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.
An AI operating system degrades through four specific, nameable failure modes as it scales, and staying accurate comes down to five concrete habits: a router-style CLAUDE.md, scheduled self-audits, update automations, knowledge segmentation, and backtracking every mistake to its root cause.
Read if. Skip if.
READ IF YOU ARE…
- You're running Claude Code (or a similar agent) against a growing folder of wikis, indexes, and routing files, and you've started getting inconsistent or outdated answers.
- You manage client work inside the same AI operating system you use for your own business and aren't sure how to keep the two segmented.
- You want a repeatable, scriptable way to catch context rot before it produces a wrong answer to a real business question.
SKIP IF…
- You haven't built a second-brain / wiki structure for your AI tools yet — this assumes an existing, growing system, not a starting point.
- You're looking for prompt-engineering or model-selection tips rather than information-architecture and maintenance habits.
The full version, fast.
As an AI operating system grows, it fails in four specific ways: poisoning (a false fact gets confidently repeated), bloat (too much data buries the relevant part), confusion (missing or irrelevant facts get filled in with guesses), and clash (two contradictory data sources and the agent can't tell which to trust). The fix is treating context as two types — expertise (always-loaded rules, like a system prompt) and situational (pulled in just-in-time) — and running five maintenance habits: use CLAUDE.md purely as a routing table rather than a system prompt, have the AI audit its own project on a schedule, automate the recurring data pulls instead of doing them manually, segment knowledge into separate wikis/folders as each area grows, and when the agent gets something wrong, make it backtrack and explain why before you let it fix the routing.
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 – 00:50
01 · Cold open
States the most common question he gets — how his AI operating system is organized — and previews the five tricks.
00:50 – 01:44
02 · Live audit demo
Runs a custom OS-audit skill against an older version of his 'Herk 2' project in Claude Code while the rest of the video plays out.
01:44 – 04:36
03 · Four failure modes
Defines poisoning, bloat, confusion, and clash using a whiteboard tool, with a refund-policy example for clash.
04:36 – 06:44
04 · Expertise vs. situational context
Frames always-loaded 'expertise' context (the rule book) against just-in-time 'situational' context, using a principal/teacher analogy.
06:44 – 08:06
05 · Sponsor: HyperAgent
Sponsored segment for HyperAgent — cloud agent infrastructure used to build a 5-persona review council.
08:06 – 11:52
06 · Audit results walkthrough
Reads the finished audit report: routing integrity, index truth, freshness, bloat/duplication, and hygiene all flagged, plus a proposed fix list awaiting approval.
12:07 – 17:06
07 · Hack 1 — CLAUDE.md as a router
Shows his real 'Herk 2' folder structure and explains treating the root CLAUDE.md as a routing table to sub-folders rather than a system prompt.
17:06 – 19:09
08 · Hack 2 — Have AI audit itself
Explains the habit behind the audit skill shown earlier and how Claude itself once suggested splitting a single wiki into two.
19:09 – 20:18
09 · Hack 3 — Build automations to update data
Argues for scheduled crons (weekly Q&A, meeting transcripts) over manually telling the agent to go pull data every time.
20:18 – 22:53
10 · Hack 4 — Segment knowledge
Covers splitting wikis by topic as they grow, and how to structure a dedicated clients folder for internal vs. client-facing data.
22:53 – 23:59
11 · Hack 5 — Backtrack
Explains making the agent explain and prove its own mistake — where it searched, why it missed data — before fixing the routing.
23:59 – 25:03
12 · Closing thoughts
Frames team-level second-brain syncing as a people/habit problem rather than a tooling problem, and signs off.
Lines worth screenshotting.
- An AI agent doesn't intentionally lie — it grabs whatever's in its loaded context and presents it confidently, so a false fact anywhere in that context becomes a confidently wrong answer.
- Context poisoning is the easiest failure mode to fix because the fix is just verification: web search, cross-checking a live database, or a human-in-the-loop step when confidence is low.
- Context bloat isn't only about the context window filling up — it's the needle-in-a-haystack problem, where relevant data gets buried in irrelevant data the agent still has to sift through.
- Context confusion looks like poisoning but isn't: poisoning is a wrong fact confidently delivered, confusion is the agent filling a gap — missing or irrelevant data — with its own guess.
- Context clash happens when two data sources disagree because one is stale, such as an old refund policy and a new one both still living in the knowledge base.
- Expertise context is the rule book that should load on every single call — who you are, what your business does, your policies — the AI equivalent of a system prompt.
- Situational context is pulled in just-in-time, like a single customer support ticket from yesterday — loading it permanently would only add bloat, confusion, and clash risk for no benefit.
- A CLAUDE.md file works better as a routing table than a system prompt once a project gets large: point to where each type of data lives rather than trying to hold the data itself.
