retrieval-Augmented Generation (RAG) is the pattern of feeding those retrieved chunks into an LLM prompt.
- Embeddings model
- A model that maps a text chunk onto a coordinate in semantic space — chunks that mean similar things land near each other. The embeddings are stored in a vector database for similarity search.
- Knowledge graph
- A data structure where entities (people, companies, projects) are nodes and their relationships (works at, endorsed by, competes with) are typed edges. Unlike a wiki, the connections carry explicit meaning that agents can traverse as chains of reasoning.
- LightRAG
- An open-source library that builds a knowledge graph from unstructured text, enabling both vector similarity search and graph-based relationship traversal. Used as the Level 4 example in this video.
- gbrain
- Gary Tan's (Y Combinator CEO) personal brain-OS concept: a system that continuously syncs memories, refreshes notes, and enriches context autonomously — the 'always-on' Level 5 architecture.
- OTAs
- Abbreviation used for quarterly objective/project tracking files in the presenter's second brain. Functions as a living decision log organized by quarter.
- Grill me skill
- A Claude Code skill (originally from Matt Pocock) that runs a relentless interview session on a topic — asking questions until the AI has extracted everything the user knows — and writes the output to a structured brainstorm file for ingestion.
Things they pointed at.
13:40toolLLM Wiki (Karpathy-style)
23:02toolGrill Me skill (Matt Pocock)
25:45toolgbrain (Gary Tan, Y Combinator)
01:08toolHermes agent
01:08toolCodex
Lines you could clip.
02:00
“Your moat is your data. It's your IP.”
02:31
“A basketball hoop and a basketball. We know what shape the hoop is, and we know that the ball needs to go through. So why would we ever design the ball to be a giant square?”
04:03
“Complexity climbs as you go up, not cost. Most people land at 1 through 3.”
13:20
“My wiki has links, isn't that a knowledge graph? Not exactly.”
19:33
“Knowledge graph: usually the skip rung.”
27:18
“In a year, will it be good for me to have this memory in here? If no, it's just noise.”
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.
metaphoranalogy
00:00Today, I'm gonna explain the different levels of building your own AI second brain. You can see here we have a visual of three very different types of data. This one is where we have our context really starting to form, and we're starting to see some relationships, and we're starting to see some different nodes and entities form.
00:15And then as we continue to scale this up, add more knowledge, more knowledge, more relationships, we start to get something that looks a little bit more like this, where we have clearly different clusters. And inside of all of these nodes, we can see how they relate to each other.
00:26And then over here, we're taking all of those relationships a step farther, We're able to then start to see how everything really pieces together rather than just having files that sort of link back to each other. This is relationship mapping.
00:37And so really the idea of an AI second brain has blown up because we're all trying to get as much information out of our heads into our systems as possible. That's the true value. .
00:48It's your IP. But the process of organizing that into a system so that you can use it with a bunch of different AI models and so that it can actually recall things in a way that makes sense rather than just hallucinating or spending a bunch of your time and tokens trying to look through everything. That's the issue.
01:02So clearly, all of this is my real data, and this is what the actual project looks like. It is my HERC two project. I have a bunch of folders and files here.
01:08And at the end of the day, that's basically all it is. It is markdown files that are organized in a way that I understand and that my agents understand. And so, yes, I'm gonna walk you guys through what I have here and how it works, but I also have this other project where I'm gonna show you if you're starting from scratch or if you feel like maybe you're in between level two and three, how we can actually look at the differences and what it might look like to scale up your own systems and start to add context in different ways.
01:31So super excited to dig into this today, and I don't wanna waste any of you guys' time. So let's just start looking at these five levels and how they differ. Alright.
01:38So every level of a Claude code second brain, and I'm gonna be obviously kind of referring to Claude code a lot, but keep in mind, this can be used with any AI model. I use my second brain all the time with Codex as well. I use it with Hermes agent.
01:49This can be used by different agent harnesses because it's just files and folders. So what is the actual job of a second brain?
01:56A lot of people probably define this differently, but the way that I think about it is that it's a place for me to save notes, meeting recordings, ClickUp threads, stuff like that. I can save it there, and then it helps me basically ingest it and get it into the right spots so that it can actually find it later.
02:11And so that's really the thing to think about is can your agent find it again, and could you find it again? Because if the answer is no, then you probably don't have the right routing or folder architecture set up, which is what I'm here to talk about today. And one other sort of mindset thing that I want to get out there before we dive into these five levels is that you kind of have to work backwards.
