The argument in one line.
Building an AI operating system is not about choosing the right model - it starts with one default habit and compounds through four ordered layers that together turn a chat tool into a second brain that acts on your behalf.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code or a similar agentic IDE daily but still open separate AI tools for different tasks.
- You want a reusable knowledge and skills architecture that survives model switches and subscription changes.
- You manage a small team and are thinking about how to extend a personal AI OS to colleagues.
- You are already experimenting with custom slash commands and want a principled framework for organizing them.
SKIP IF…
- You want a step-by-step setup guide - the actual hands-on course lives in a free external community, not this video.
- You are not using an agentic harness like Claude Code, Codex, or equivalent - the architecture advice assumes one.
The full version, fast.
An AI operating system begins with one decision: defaulting to your agentic harness instead of scattered tabs. The four-Cs framework - Context (your CLAUDE.md as a routing tree), Connections (live vs. static data via APIs and CLIs), Capabilities (skills encoding repeated prompts), and Cadence (automated triggers) - describes the build order and explains why skipping layers creates brittle systems. The creator walks through his actual HERC 2 project structure live, including a permission-layer principle that no prompt is a security boundary: only scoped API keys are.
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 – 01:47
01 · Intro: Claude Fable and My Second Brain
Obsidian graph and Claude Fable benchmark table. Frames the video as a live system walkthrough, not a getting-started guide.
01:47 – 02:47
02 · The Mindset Shift
Close the tabs. Default to one harness. Before/After slide: scattered tools vs. single compounding context.
02:47 – 04:55
03 · The Four Cs Framework
Diagram: Context to Connections to Capabilities to Cadence. Second brain equals first two layers; OS equals last two.
04:55 – 09:51
04 · Context: Your Routing Tree
CLAUDE.md as a router pointing to rules, skills, wikis, other projects. Live VS Code walkthrough of HERC 2 folder structure and OtherWorlds pattern.
09:51 – 11:51
05 · Connections: Static vs Live Data
Tier-1 map of weekly apps by category (Revenue, Calendar, Comms, Tasks, Meetings, Knowledge). APIs and CLIs preferred over MCP servers for control and cost.
11:51 – 16:48
06 · What Fable Already Knows About Me
Live demo of /goal slash command generating a who-I-am video from transcripts. One-shot relationship-map app. Explains why static data goes stale and how live connections fix it.
16:48 – 22:53
07 · Capabilities: Skills and Workflows
Repetition is the signal for a skill. Assembly-line sessions: one clean job, then pass output. 20+ skills shown live. Update every skill after each use.
22:53 – 26:15
08 · Cadence: Automating While You Sleep
Three trigger types: manual, event, schedule. Autonomy is earned through battle-testing. Cautionary story: agent sent a discount code to 150-200K people unintentionally.
26:15 – 31:05
09 · Usage Tips
Thought partner with devil's advocate framing; grill-me skill for knowledge extraction; have agent verify its own work visually or with Playwright.
31:05 – 34:02
10 · Lightning Round Q&A
Cost, data privacy, no-code entry via GitHub starter repo, handling confident mistakes, live connections via APIs, team rollout.
34:02 – 34:20
11 · Final Thoughts
You are building your own personal OS, not a Claude Code OS. Folders, markdown, skills, routing logic - that is the IP that survives every model switch.
Lines worth screenshotting.
- An AI OS does not start with architecture - it starts with a default; closing tabs and opening your harness first is the entire first step.
- The second brain and the AI OS are two distinct layers; conflating them is why most setups feel like they never quite work.
- CLAUDE.md is a router, not a knowledge base - it holds pointers to where content lives, not the content itself.
- Static context and live connections require different handling; mixing them into one file creates staleness bugs you will spend weeks debugging.
- Any prompt you type three times is waiting to become a skill - repetition is the signal, not complexity.
- Skills are never finished - every use is data; updating the skill after each run is how the system compounds instead of plateauing.
- The assembly-line mental model applies to sessions: one clean job per session, then pass the output forward, rather than doing everything in a single context window.
- A prompt is never a permission layer - if the agent has a credential, assume it will use it; security is scoped API keys, not instructions.
- More autonomy in an automation means more monitoring, not less - the maintenance burden scales with the capability.
- The model is just the engine - folders, markdown files, skills, and routing logic survive every model switch; that is what you are actually building.
- Context rot is real: too much work in one session blurs outputs; phasing tasks with a clear between them keeps outputs sharp.