- A useful self-audit test: try to find a past deliverable in your own file explorer without asking the AI. If you can navigate to it intuitively, the structure is probably intuitive enough for an agent too.
- Splitting one master wiki into separate topic-specific wikis (e.g. YouTube transcripts vs. meeting transcripts) lets the agent search a fraction of the files, which is faster, cheaper, and more accurate.
- For client work, keep internal knowledge about the engagement (contract dates, pricing, scope) inside your main AI operating system, but keep client-facing deliverables in a separate repo you could hand the client access to.
- When an agent claims it doesn't have access to something you know it has, don't just correct the answer — make it backtrack, explain exactly where it searched and why it missed the data, then have it fix the routing based on its own diagnosis.
- Recurring data types — weekly meeting notes, a Monday Q&A, quarterly reruns — are good candidates for a scheduled automation instead of a manual 'go pull that data' request every time.
- Syncing a second brain at the team or department level is described as more of a habit and permissioning problem than a technology problem — the tooling (Drive, Notion, GitHub) already exists.
Context fails in four specific ways — and five habits keep it from happening.
WHAT TO LEARN
An AI knowledge base doesn't get worse randomly — it fails through poisoning, bloat, confusion, or clash — and staying accurate as it scales is a maintenance discipline, not a one-time setup.
02Live audit demo
- Run an OS audit against your own AI project periodically — a read-only check that flags routing errors, stale indexes, and outdated data without changing anything itself.
- Treat the audit's output as a proposed fix list awaiting your approval, not an autonomous cleanup — you decide what actually gets changed.
03Four failure modes
- Context poisoning is a false fact confidently repeated; it's the easiest failure to fix because the fix is just verification (web search, cross-check, or human-in-the-loop).
- Context clash happens when two data sources disagree because one is stale — retire or timestamp old policy data instead of leaving it live alongside the new version.
04Expertise vs. situational context
- Expertise context (identity, policy, goals) should load on every run, like a system prompt; situational context (a single ticket, a one-off lookup) should only be pulled in just-in-time.
- Deciding what's 'always loaded' vs. 'pulled just-in-time' is the single biggest lever for avoiding both bloat and clash at the same time.
06Audit results walkthrough
- A stale index doesn't just miss new data — it actively fabricates confident-sounding answers pinned to whatever date the index last updated.
- When you find a real gap in your knowledge base, list every specific broken query it would produce (routing, freshness, hygiene) rather than fixing vaguely.
07Hack 1 — CLAUDE.md as a router
- Use your root CLAUDE.md as a routing table — a table of contents pointing to where data actually lives — instead of trying to cram the data itself into it.
- One big folder structure that mirrors how you'd navigate it manually (and pushes cleanly to a single GitHub repo) is a legitimate, simple default — there's no single correct layout.
08Hack 2 — Have AI audit itself
- Make self-auditing a recurring habit, not a one-time build: ask the agent to review its own routing rules periodically and suggest structural changes based on your actual usage patterns.
- A good sign your knowledge structure is intuitive: you can find a past file yourself, by hand, in your file explorer, without asking the AI to search for it.
09Hack 3 — Build automations to update data
- Identify recurring, predictable data types (weekly meeting notes, a standing Q&A) and automate their ingestion with a cron instead of manually prompting the agent every time.
- Automating the boring, repeatable pulls means a missed manual step doesn't create a stale answer later — the data is already there next time you ask.
10Hack 4 — Segment knowledge
- Split a single growing wiki into topic-specific wikis (e.g., video transcripts vs. meeting transcripts) once you notice a distinct, recurring category forming.
- Segmenting knowledge narrows what the agent has to search through for a given query, which makes answers faster, cheaper, and more accurate.
- For client work, keep internal engagement details (contract dates, pricing, scope) inside your main AI system, but isolate client-facing deliverables in a separate repo you could hand off.
- Still give your main AI operating system context that a client engagement exists — segmentation isn't the same as hiding the relationship from your own agent.
11Hack 5 — Backtrack
- When an agent wrongly claims it lacks access to something, don't just correct the fact — make it backtrack: explain where it searched and why it missed the data first.
- Only after the agent proves its own mistake and states the root cause should you have it fix the routing — this produces a durable fix instead of a one-off patched answer.
12Closing thoughts
- Syncing a second brain across a team is treated as a habit and permissioning problem, not a missing-tool problem — the storage layer (Drive, Notion, GitHub) already exists.
Terms worth knowing.
- Context poisoning
- A false fact sitting inside an AI's loaded context that gets confidently repeated back as if it were true, because the agent has no way to taste-test what it's serving.
- Context bloat
- Too much data loaded into context at once, which buries the relevant information and makes it harder for the model to isolate what actually matters to the current question.