02:31You want to reverse engineer based on the question. So this will start to make more sense as we get into it, but really what you should be thinking about is how do I want to use this data in the future?
02:40Because how it's going to be accessed and recalled determines the way that you put it in in the first place. For example, a basketball hoop and a basketball.
02:49We know what shape the hoop is, and we know that the ball needs to go through. So why would we ever design the ball to be a giant square? Because it just wouldn't fit through the hoop, so that would make no sense.
03:00So you need to start with the end in mind a little bit. Once again, I will show you exactly what I mean by that as we continue on. Because remember, we're trying to get to the point where your second brain knows everything about your business, about you, your relationships.
03:12It knows everything to the point where it probably can recall stuff better than you can because it has a better memory and it can search through things way faster than you can. So we've got five different levels to talk about, and they each kind of have different questions.
03:25So level one is, can you find the file or the info by looking for an exact word or name? Level two is can you pull everything on a certain topic together? Level three is I search for different words than I wrote.
03:36So semantic search. You're searching for meaning rather than an exact word match. And then trace relationship chains.
03:43Can you ask about topic x and then trace that all the way back to topic a? And then level five is just kind of making this whole second brain thing super autonomous to the point that you don't even have to think about it. And by the way, this isn't me saying that number five is best.
03:57I have some arguments about why I do not currently sit on level five. The point I'm trying to make here is each level is different, and you wanna find the simplest level or the lowest level that actually fits your needs. If you don't have a pain point in your system, then I don't really think there's a need to go experiment or develop a new sort of, you know, architecture.
04:16If there's not pain, then why create more? Okay.
04:20So level one is pretty simple, and this is where you always start. So you start with a claw.md, or if you're using codex or something, you would start with an agents.md. But you start with a claw.
04:29M d, which is kind of, you know, that gets loaded up. That's almost like the system prompt for that session for that project. And then you've just got a bunch of folders and files.
04:35But the key part there is the Cloud. M d is kind of treated as a router. So, yes, you've got some, hey.
04:41This is your role. Here is what's important, but you also have routing rules. If you ever need to find information about me personally, look in this folder.
04:47If you need information about our quarter one priorities, look in this folder. Because if you've ever had a point where you ask Claude to do something and then it asks you, hey. Can you give me more info?
04:56I don't know what you're talking about, but you know there's files and folders in your project, then you probably just didn't give Claude the knowledge to go look there. It's not just gonna go search your entire code base automatically.
05:07I mean, you wouldn't want it to do that because it's gonna waste your time and your tokens. So if it doesn't know if something lives somewhere, then it's probably not gonna be able to find it. So when this is properly set up, you will stop having to re explain things.
05:17You will talk to it, and it will just know where to go look and why. But the problems with this is that if it grows too big, it can start to get messy and feel ignored, and this is typically more of, like, an exact words type of search depending on the way that you route.
05:29So if I open up my example project here, let's open up level one. So in level one, what you can see, pretend this is its own Claude project.
05:37We've got a Claude NMD. So let me click into that. We can see here, it says, this file loads automatically every time you open Claude code in this folder.
05:44It is the one file that tells the AI who you are, how you work, and where things live. At level one, this file plus a few folders is your entire second brain. So here's kind of like that basic knowledge.
05:52And then right here, it's this simple, where things live. In the context folder, always true background about you and how you work, read this first. Projects.
06:00Decision log. And that's basically it. So right here, can see there's a context folder.
06:04We have an about me file, which you could grow. We have stack and conversations file. We have decisions.
06:09So this is a decision log where you can have your cloud.md always append new decisions and the dates whenever you make a big change to your project or to your life or to your business. And then we have projects. So this is where you could have a markdown file or even folders within the projects for all of your ongoing projects, all of your ongoing clients, whatever it is, however you wanna organize it, that's where you can have some projects.
06:29And you can even start to organize these things by dates if you want. So if you want to just have one that's for, like, May, and then you have all of those stuff, and you have one for June. The thing that I really wanna stress here with level one and the thing that I answer a lot in my community and in the comments is that there is not yet a standard way that has been proven the best way to set up your projects or your second brain, besides some of the most common things like your context and your Cloud dot m d and your, you know, whatnot.
06:52But the point I'm trying to make there is don't see what I do and think that that's the right way, or see what someone else you watch does and think that that's the only right way. All that matters is do you have proper routing in place, and does it make sense to you, and does it make sense to your AI?