- Having the agent verify its own work is a concrete instruction, not a metaphor - give it a browser, have it click through, require it to show sources.
The model is the least interesting variable in your AI setup.
WHAT TO LEARN
Every capability gap you feel with AI tools is almost always an architecture gap, not a model gap - and fixing it follows a specific order.
02The Mindset Shift
- Default first: close every other AI tab and commit to one agentic harness before building anything else; context compounds only when it accumulates in one place.
03The Four Cs Framework
- The second brain and the AI OS are two distinct layers - Context and Connections build the brain, Capabilities and Cadence make it an OS; skipping layers produces systems that feel fragile.
04Context: Your Routing Tree
- Treat your system prompt as a router of pointers, not a knowledge dump - it should say where things live, not contain everything itself.
05Connections: Static vs Live Data
- Separate your static knowledge from your live connections - they go stale at different rates and need different refresh strategies.
07Capabilities: Skills and Workflows
- The signal for a skill is repetition: any prompt you type three times should become a reusable slash command, and you should update it after every run.
- Run one clean job per AI session, then pass the output to a fresh context for the next phase - mixing research, drafting, and publishing in one session degrades all three.
08Cadence: Automating While You Sleep
- Automate only what is battle-tested: the cost, risk, and maintenance burden of an automation all increase together, and set-it-and-forget-it is how agents send wrong emails to 200,000 people.
- A prompt is never a permission boundary - if your agent has a credential, architect around the assumption it will use it; scoped API keys are the real guardrail.
09Usage Tips
- Having the agent verify its own work is a concrete instruction you give inside the prompt - specify how (open a browser, click through, check accuracy) the same way you would with a new employee.
Terms worth knowing.
- AIOS (AI Operating System)
- A personal computing layer built on top of an agentic AI harness, combining a knowledge base, live data connections, reusable skill automations, and scheduled tasks - as distinct from a simple chat interface.
- Second Brain
- The knowledge layer of an AIOS - all the static and semi-static information the agent knows about your business, life, and context, stored as markdown files rather than a database.
- CLAUDE.md (as router)
- The primary instruction file in a Claude Code project, used here not as a knowledge store but as a routing tree of pointers that tells the agent where rules, skills, wikis, and other projects live.
- OtherWorlds folder
- A convention for nesting multiple distinct code projects inside one Claude Code workspace so a single session can reach across repos without switching projects.
- Context rot
- The degradation in output quality that occurs when a single AI session handles too many distinct tasks, causing earlier context to dilute or confuse later responses.
- Grill Me skill
- A custom slash command that instructs the agent to ask the user an extended series of questions to extract knowledge from the human into the second brain markdown files.
- Cadence
- The fourth layer of the four-Cs framework - automations that trigger on a manual call, an event, or a schedule, running without the user being present.
- Scoped API key
- An API credential with explicit read/write/delete permissions restricted to specific endpoints, used as an actual permission boundary rather than relying on prompt instructions to prevent unwanted agent actions.
Things they pointed at.
Lines you could clip.
02:41
“An OS doesn't start with architecture. It starts with a default.”
05:19
“I think of my CLAUDE.md file as my router. It's like, okay, this is where the files live.”
18:16
“Your skill will almost never be perfect on the first try. Every time you use your skill, that's data.”
24:33
“A prompt is never a permission layer. If it can, it will.”
29:31
“You're not building a Claude Code AIOS. You're building your own personal operating system.”
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.
00:00So this right here is basically my entire life and my entire business all in one second brain. And Cloud Fable is able to understand all of this better than I do. And not only that, but it's helping me automate a lot of this.
00:12So if you haven't heard, Claude Fable just dropped, and it is basically just Claude Mythos five, but there are more cyber guardrails baked in. And Claude Mythos is the model that Anthropic has been teasing for months now. It's the one that's so powerful that it's not generally available.
00:27It's only available to heavy cyber and infrastructure partners that are in project class wing. And what you'll notice here is there aren't a ton of, like, major major jumps, but a lot of the feedback that I've been seeing from Andre Carpathi, Boris Cherny, and also just the general community is saying that this model is definitely a major step.
00:45However, today is June 9, and this thing's only available from today through June 22 on your subscription, and then it's going to switch to usage credits. Now I'm sure that is going to come back to the subscription at some point, but you definitely wanna get in here and play around with this model for the next two weeks ish before it goes to usage credits.