- Context confusion
- A failure where a fact is missing or irrelevant rather than wrong, and the agent fills the gap with a guess instead of flagging the gap.
- Context clash
- Two contradictory pieces of data in the same knowledge base — often an old policy and a new one — that leave the agent unable to tell which source to trust.
- Expertise context
- The always-on rule book an agent should load on every single run: identity, policies, goals, and tools. Functions like a system prompt.
- Situational context
- Data that's only relevant to a specific moment or query, pulled in just-in-time via a live lookup rather than kept permanently loaded.
- OS audit
- A read-only skill that checks an AI operating system's routing files and indexes against what actually exists on disk, then reports what's out of date without making any changes itself.
- CLAUDE.md as a router
- Using the CLAUDE.md file as a table of contents that points to where data lives, instead of trying to cram the data itself into the file.
Things they pointed at.
01:36toolOS Audit skill (free download)
07:00toolHyperAgent
01:36channelAI Automation Society (free Skool community)
Lines you could clip.
00:00
“The number one most common question I've been getting lately is about how I have my AI operating system organized.”
02:08
“The agent inherently is not going to intentionally lie. It's probably going to grab something and then use that as data, but the problem is the dataset was poisoned with an incorrect fact.”
04:11
“In March, your policy was always refund. In June, your policy is now to never refund. So when a refund question comes up, the agent doesn't know which source to trust.”
11:47
“Any post June 29 business questions would give me a confident June state answer, which would be wrong.”
17:31
“The only way that you're doing this wrong is if you're constantly getting wrong answers and you're not doing anything about it.”
22:47
“Have it backtrack... have it prove its own mistake and tell you what it did wrong and how to fix it. That works a lot better than just saying, make no mistakes.”
24:20
“I don't think it's a tech problem. I think it's a people problem.”
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.
00:00The number one most common question I've been getting lately is about how I have my AI operating system organized. Questions about how I should be routing files and how I should have wikis organized and folders organized and where do I put client projects and so many questions around the organization of it. Because the problem is if you don't have it organized, then your agent is gonna start to hallucinate things not only to you, but potentially in your skills and when you're having it build automations and stuff like that that could get pretty bad.
00:24So today, I wanna talk to you guys about these five different tricks that I've been using that have helped me keep my AI operating system super accurate, super up to date, and it allows me to add more and more data every week without sacrificing quality or memory. So let's not waste any time and just get straight into today's video.
00:39Alright. So here is an older version of my HER two project that I pulled in just to show you guys a quick skill that I built that I'll be giving you all for free. So look at this.
00:47I'm gonna go in here and I'm gonna run this OS audit skill. Now what's gonna happen here is it's going to look through my entire project. It's gonna read everything.
00:54It's gonna look at all of the routing rules, and it's gonna tell me all of the areas where there are things that are weak, where I need to make some improvements, and where I need to update some data, that kind of stuff. Essentially, the deliverable of this audit is an audit. It says, hey.
01:07Here are 10 things I noticed. Here are the 10 fixes. Do you want me to do them?
01:11Yes or no. It won't actually do anything yet. This is just kind of an exploratory phase.
01:15It will also create this folder at the root of your product if you don't have one. If you do, it's called audits, and then it will just chuck in a markdown file that tells you what it found and what the fixes are. And by the way, if you guys wanna get this skill for free, just go to my free school community.
01:27Link for that is down in the description. Come into here, click on classroom, click on all YouTube resources, and you'll find it in there. So while this is running, what I wanna do real quick is talk about why this is so important, and then we'll come back and we'll look at the actual audit.
01:38So what I wanna talk about today before we get into these five tips are the different methods of context failures. So four different failure modes of context and then two different context types. So understanding these six different things is going to make the audit make more sense, and it's gonna just help you in your day to day when you're using your AOS way more.
01:56So the first thing is the four failure modes, poisoning, bloat, confusion, and clash. Basically, meaning when your agent tells you something that you know is incorrect or maybe you don't even know. Maybe you find out later that it's incorrect.
02:06When it makes a mistake because of the context, there are these four reasons. So let's start with poisoning. Poisoning means you have a false fact somewhere in the context.
02:14So imagine this is the context. Imagine this is a false fact that gets dropped in amongst these green facts, are the right ones. The agent looks in there, and it will display that back to you or put it in an email to the customer or whatever it is because it was in the context.
02:28So the agent inherently is not going to intentionally lie. It's probably going to grab something and then use that as data, but the problem is the dataset was poisoned with an incorrect fact.
02:39Now luckily, poisoning is the easiest one to fix because, basically, it's just a matter of having some verification. So making sure that it fact checks everything with a web search or maybe two web searches or fact checks and cross checks across your live database. Or maybe if it's not a 100% confident, it has to just human in the loop.