07:09Okay. So let's say I have my HER two project right here, and I need to find something in here, but I can't ask AI for some reason. What I need to find is easy because I understand the drill downs.
07:19You know, I understand my base folders. And let's say I'm looking for the HTML slide deck I built for my ranking Cloud Code features video.
07:26I would come into here and I say, okay. Know that's a project, I'll go there. Within my projects, I've got another project for YouTube videos.
07:31I'll open that up. And now I know I made this video right here, May 30 Cloud Code top 50 features.
07:38In here, I have the actual tier list deck. And when I open that up, now I have the slide deck. And not only can I find it easily, but my agent can find it because it all makes sense, and I have routing rules?
07:47Real quick, guys. If you're watching this video, you're probably interested in building your own AI operating system. Lucky for you, I have a full free course on that in my free school community.
07:55The link for that is down in the description. Join the free school community. Hop in here.
07:59Take the seven day challenge. Build your own AI operating system and apply these principles into building your second brain, which will make your AI operating system even more powerful. So link's in the description.
08:07Let's get back to the video. Awesome. Okay.
08:09So that is how you start. Now as you move up to level two, you might be able to start to work in some things like the LLM Wiki, which is what I've got set up for a few different things. This is the whole Carpathi LLM Wiki, which I did make a full video about.
08:21If you wanna check that out, I'll tag that right up here. But this is when you start to have more files and and they start to take a bit of a different shape, and you wanna organize them together in a bit of a different way. So it could be really good for researching all on a certain project.
08:34It could be really good for, you know, a few of the ones that I've got set up is my YouTube transcripts all live in their own Wiki. I've got all of, my meeting transcripts that live in their own Wiki. So for example, this is the Obsidian view of my Wiki for all of my YouTube video transcripts.
08:47You can see here if I go to Wiki, you can see there's main concepts like Agensic workflows, AI coding market, context window. And all of these in here start to relate back to other tools and concepts and videos and stuff like that.
08:59So we've got the sources. We've got platforms. We've got context management techniques, and all of this was auto created by our Cloud Code when I told it to ingest this YouTube transcript into our Wiki.
09:10So I'm not gonna dive super, super deep into all of this right now, but definitely check out that YouTube video I linked. Now what else is cool about this is this transcript Wiki actually lives within my main HERC two project.
09:21So here's HERC two. If I go right here to other worlds and then I go down to YouTube OS and I click into the transcript wiki right here, this is what we were just looking at in Obsidian. We could see the concepts.
09:32We could see the comparisons. We could see sources, techniques. This is what we were looking at in Obsidian.
09:37So all Obsidian is is it basically just visualizes your markdown files. You see here?
09:42Wiki, concepts, comparisons, techniques. This is what we were just looking at. All we get now is we just get a visual view of all that.
09:48And so the reason I wanted to bring that up as well is because I think a lot of people obviously get pretty infatuated by that visual view. And obviously, started the video with that because I think that's what hooks a lot of people in. But all that really matters is can your system grab that and give it to you?
10:03If you are a visual person and you really want that view, then by all means, install Obsidian and set it up. It's super easy. But I'm saying that you don't always need that visual layer if it's not beneficial to you.
10:13I hardly ever open Obsidian, just to be honest, because I know that it all lives here, and I know that my second brain and my OS can find all of that. So, anyways, in level two here, let's look at this. It's very similar in shape to level one.
10:25It's just building on top of it because now we have our cloud dot m d, which starts to route to some other things because it routes to the Wiki, and it still routes to context, projects, decisions, but it's also routing to references and memory dot MD. So we're just starting to add a bit more of these routing rules inside of the Cloud dot MD.
10:41We can grow the context. We can grow the decisions. We can grow projects and references, and we can also start to get this idea of memory.
10:48And what's really cool about this is you can turn on auto memory in Cloud Code, and the AI will basically start to write this file and update it on its own. So you don't even have to think about it. If you come in here and you do slash memory, it'll say auto memory on or off.
11:00And if it's off, if you wanna turn that on, just turn it on. And now one thing to think about is I mentioned earlier that we wanna make our second brains tool agnostic. And this is one thing that's pretty specific about Cloud Code is it uses claw dot m d, and it uses this memory dot m d, and it keeps that updated on its own.