01:02And this model is two times as expensive as Opus, so $10 for a million input tokens and $50 for a million output tokens. So it will eat your subscription faster than OPUS. But anyways, what I want you guys to understand in this video is that this is kind of my second brain, and then I've obviously got my operating system, which is called HERC two.
01:19So if you've been following my channel for a while now, you probably are aware that this exists. And this isn't something that I just built today because Cloud Fable dropped. This is something that I've been working on for months.
01:28But today, I have been playing around with Fable all day long to see how it's able to adjust into my AI operating system, and that's what I wanna talk about with you guys today. Basically, show you exactly how I've set everything up and exactly how I keep improving it over time and how I'm able to be so productive and efficient.
01:42So today, I'm gonna go over my CloudFable AI operating system and how this thing is my second brain. So the first thing that I wanna start off by talking about is a lot of this is a mindset shift.
01:52A lot of this is basically an adoption issue. It's a it's a habits shift. So, basically, before I started using this AI operating system, I was opening different AI tools, different subscriptions I paid for, and I was using like custom projects or custom GPTs, and I thought that that was being really productive, and it still was, but I was repeating myself, and it didn't feel like I had an actual kind of like a cofounder that I could talk to that always knew what was going on in my life and my business.
02:18And so the first step is basically close out of all those tabs and try to have the mindset shift, try to default to using Cloud Code or Codex or whatever you want your harness to be. Most of you guys clicked on this video, you probably want Cloud Code, but just default to using Cloud Code. I don't care if you use that in the desktop app or in Versus Code like I'm gonna be showing you in this video, just default to doing everything through Cloud Code, and you start to build up so much context and so much memory and preferences, and that's trust me, that's just the way to do it.
02:43So anyways, an OS doesn't start with architecture, it starts with a default. From there, we've kinda got two layers, and you probably hear people talk about my agentic OS or my second brain or my executive assistant, whatever they wanna call it. I think of it in two distinct ways.
02:57I think of the second brain as the first piece. Without a second brain, you can't have an AI operating system. Second brain is basically your knowledge.
03:03Does this thing know what's going on in your business, in your life, with your clients, you know, with your YouTube channel, whatever it is, does your second brain know that and can you ask it questions? Once you have the second brain, you're able to then build on top more of the OS infrastructure, the architecture of hey, now that we have all this knowledge, let's start to build some skills.
03:21Let's start to automate some things, and let's start to actually work out of here as my operating system. Rather than Apple or Windows or whatever you normally operate in, let's operate inside of this second brain instead.
03:32And so everyone has different definitions of second brain in AIOS, but that is the way that I think about it, and I think it's pretty simple. So from there, the framework that I like to use to think about building and maintaining my AI operating system is called the four c's, and we take it in order.
03:47You can see the first two are the second brain. The second two are the AI operating system. And we start with context, which is who you are, who your business is, stuff like that.
03:55And then we have connections. Can your second brain actually reach out for live data? Because sometimes you have data that's static.
04:02Right? Like your background, your meeting transcripts, your whatever, your 2025 progress and achievements.
04:10But then you have connections, is real data, like your ClickUp messages to the team or your emails or your, you know, your QuickBooks p and l. Any data that's less static and that's constantly changing, that's what I want to use as my connections.
04:24And then from there, we have capabilities, which is where we get into AIOS territory, building skills, building agents, building automations, building pipelines, and then turning those capabilities into cadence, meaning they can run on their own when you're sleeping, not just when you're sitting there, triggering them manually, and babysitting your AIOS.
04:39So this is the order. This is what I like to teach. And it let's say you already have your AIOS set up.
04:44This is still something really important to think about because maybe in the future, you need to help your team get upskilled with an AIOS, or you maybe actually wanna go sell AIOS setups to a business. This is the framework to teach. So starting with the context.
04:57This is your routing tree. You know, I'm gonna try to basically make this video by answering all the questions that I've been getting in my communities and my comments about how I use my AOS day to day. So I'm gonna show you guys.
05:07I'm gonna open it up, show you guys what's in there, but it's a routing tree. And so I think of my Cloud. Md file as my router.
05:13Meaning, yes, I have some stuff in there about like, hey. This is your goal. This is what I do.
05:18This is your, you know, your your processes. But more so, it's like, okay.