02:54You know what I mean? So poisoning is the easiest to fix. Now let's take a look at bloat.
02:59Bloat is when there's so much stuff. There's just way too much data, and this is what I think a lot of you guys start to feel as you scale up your AIOSs and you start to use them more and more.
03:08Now this one is tough because as you can see, we all know about context rot as far as the window, but we also know about the idea of needle in the haystack. Right? The agent is going to look at the data that it's currently loaded in in order to make some sort of decision.
03:21And if it has way too much to look at, then it's gonna be really hard for it to actually pull out what's relevant and what's not, and there's just gonna be some stuff that bleeds in and you probably don't want it to. The tough thing about bloat is that it's a little bit harder to fix, but when I talk about expertise versus situational, that's something that's really gonna help us out here.
03:37So just hold on to that thought for a sec. So anyways, that's bloat. Then we have confusion.
03:42This is where there is something in the database. So it's a little bit like poisoning, a little bit like bloat, but it's basically an irrelevant fact or something is completely missing. This is more of the classic hallucination because it tries to fill in the wrong or missing data with its own, which is I know it sounds very similar to poisoning, but poisoning is more so it grabs a fact that is just inherently wrong and confidently delivers it.
04:03Whereas this one is it gets confused because of the facts that it sees and because of facts that it potentially knows are missing and it just answers instead. And then the last one is a clash, basically meaning you've got old data or you've got two different data sources that have clashing information, so it doesn't know which one to choose between.
04:20Very simple example. In March, your policy was always refund. In June, your policy is now to never refund.
04:25So when a refund question comes up, the agent doesn't know which source to trust, so maybe it will trust the old one. Maybe sometimes it trusts the new one. Maybe it just makes something else completely up.
04:34Now this next piece that I wanna talk about is expertise versus situational context. And this is also very important, and it relates back directly into these four things.
04:44So if you guys remember, when you think about the whole idea of your second brain and if you use my framework on building your AIOS, which is the four c's, where we basically split this up into having context, connections, capabilities, and cadence, and you think of the context and the connections as the second brain piece of your AOS, I think of context as expertise, and I think of connections as situational.
05:05So let me break that down so it, you know, actually makes more sense. So expertise context is the things that you actually need all the time.
05:12So, you know, who are you? What are your goals? What does your business do?
05:16That's expertise context. The way I like to think about this is an analogy of a principal and a teacher. Pretend they're trying to make a seating chart.
05:24The principal knows how classrooms should run. They know where the whiteboard is and where the, you know, the door is, and they understand how you should build a good seating chart. But the teacher has the situational context of every student.
05:36Which student has bad vision and needs to sit closer? Which two students are loud? And if they sit next to each other, they're gonna laugh the whole class.
05:42The teacher has that situational context that you load in just in time. So here, the way I think about this is expertise context is the rule book. Right?
05:50Your policies, your preloaded information that needs to be in every single run, sort of like the system prompts. Whereas the situational context is things that you need just in time. So for example, let's say you have this AIOS and you're curious about, you know, a customer support ticket that came in yesterday.
06:07You don't need that customer support ticket to always live in the context. There's just no reason for that. What would that do?
06:12Well, that would add in probably a bunch of bloat and confusion, right, Or maybe even some clashing too. So what you do is just in time when you realize you have to answer something about this customer, for example, like Thursday at 2PM, then you would go use that live lookup, pull the data in because you need it, and then you'd actually be able to enhance or augment the agent's response because it had real time situational context.
06:37So, hopefully, those six different concepts make sense. All of that contributes to this audit that we're currently running right now in my AIOS. And besides that, it just really helps you think about the way you prompt your agent, you organize your agent, and what you give your agent as far as, like, here's data you should always have in the AIOS, and here's data that you only need to grab every once in a while.
06:56Real quick, guys, a quick message from today's sponsor, HyperAgent. So a couple of videos back, I showed you the council that I built on HyperAgent, and enough of you guys asked for it that I actually put the whole thing up there for you to just go and grab.
07:09So a quick refresher, HyperAgent is built by the Airtable team, and it spins up a real cloud machine for every single agent. You get a full browser, a shell, and all the tools to go do actual research. So I use that to build a council of five agents each with its own persona.
07:24So for example, this week, I dropped a course idea into Slack, and then the council went at it. One played the skeptical buyer, one pulled what competitors already charge, and one stress tested the math. Then they came back to me with a verdict, where it was weak, the cheapest way to test it before I build anything, and a few things I honestly didn't wanna hear.
07:40Now you could go build this council yourself just like I did, but now you don't have to because I published my exact council on the HyperAgent Marketplace. So you can sign up, you can install it, and you can run your own idea through it in just one click. And my link gets you $1,000 in free credits to start, which is down in the description.