11:16So if you wanted to move this over to Codex, what you would do is you would, first of all, transition your claw dot m d. You'd make a copy of it called agents dot m d. As you can see here in my Herc two, I've got my if I scroll down, Cloud dot MD right here, and then I've got agents dot MD right here.
11:29And they're essentially the exact same file just so Codex can read this one and Cloud Code can read this one. But because Cloud Code keeps that auto memory, all you need to do is make sure you have that memory m d file and just tell Codex, hey. By the way, for memories, look in our memory dot m d file.
11:43It's all about the routing there. Anyways, just felt like that was important to throw out. But at a certain point, when you have these, you know, wikis, they do start to degrade a little bit because what's what's great about them is that they have indexes.
11:53Right? So when your AI starts to look in the wiki, it knows, okay. If the user's asking about agentic workflows, I'm probably gonna start here.
12:01And then from here, I'm gonna drill down and read this to see what else is important to them. Maybe they're asking about the WAT framework, and then I can drill into that. And maybe from there, I need to learn a little bit more about the Cloud NMD system prompt, and then I will drill into that.
12:14So there are relationships here a little bit, but this isn't the same as, like, semantic relationships or knowledge graph relationships that have more meaning. This is more about just actually following a trail and reading the page in its entirety. And I'll be fully honest with you guys.
12:30I pretty much sit my entire perk two project in this level, in level two, because this has been working really well for me. Like I mentioned earlier, I haven't felt a pain yet big enough to switch over to level two. And here's what I meant by that.
12:42My Wiki has links. Isn't that a knowledge graph? Not exactly.
12:46Because this doesn't have connections of how they are related.
12:50Like, this is endorsed by this or this has cron to here. These just have connections because it's like a a c also. It's like backlinks.
12:58So they're very similar, and, yes, they can achieve a similar effect, but it's still a little bit different. Anyways, let's take a look at level three, which is where you start to do things like semantic search. Whether you do that in Obsidian, whether you do that with Pinecone or Supabase, however you start to grab the actual semantic search, that is what level three is.
13:17And so just as a quick visual for you guys, let's take a look at this quadrant cluster of images. So every one of these vector points is an image. And what we see in here is the payload is stuff like the file name, the URL, the name of the author or the artist, and the URL, but we don't actually see, like, what's in the image.
13:35We don't get a description. So what we have to do is we have to organize these images by meaning or by similarity. So when I open up this graph and we start to visualize this stuff here, what you see is that we have this main image, these owls, these kind of like I I don't even know.
13:49It's a very trippy style, like hallucinogenic style. Anyways, then this one is kind of similar. Right?
13:54It's got those colors. It's got the paints. This one is also similar, but they're not the same.
13:59They just share similarities. And as we start to expand these more and more, we can start to get into different styles. So this one has, like, some creepy eyes and mushrooms or whatever.
14:08This one is kind of more down that fantasy lane. As we start to build out more of these relationships and meanings, we can expand and grow away from them.
14:15And so quadrant really just gives you a visualization here. I mean, it's a it it has clusters and vector store. But the reason I pulled this up as a demo is just because we start to see the actual relationships form here based on meaning.
14:27And that's what's important about semantic search is that we're no longer doing keyword matching. We're searching based on meaning. So here in my YouTube transcript second brain, if I go to the smart lookup over here, this is very different from just the regular search.
14:41So for example, if I search here for feedback, let's say, we're actually doing a match on the word feedback, and it's only showing me where that word actually appears inside of our second brain.
14:54But if I come over here in the smart lookup and I search for feedback, we are getting matches that have things in here that mean feedback. So live test results, Cloud Code skills, which was talking about evaluations and stuff.
15:06So there's a big difference between keyword matching and semantic search, you know, similarity matching. This one over here is saying x equals x, and this one is saying x is similar to x, y, and z. And so this all just goes back to vector databases.
15:18I've talked so so much about vector databases, so I'm not gonna dive super deep in. I've got so many resources on my channel. But, basically, what it is is we take a document, so let's just say a YouTube transcript.
15:29We chunk it up, and then each chunk is ran through an embeddings model. And the embeddings model puts that chunk of text onto, like, a three-dimensional space where space is related to meaning.
15:40And so it decides, okay. This chunk is about a company, so we're gonna put it up here. This chunk is about finances, so it's gonna go here.
15:46And we start to see these vectors form near other similar vectors. Now do you guys remember how I said earlier, like, you wanna think about how is the data gonna be used? What type of questions are you gonna ask?