07:55So go get your ideas roasted, and let's get back to the video. Okay. So this audit finished up in about two minutes.
08:01Let me go to the top, and let's just read through this real quick. So here is the OS audit on July 22.
08:08It was my third run today. I was testing it out, making some tweaks. But, anyways, the knowledge here is current through June 29, so almost a full month ago.
08:16What we have here is the routing integrity is red. There's an OTA misroute, and up at OTAs are unmapped. We have index truth also red.
08:23Red being bad, green being good. Index has 55 folders, but disc has 79. Index has 52 rows.
08:29You can see freshness is a red, bloat, duplication is a yellow, hygiene is a red, context placement, red, a bunch of problems that we need to solve here with our AIOS. It tells us basically why it found all of this information and what would wrong answer you today. So any post June 29 business questions would give me a confident June state answer, which would be wrong.
08:48If I asked about how q three OTAs are going, it would miss routes, and it would give me wrong info. If I asked what's in the projects, it would give me stale data. So it's showing me all of these issues that are currently living because it did a full audit of our OS.
09:02And then it gives us a fixed list, which says await approval, but we could say, okay. Yeah. Let's go do a, b, c, and d right now to get everything back up to date and everything actually working the way that it should be.
09:12So look what it recommends. Finish in flight, so commit the cleanup decisions, rename the row skill, blah blah blah. Those are all things that we could say, yep.
09:19Opus 4.8, go ahead and do that. We have routing plus index truth, so just fixing some of the stuff. We have data catch up, so ingesting more Wiki from my q and a's and from my meetings and stuff like that.
09:31And then it also suggests durability. So having a weekly cron for Fireflies, which is, you know, my meeting transcripts, YouTube polls, archive sweeps, quarterly reruns. So So it's helping us constantly build this thing to make it better and better.
09:43So that was the actual audit deliverable. Let me real quick show you guys the actual skill. So in my dot cloud, in my skills, if I go to my OS audit and I pull up the skill at MD, I can do control shift v if you guys are in Visual Studio Code so we can just preview it a little bit better.
09:57We have the skill, the description, and the argument hint. And, basically, I'm not gonna read all this, but let's take a look at what I'm telling it. So OS audit, is your AIOS still true?
10:05This is your operating manual. Indexes and wikis are claims about what exists and what's current. The audit checks every claim against reality.
10:12Read only, never fix, or rename or delete. Just give a report of what needs to be changed, and this works on any Cloud Code project. So here is why I wanted to tell you guys about those different things.
10:23Here, it goes through the difference between poisoning, bloat, confusion, and clash, and then it also goes through expertise versus situational context. And then we get into the actual steps that this audit will be running. So step zero is prior report and recent evidence.
10:35So look for early reports inside of the audit folder. If one doesn't exist, then create that audit folder real quick. You can see that for a large project of a 100 plus folders to fan out one explore subagent per check below, giving each the check's instructions verbatim plus the project routes, and then you merge their reports.
10:51So utilizing some subagents here, the bigger that your project gets. So the very first check is routing integrity. Basically, does everything it points to exists?
11:00So it'll read through the Cloud. M d, the local, the agents. M d.
11:03It'll read through these different routing projects or sorry, routing files that are basically a table of contents because it's probably more than just one that you have in here, especially if you've added in some things like a Carpathi wiki or even multiple Carpathi wikis. It will also look in the reverse direction. It will see if there are things that are completely misrouted, which sometimes happens, and it will spot check everything.
11:22Then it will go to index truth. Do the indexes match the disk? Then it will go to freshness.
11:27Are all the data feeds current? So it will look and see if they are fresh, drifting, frozen, retired, or if they're on demand. And it will go through all of your different connections and understand when do you need to pull things in or, you know, do you only do that when kind of like a situational context presents itself?
11:44It will look at memory. It will go into looking for bloat, duplication, organization. So that's basically how this works.
11:49I'm not gonna read this entire thing out, but the skill is downloadable for free inside of my free school community. The link for that is down in the description. Okay.
11:56So now that we've seen a quick example and some background context on why I built it like this and why that's important, let's talk about these five hacks. So this first one is CloudMD as a router.
12:09So a lot of you guys are probably when you're building very specific projects, you're using CloudMD sort of like a system prompt. Sort of like, hey. Here's your background.
12:16Here's what you do, blah blah blah. And for the most part, that is correct. But the way that I have my Herc two iOS set up is I have it all under one massive folder.
12:28So if you see right here, let's just call this one my Herc two project. Now what I have inside of Herc two is a bunch of folders, literally just a bunch of folders. And then some of those folders have their own folders inside.
12:39And some of those folders have their own folders inside there. And that's how I have mine set up. And the reason I like to do it this way is because I like to just be able to push this main folder to GitHub and everything backs up.
12:49So for me, that was the easiest, and I like it that way. You don't have to do it, but that's honestly the way that I recommend. That also means I could c d into this directory, and this has its own Cloud dot m d, and so does this one, and so does this one.
13:00So that's where I can set more, like, specific project level sort of system prompts if I want. But this main CloudNMD that I use up here, I treat this almost purely as a router.
13:09So let me show you guys that. If I go into my HERC two and I go to the CloudNMD and I preview this here, what do we see is, basically, this is routing.
13:17So I start off saying, yeah. Hey. You're Nate Hurt's AI operating system.
13:20Your job is to help him spend less time in operations, blah blah blah. But here's where you actually go to find data. You know, here's where things live.
13:26If you need this, you go here. If you need this, you go here. If you need this, you go here.
13:31If you need this, you go here. All of this is basically just a routing table. And then I get into, like, the knowledge base.
13:36Here's the Wiki path. Here's the hot cache. Here's the index.
13:38Here's the GREP fallback. Right? Here's the memory system.
13:41Here's this. Here's this. Here's this.
13:42Here are the tools. Here's this and this. You know, there's obviously more than this, but here's where my API keys live.
13:47Here's where skills and agents live. Here's where decisions live. Here's templates.
13:51Here's references. Here's projects. Here's other worlds.
13:54And other worlds are basically massive other repos that all that all have their own individual GitHub repo because I like to keep everything synced. Right? And so as you can see, my CloudNMD is essentially just a master routing file, just a master table of context.
14:08And so I just wanted to show you guys, like, the way that I kind of have mine set up with my main folders. And once again, this isn't the most optimal way, and there's probably some ways that I could change this up, but this works for me right now.
14:21So in my HERC two, right, I've got obviously a Cloud that MD. I've got my dot Cloud with all of my pretty much global skills and global sub agents in my settings here that apply to basically this whole project unless, of course, I CD into one of these projects. But, you know, I've got my brainstorms folder, which is anytime I run a grill me session, it saves it here.
14:40I've got the Herc Brain, which is like my first overall wiki. I've got my other worlds, which are all of my other worlds. These are massive projects that I would consider standalone Claude code projects.
14:50So maybe if you have clients and stuff, this is where you can put that sort of stuff. And then I've got a folder called projects, is the largest one. Every single little time that I spin up a new chat and I start brainstorming about something or I wanna create some sort of deliverable, it will save it in here.
15:02So I'll show you guys that in a sec. And then I've also got at the root level something called brand assets where I have, like, brand guidelines, logos, pictures of me, things like that. So just to show you guys that in here.
15:12I'm on the desktop app this time. Obviously, you got, you know, different things. I've got dot agents.
15:16I've got dot Claude. I've got dot codecs. So keep in mind, all we're building here is a bunch of files and folders, which means you can plug in Hermes.
15:22You can plug in codecs. You can plug in anything else. So just keep that in mind.
15:25Right now, obviously, we're kinda focusing on Claude. But either way, here's the brand assets one. Right?
15:29I've got some fonts. I've got some pictures. I've got, you know, PNGs here, stuff like that.
15:34If I go to my brainstorms, this is where I have some of my brainstorming sessions from the grill me. I can see I've got some audits. I've got my decision log.
15:41I've got other things here. Here's my Hercbrain wiki. Here is my other worlds with all of these other projects.
15:46And pretty much all of these other projects have their own GitHub repos associated with them. And let's see. Here's my projects one.
15:53So this is where it starts to get big. Right? Like I said, almost everything that I do inside of this project goes in here.
15:59So that's why there's so many things in here. And what you'll notice guys is there's also one right here called YouTube videos, which if I open this up, this is also massive because every single one of these videos, obviously, I've got different things. So, you know, in here, I've got, like, transcripts.
16:13I've got, you know, scripts. I've got tests. I've got you know, sometimes I run a bunch of, like, research, and I run a bunch of, like, Python skills for this kind of stuff.
16:21So, anyways, the point being, there's just so many ways I can drill down in here. This is the way that I currently have mine set up.
16:28Obviously, I'm trying to optimize it every day and playing around with different things, but this is the way that I've got it set up with one big projects folder with a lot of my work in there, and then I save other things where it makes sense. But like I said, it doesn't really matter if you have it super flat, so you've got tons of products at your root, or it doesn't matter if you have one main one that you drill down everything into.
16:47All that matters is that you have the right routing rules in place so that you and your agents can find it. Now moving on to number two, have AI audit itself. So just what I did there, as you guys saw with this audit, just do this.
17:00I've obviously built a skill around this, but what I used to do is, like, at the end of every week when I made a bunch of changes or the end of every month, I'd be like, hey. Look through everything. Like, open up every file, look through my routing rules, and just make sure that everything's still accurate.