15:56This is a reason why that's so important. So think about this. Let's say I put my meeting transcript of March 5 meeting into my second brain, and I put those in as, you know, vectorized chunks.
16:07So let's say when I vectorize that meeting, we actually get, you know, like, 20 chunks. It actually creates 20 chunks or however many that is.
16:15And then when I say, hey, mister AI agent, can you summarize the meeting on March 5? It will basically search for March 5 meeting summary, and it will pull chunks that are similar to March 5 meeting summary. And then even if it gets the right chunks, it's going to only summarize those five chunks.
16:31It's not able to look at the entire meeting summary or sorry, like, meeting transcript entirety, so it doesn't really know a summary.
16:38It might be missing a lot of key information. Now, yes, there are things you can start to play with there, like metadata and other things like that to make these results better. But at the end of the day, people kind of assume that a vector database was some magic solution where it could always pull back what you need, but that is very false.
16:53And I mean, think about it like this. Let's say we have a table, and we say, hey. Which week do we have the highest sales?
16:58Okay. The agent looks for highest sales. It maybe grabs this chunk outlined in gray of data, and then it looks like, okay.
17:05Week six here was the highest sales, so that must be the answer. But in reality, you can see week 14 was higher. Week 19 was higher.
17:11So when you need something that has actual full context, then you can't do the vector database chunking.
17:19That's where you'd rather just have a markdown file of March 5, and then all this agent would have to do is read that entire markdown file and then give you a summary. And that's just going to be more accurate.
17:29So in this project, if we open up level three, you can see it's very similar because you can still have context files, decision files. You can still have all that, and then you might identify, okay. Actually, this one specific unit of my business, maybe my YouTube transcripts, maybe I want just that to be a vector database, but I still want my context and my projects and my decisions to be marked on files.
17:49So another point I'm trying to make here is just because you have a second brain and just because you have a massive, you know, folder here with a bunch of folders and files doesn't mean that the whole folder needs to be one style.
18:01Doesn't mean that everything needs Graphrack. Doesn't mean that everything is just LLM Wiki. It means that you're able to decide based on the type of data and the way you use it, how can you structure this specific folder in the way you want it.
18:12So here we have a vector index folder, and we click on the how search works. It works by chunking, embedding, search, hybrid, reranking. There's some things you can get really, really nitty gritty on when it comes to semantic search.
18:23But what vector retrieval is really, really good at is looking at tons and tons of data, typically just like a lot of text, and when you need a very specific answer, something that's very similar. So if you had a thousand rules that you needed to store and you basically said, hey.
18:37Can you remind me what rule 17 was? That might be a really good use case for vector search because it's able to search for rule 17, pull in those chunks, and just give you a little snippet because it would be a waste of time and tokens for your agent to read the entire markdown file of all a thousand rules if you just needed rule 17.
18:54So that's kind of the difference there. Like I said, I've got so many videos on vector stuff on my channel. But, really, you could say, hey.
19:00To your Cloud Code agent, I have this data. Here's how I wanna use it. Do you think this would be better for now as markdown files, or should I do semantic search?
19:07Like, what would actually make more sense here? And it will help walk you through the way that you should actually set that up. So now I hope you guys are starting to understand why I said, you know, moving up on or I'm sorry.
19:17Like, moving up on levels, moving down doesn't necessarily mean better. It's all about figuring out what is the pain point with what you're currently doing and where would a different level help you out and fix that pain point.
19:27Okay. Now So let's take a look at level four. This is where we start to get into, like, knowledge graphs and relationship graphs, which typically are gonna be the most complex and sometimes the most expensive as well.
19:37If you're doing it on a certain platform, you could always use open source software, but, anyways, knowledge graphs. And I also wanna be upfront. I've played with these lot, but I do not actually use these on the day to day because I found out just other ways to use routing files and wikis that fit my needs.
19:52Now my work is very different than what a lot of your guys' work may be. Mine is very project based and it is very, you know, content heavy. I don't have a massive CRM to manage with a bunch of different businesses and clients.
20:04You know? And if I did, maybe a knowledge graph would make a lot more sense and it probably would. But typically, the cool part about that is if you identify that you needed a knowledge graph, let's say for all your projects, you needed you wanted to put all of this in a knowledge graph, the data probably already exists here.
20:19And that's the thing about building out these relationships in your knowledge graph is that the system, whatever software you use, is typically gonna be pretty good at embedding that and creating that, but the problem that you have to solve is you have to give it enough data. And so one thing that I really like to do is I like to have these brainstorm sessions as you can see.
20:35And what I do with these brainstorm sessions is I use a skill called grill me. So if you see here, I have a skill called grill me, which I originally got from Matt Pocock. I customized it a little bit.
20:44I'll leave the skill for grill me in my free school community. The link for that is down in the description. All you have to do is hop in here, go to classroom, click on all YouTube resources, and you can find all the skills and everything like that.
20:55But the skill, what that does is it basically just grills me. It interviews me relentlessly about a certain topic, and it creates a brainstorm file here. It only stops when it knows everything about it.
21:04So if you wanted to start building up a knowledge graph for all your clients and businesses, just say grill me about client a. Grill me about client b. Grill me about business a.
21:12And it would just ask you questions, and you can feed it files. You can give it stuff. You can feed it in transcripts.
21:16You can feed it in, you know, contracts, whatever it is. And that's how you can start to form a lot of data. Hey, guys.
21:22Me again. Real quick. I'm editing this video, and I realized that I needed to throw out one thing here, which is that, obviously, if you're putting all of this data and you're sending it all to Anthropic, to Claude models, then that's not private.
21:34So if you feel comfortable with that, that's fine. I am putting a lot of my data in there, and it is my business stuff, and that's what I'm doing. But if you don't feel comfortable with that or you, you know, don't wanna send client data, of course, you don't, then maybe you want to do that through open source models.
21:49And maybe Cloud Code isn't where you have the second brain that has every single piece of information about you and your business and your client's business. So the point I'm trying to make here is just this is what I'm doing. I'm obviously aware of the fact that my data goes to Anthropic when I process it through Claude.
22:04And if you guys are doing that, then you should also be aware of that. But there are other options if you can't do that. So how to throw that out there?
22:09I am planning to make a ton of videos here soon about local AI and open source models and all this stuff because it's a really, really exciting space that I think is going to start becoming bigger and bigger. So, yeah, keep that in mind. Back to the video.
22:21I think sometimes that's a misconception about how I got here and how people build their own AIOS or Second Brain is that they think the problem is the system not retrieving it great, which sometimes it is.
22:34But sometimes it seems like the bigger problem is getting everything out of your brain into the system. So before you blame AI, take a look at your folders and files and say, is this actually holistic?
22:44Is this does this have all the nuance that I have in my brain? Anyways, from there, when you open up level four, you can see that it's it's, you know, very similar still.
22:52We're just adding on a few things. You can see here we've added an agents.md, which which is the exact same as the cloud.md.
22:57And what else is cool is you can literally just reference inside of your cloud.md at agents.md, and then you can delete all this because this basically just, like, injects that file into here. But I just wanted to show that.
23:07But, anyways, you can see we're still following the same principles. We have a wiki. We've also added a knowledge graph layer.
23:12We've still got the same where things live with the routing. With all these just regular folders and boring markdown, but boring is beautiful. You can see that our memory is still here.
23:20It's starting to grow, and we just keep building on top of this. So one thing we added here, as you can see, was our knowledge graph folder. And so what happens here is we get different entities.
23:29Right? So, like, we can see, okay, Jordan is a person. Acme is a company.
23:32And then we can start to form relationships between all these things. So Jordan works at Acme. Acme is endorsed by Postpilot.
23:39Postpilot is a competitor of Cadently, and it starts to build out not only these entities, but it shows you how they're all related. And so that's why when I said that I really like using, you know, this, what's it called, LLM Wiki, is because I have enough of that feel of all these relationships because I've put so much time and effort into ingesting these in the right way and giving it context.
24:00The thing about this one is that it has to read every single file it wants. Maybe it was looking at AI video production and all it needed to know was 11 labs. It still would have read this entire file first.
24:12And so that's where sometimes the knowledge graph is actually more lightweight in that sense. And this is the example I showed at the beginning of the video where we have LightRag. And forgive me.
24:20I'm gonna have to blur some of this stuff out because this is, like, legitimately my entire second brain in our business. But as I really zoom in here, and this kinda slows down my computer because there's so much, but what you'll notice is that we actually start to get relationships.
24:32I probably shouldn't have done this with so much data, but you can see, like, we have this collaborates with that. We have this builds that. And so if I really started to open up all of these little, you know, circles, we could see what was going on and how they're all related.
24:46We could see that our seven day AIS challenge, it was provided from YouTube. It connects to the onboarding process of AIS plus.