17:13Make sure that this all makes sense. And you know what? If you don't like how this is set up based on your best practices or based on the way that I talk to you every day, then suggest some changes.
17:21And let's just, you know, keep iterating and keep scaling this thing up because I know I add a lot of data to you every single week. And right now, you guys probably know that I have a few different Carpathi wikis. I've got my two main ones.
17:33Right? I've got my main one for my YouTube transcripts, and I've got my main one for my meeting transcripts. And I originally had all of that in one master wiki.
17:41And then it was actually Cloud Code that suggested, hey. You know, I've noticed that you're putting in these two specific types of, you know, meeting transcripts or specific types of data every single week, like, a specific cadence.
17:52You might as well just split those up so that it makes it easier for me to search through data so I'm getting less of that, you know, what do we call it? I'm getting less of that bloat. I'm getting less of that confusion, and I'm able to answer you not only quicker and more accurately, but also cheaper because I'm not spending so many tokens, so many of your tokens looking through everything.
18:09So, yes, I've made sort of a more formal audit process with that skill that you guys can check out, but you can also just have a conversation. Right? I think a really important mindset shift is to realize that everyone's building their own second brains.
18:21Everyone's doing a little bit differently, and there's not a right way. There's the only way that you're doing this wrong is if you're constantly getting wrong answers and you're not doing anything about it. But besides that, like, don't get so stressed out about, oh, you know, I saw Nate do it this way or I saw this other YouTuber do it this way.
18:36Just do it however it works for you. I think a really good test is, could you pull up your file explorer? You know, could you pull up your file explorer right here?
18:43Could you find like, think of something that you did and see if you could find it without searching, without asking Claude, see if you could find it. And if you can follow your different folders and you can find the path yourself, then it's probably set up pretty intuitively where an agent could also do so, especially if it had some routing rules in place.
18:59So I'm constantly making deliverables in here, and then I'm just finding them manually because I wanna make sure that this is still intuitive to me and that I actually understand my own second brain, my own AIOS. Okay. Moving on to number three, build automations to update data.
19:13If you guys remember here, right here, in the audit, it said durability, weekly cron for these different things. What you'll notice is as you start to use this thing more and more, you're probably going to be telling it frequently, hey.
19:24Can you go pull that data? Or, hey. Can you go do this?
19:27And you might realize that this isn't sort of like a just in time sort of situational context thing. Like, for my Wiki, every single Monday I have a q and a, I want that in there. Every single Tuesday when I meet with my leadership team, I want that in there.
19:39So why not just set up crons to automatically pull that stuff in there so that if I, for some reason, forget to after one of my meetings, I don't have to worry about it. And next time I go talk to my agent, it will already be there.
19:50And I know that sounds simple, but it's something that I've noticed a lot of people aren't doing. So definitely set up some sort of crons to pull in the data that you want to always be living inside of your local project. As you guys know, that's super simple.
20:02All you need is basically the API key, and you can say, hey. You know what, Claude? Go do this for me.
20:07Set up this cron. Set up this routine or help me push this, you know, script on the modal or whatever it is. You can set that up really easily with natural language.
20:15And number four is to segment knowledge. So as you guys know, I have two different Wikis. I've got my YouTube one.
20:21I've got my meeting transcript one, and I segmented them because they were both starting to grow, and I planned to keep growing those and growing those. So the more you can segment stuff out, the better.
20:31If you find yourself having little, like, nodes of knowledge that are gonna keep growing that are very distinct and different, then it's so much easier to split those up because now your agent knows, okay. If I have to find something related to, you know, Appify for some reason, let me check if Nate's ever made a video about that.
20:49And I know exactly to go to his YouTube video transcripts wiki rather than having to search through potentially five or 10 x more files in order to find what I'm looking for. So it's just about how can you narrow the actual context that your agent is going to be looking through. And a very common question I get around this whole segmenting knowledge idea is where do client projects go?
21:08And, like, should I put everything under one folder? And the way that I answer this question is basically that if I was working with a bunch of different clients, I would 100% keep information internally inside of my AIOS, and I would segment that by client. So I'd probably have one folder in here.
21:24I'd have one folder here called, like, you know, clients. And then I'd have client a, client b, client c. And then in there, I'd have different files that refer to what I'm working on with this client.
21:36But the caveat here is if I was working on something like a very specific deliverable for this client, I would probably build that out over here because this would maybe be like my client facing, you know, repo. And what I do is I would say, okay.
21:50In here, this is where it lives. Right?
21:53This lives in my it's typing in a different color. This lives in my c users, you know, Nate h, whatever, inside of a folder called clients a. But what I'd have in here would be the internal knowledge.
22:03Like, hey. You know, we started working with this client on June 22. He signed the contract on June 25.