24:54It was developed by Aiden, and so we can basically follow around these relationships as you see. And even though it's pretty much the same data that you see here in Obsidian, we're not getting that same level of relationships between these different entities. So anyways, if you guys wanna see, you know, a full breakdown video on something like LightRack or, um, Graphify or all the other solutions that there are out there for more of a knowledge graph relationship graph, then let me know.
25:15But that is kind of the difference there. So if you don't need those sort of relationship chains and you're not worried about that semantic type of relationships, then you probably don't need to use something like a knowledge graph. And then level five, we have more of the always on brain OS and something like gbrain.
25:30Gary Tan, CEO of y combinator, he created this thing called g brain, which pairs really well with g stack. But g brain is kind of the idea of everything we've talked about here.
25:39Wikis, routing, relationships, tools. But g brain has kind of that always on element because it is, like, constantly syncing and refreshing memories and adding more stuff. So adding in g brain to something like a Hermes agent would be really, really good.
25:53You could still do it in Claude code, but you'd have to handle those crons and get all that stuff set up, which is why I don't currently run g brain at the moment, but I have been playing around with it with my Hermes agent. So, anyways, the point here is that it's very similar to everything else we've just talked about.
26:05It's just having that auto updating feel, more of the autonomous always on feel. But I will say another thing that I kinda that kinda scares me about that is you have this whole dilemma of, you know, when do you have too much context, and when does it get to the point where it's actually doing more damage than it's doing good?
26:23And the reason I bring that up is because I am in complete control of what my second brain ingests. I will run a skill to go grab all of my meeting transfer from the week. I will say, hey.
26:32Here's something. Help me figure out like, help me brainstorm about this, and then let's ingest it together. And for me, I really like being in that control because in my mind, there's a big difference between a few types of data.
26:41If you guys remember in my, like, AIOS videos, I've talked about the four c's. So context, connections, capabilities, and cadence. And for the second brain, I mainly think about it as just these first two.
26:50So context and connections. And so when I think of context, that's stuff like, you know, what my business has done. So if I come into here, into my my second brain, and you can see here if I go to up at OTAs.
27:02So OTAs are basically just our projects for the quarter. And so here I can see all the q '1 ones. Right?
27:07I can look at all those, and I can click at them and see decisions that we've made and the statuses. And I can also see q two OTAs, so I can see what's going on here. And my second brain's able to see that because that has been basically, those are locked in decisions.
27:17This is what we're doing this quarter, and then I'm updating the statuses of that stuff. So that's like context. That's what's going on in the business.
27:23But when it comes to connections, if I go back to this, this is more of like the real data that isn't as evergreen. This is stuff that changes. This is like Slack threads.
27:32This is emails. This is customer data. And that type of data, you don't wanna ingest into a second brain because that's just noise then.
27:39Then you have to go back every month and like delete old stuff. So the way that I like to think about my actual second brain is stuff that I'm not gonna delete. This is stuff that is like, okay.
27:48In a year, will it be good for me to have this memory in here? Yes. Otherwise, it's just adding noise.
27:53So when you're adding data into your project, think about it like the context and connections. Think about if this is kind of like more evergreen, holistic data, or if this is more things that are gonna change next week.
28:04So you probably shouldn't pull it in, but you should make sure that your second brain has access to go grab it. So that way if I said to my second brain, hey. Can you just take a look real quick at what John and I were talking about last week about, you know, OTA number seven?
28:17It would first go to our OTA file, and it would search through there, and it it would try to find it there. If it couldn't find it there, it would look through the wiki, and it would look through meeting transcripts and see what we talked about there. And if it couldn't find it there, it would finally go to ClickUp itself, pull real data in from me and John's conversations, and see if the answer lived there.
28:32And so that in my mind is still a second brain because I'm able to ask a vague question, and the second brain knows exactly where to look in what order to find that real time data and then give me back the answer that I need. That's the question I ask myself is does this thing understand where my data lives and where to look, and can it give me accurate answers?
28:48So as far as finding your level, remember, your whole project doesn't fit into one level. Maybe this folder's level two. Maybe this folder's level four.
28:55Maybe this folder's level three. Here's some things to think about. If you are reexplaining your setup and you need to find things by exact words or files, look at level one.
29:03If you have 30 plus notes and you keep forgetting what's in them, look at level two. That's where you sort of, like, ingest them and get that wiki with relationships. If your project is just completely whiffing on notes that you know exist and your routing isn't working, then maybe you wanna look for something more like a semantic search that doesn't rely on an exact word level match.
29:19If you're looking for relationships and to be able to follow chains of questions and thoughts, then you probably wanna look for something like a knowledge graph. And if you're running agents offline and you've got so much data and you wanna sync up a bunch of Hermes agents together, then you probably are looking for something like level five, something like g brain.
29:33And And another topic that I get some questions about, which I'm not going to fully address in this video, but I will briefly bring up, is the fact that you are building your own second brain OS, so are other people on your team. The next question is how do you actually make sure that everyone's data is syncing together, and how do you have more of, your team's second brain?
29:51There's a lot of different ways to solve that. I think once again, it's not an issue of, oh, do we use Google Drive or Notion or GitHub or cloud plugins? I think the issue to figure out with your team is how do we actually make sure that we all habit shift so that this stuff is actually useful and not just noise?
30:07How do we make sure that process owners are updating their docs and syncing their stuff there? How do we make sure that other people are pulling from that rather than always just pinging the same people for questions and answers all the time? I think the adoption and the change management question is the bigger one.
30:21The tech and the way it actually functionally rolls out is a little bit less. But what I do know is that you getting set up with your own first and understanding how it works, how you should route, how you should make the decisions of where the data should live, that's the first hurdle. You can only solve the team wide problem once you feel comfortable about the way you run it every single day and then it works for you.
30:41That is gonna do it for today. Like I said, you guys can grab all the skills and everything that you need from this free community. The link for that's down in the description.
30:48I will also include the slide deck if you guys are interested in flipping through. So if you guys enjoyed the video or you learned something new, please give it a like. It helps me out a ton.
30:54And as always, I appreciate you guys making it to the end of the video, and I will see you all in the next one. Thanks, guys.
The Hook
The bait, then the rug-pull.
Three knowledge graphs appear side by side — sparse dots, then clusters with visible relationships, then a dense web where everything connects. Before a word is spoken, the visual argument is already made: there are meaningfully different levels of organization, and each one unlocks something the previous one cannot. The question this video answers is not which level is best, but which level solves your actual pain.
Named ideas worth stealing.
04:03list
Five Levels of a Second Brain
- Level 1: The Folder + CLAUDE.md — exact word/file search, free, no terminal
- Level 2: The Curated Wiki — topic-level retrieval, index + summaries, auto-memory
- Level 3: Semantic / Vector Search — meaning-based retrieval, RAG, embeddings
- Level 4: Knowledge Graph — typed relationship chains, LightRAG, entity traversal
- Level 5: Always-on Brain-OS (gbrain) — autonomous sync, 24/7 server, dream cycle
Each level answers a different retrieval question and adds complexity only where the previous level creates pain. Most builders land at 1–3.
Steal forDiagnosing why your AI agent keeps missing context and picking the simplest fix
02:31concept
Start With the Question, Not the Tech
Decide how you will query your data before deciding how to store it. The basketball-hoop rule: design the ball (storage format) to fit the hoop (the question you will ask). Applies to folder structure, chunking strategy, and graph schema.
Steal forAny time you are tempted to add infrastructure before you have a retrieval pain
27:18concept
Evergreen vs. Ephemeral Data Filter
Only ingest data your second brain should hold in a year. Slack threads, emails, and customer data belong in live-access tools, not the brain — they become noise. Evergreen data (decisions, project context, personal context) belongs in the brain.
Steal forDeciding what to ingest before running any intake skill or script
26:55acronym
The Four Cs
- Context
- Connections
- Capabilities
- Cadence
The presenter's framework for categorizing what a personal AI OS contains. For the second brain specifically, only Context (evergreen business/personal history) and Connections (relationships between entities) are relevant — Capabilities and Cadence live elsewhere in the OS.
Steal forStructuring an AI OS beyond the second brain layer
11:25concept
Tool-Agnostic Routing via Dual Files
Keep CLAUDE.md for Claude Code and agents.md (identical content) for Codex. Each agent harness reads its own file. Memory stays in memory.md and both files point to it. Same knowledge base, multiple AI front-ends.
Steal forAnyone using more than one AI coding agent on the same project
How they asked for the click.
VERBAL ASK
30:00link
“The link for that is down in the description. Join the free school community. Hop in here. Take the seven day challenge.”
Mentioned twice mid-video and once at close. Soft, non-pushy. Free community with 7-day AI OS challenge and downloadable skills/slide deck.
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.