22:08This was our project price. Here are the, you know, discovery calls. Here are the here's the scope of work.
22:13But here is where I'd have the external stuff, which would be, like, the actual deliverables that I'd be giving this client so that I could, you know, potentially have him be a collaborator on this repo, or I could, you know, push this to a certain environment. Whatever it is, I would keep it segmented a little bit away from my actual internal AIOS, but I would a 100% still be giving my AIOS context of this because it still needs to know about this engagement.
22:35It just maybe doesn't need to own all of the stuff inside of it when I'm pushing my whole repo, things like that. So it's very situational, and hopefully this is you know, hopefully, you understand, like, that there's no right or wrong way to do it. There's just wrong answers that you can get, and that's wrong.
22:50And then what I have here for number five is to backtrack. I've I've obviously stressed the importance of, like, when you realize something is wrong, when you realize it searched for five minutes for something it should have found instantly or it searched and said, hey. I don't have access to that, but you know it's in there, you have to correct it.
23:05Right? You say, hey. You know, you actually have access to that, so please make sure that never happens again.
23:10Make sure you're checking things first. But more importantly, have it backtrack. I found a lot of success in that when I say, you know what?
23:16You told me you didn't have access to this, but I know you do. So go look through what you did, where you searched, and help me figure out why you didn't find that data right away. And then once it does that, says, you know what?
23:28I made a mistake. I'm sorry about that. This is what I did.
23:31This is where I should've looked, and I would've gotten that much faster. Now that you've had it prove its own mistake and tell you what it did wrong and how to fix it, then just have it fix it. Have it update the routing.
23:40Have it maybe even move stuff or reorganize stuff based on what it found when it was backtracking. And that works a lot better than just saying, make no mistakes. Don't let that happen again.
23:49So, anyways, I know that second brains and AI operating systems, very, very hot topic right now and as they should be. And a big problem that we're looking at, we're talking to businesses about, that obviously I know every business is having is once people start building their own second brains and their own AI operating systems, how do you sync all the data together at the team level, at the department level?
24:09And that's something that I'm looking into very actively and playing around with. I don't have a great answer for you guys right now besides the fact that I don't think it's a tech problem. I think it's a people problem.
24:18I think that you could do this with Google Drive, with Notion, with GitHub. I think the problem is having people habit shift to syncing data, knowing what to pull in, making their agents read stuff, permissioning. I think it's a people and a habit problem that we need to figure out before we start pushing like, hey.
24:34This is the right way to do it. So, anyways, like I said, looking into it. I'll definitely be bringing more content on YouTube as we start to, you know, figure this out a little bit more, but that's something to be aware of.
24:43But I do think that the best thing you can be doing right now is mastering and understanding your own systems so that when you do start to bring this to a team level, it's a bit easier because you've already walked the walk. So that is gonna do it for today. I hope you guys enjoyed this one.
24:56And if you learned something new, please give it a like. Helped me out a ton. And as always, I appreciate you guys waiting to end the video, and I'll see you on the next one.
25:02Thanks, guys.
The Hook
The bait, then the rug-pull.
The most common question Nate Herk gets isn't about prompting — it's about how his AI operating system is organized. He opens by running a self-built "OS audit" skill live against his own project, then spends the rest of the video explaining the four ways that context quietly breaks (poisoning, bloat, confusion, clash) and the five habits that keep a growing second brain from confidently handing back the wrong answer.
Named ideas worth stealing.
01:44list
Four Context Failure Modes
- Poisoning
- Bloat
- Confusion
- Clash
The four ways an AI agent's loaded context produces a wrong answer: a false fact confidently repeated (poisoning), too much data burying what matters (bloat), a gap filled with a guess (confusion), and two contradictory sources with no way to pick (clash).
Steal fordiagnosing why any RAG/second-brain agent gives a wrong or stale answer
04:37concept
Expertise vs. Situational Context
Split context into always-loaded 'expertise' (identity, policy, goals — like a system prompt) and just-in-time 'situational' (pulled in only when needed, like a single support ticket).
Steal fordeciding what belongs permanently in CLAUDE.md/system prompt vs. what should be a live lookup
12:07list
5 Hacks to Level Up Your AI OS
- CLAUDE.md as a router
- Have AI audit itself
- Build automations to update data
- Segment knowledge
- Backtrack
The five maintenance habits the video builds toward, tracked on-screen as a running checklist.
Steal fora maintenance checklist for any growing AI second brain / operating system
How they asked for the click.
VERBAL ASK
07:00product
“go get your ideas roasted... my link gets you $1,000 in free credits to start”
Mid-roll sponsor read for HyperAgent, framed around a concrete use case he already built (a 5-persona review council) before pointing to a marketplace listing and a description link.
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.