05:22This is where the files live. So I'm pointing my agent to all the rules, to the references, to the skills, to the other projects, to the wikis. And, basically, the pulse check here is, you know, when I open up my iOS, you guys will see I have so much so much context in here, so many files.
05:37And people ask me the question a lot about, at what point is it too much, and how do you know when you need to, like, split some stuff up? And so far, mine has not been too much. Because whenever I ask it to find things or I ask it to, like, pull in some data, it it finds them really quick.
05:51And because I'm sitting there watching it, if it's if it's searching for, like, five minutes for a file that I know where it is right away, then that's probably an issue. And I probably need to update the architecture and the file system. You know, we've heard context engineering, prompt engineering, harness engineering.
06:05I also think that architecture engineering is going to be a new kind of, like, art. And I don't think there's a right answer, but, basically, the pulse check for me is, is it intuitive to myself? Could I manually drill through my folders and files and find what I need?
06:17And can my agent do that as well? Okay. So here is my setup.
06:21Here are all of my files and folders on the left. And before I start drilling in, I just wanted to let you guys know this changes for me all the time. I'm always changing stuff around a little bit.
06:29I'm always adding new things, and I am always updating my Cloud. Md. But let me click in here real quick and just show you what's going on.
06:35So you are Nate Herc's executive assistant. I probably need to update that. It's more of my AOS.
06:39Your job is to help him spend less time on operations, people management, and admin so we can focus on learning AI tools and making YouTube videos. Here's some information about the knowledge base. I give it the exact Wiki path.
06:48I give it the hot cache. I give it the master index, and I give it basically the way to look through it. I then start to talk about tools.
06:54I then start to talk about API keys. I show where the skills live. I show how the skills should structure.
06:59I show what's active and when to use what. Honestly, I should probably clean this up a little bit, but that is kind of the way that I am thinking about showing where things are.
07:08And you'll also notice in here, I've got some other stuff. And in my dot cloud, the two most important things for me really in here are my my sub agents and then my skills. This is pretty much my number one favorite feature in Cloud Code, and this is how I'm so productive is with all of these custom skills that I'm building out.
07:21So that's where they live, of course. And then something that I recently have switched over to, which I've gotten a ton of questions about, is I'm using a a folder that I call other worlds. And what lives in my other worlds are basically other Cloud Code projects, like things that I was normally working in a completely different repo.
07:37You know, I would open them up from the file opener right here or the folder opener, but now I've just moved everything that I use frequently into my Herc two project. Now a couple reasons for that.
07:47I think the main one, honestly, I really started to think about it is from a syncing perspective. If I wanted to push my Herc two project to GitHub so that I could go on my laptop and pull it in, I now have all of these synced up as well. And I don't have to go push like six different, you know, projects to GitHub before I move over to my laptop.
08:04So that's one reason. Another good reason is because now my main operating system has context into like other things that I'm doing. So in here, I can have it go look at my book project.
08:13I can have it go change something on the website. I can have it, you know, edit the edit videos or do something in the token dashboard, whatever I wanna do. I can kind of do that all from here.
08:21And it can CD around, and it can find its way to what it needs. And I just kind of love having that feeling of all I have to do is open up part two, and I can get to what I need to do. And these are not small projects.
08:31Right? Like, if I open up, for example, the YouTube OS, we've got projects. We've got references.
08:36We've got transcripts. We've got a ton of drill downs in here. Like, these are pretty big projects, and I'm still able to keep everything finding, you know, where it needs.
08:45I'm still able to have my agent find things. And if I go in here in my main HER two directory and I do a slash context, I'm, you know, I'm still only at 40,000 tokens starting off.
08:54The majority of that is system tools. You know, we've got a little bit of memory files here, but this has not been an issue for me. And And I don't think that these markdown files will be an issue for a while.
09:03I also have a ton of working projects in here. So if I open up my projects folder, you can see that basically most of the stuff that I'm doing is typically creating a new folder inside of my projects. And these are not usually small folders either.
09:14Like, if I go to my YouTube videos project, you can see this is a pretty big folder with a bunch of YouTube videos that I've been working on, and those drill down even farther. So the point I'm trying to make here is you guys are asking a lot of questions, and I don't think there's a right or wrong answer.
09:28There's obviously a way to tell, okay. Something's not right here because I'm spending so much tokens or my agent's searching for twenty minutes. But for the most part, there's not a right or wrong answer.
09:36So you kind of get to be creative and orchestrate your own architecture here. As long as you can follow the paths and it's intuitive to you and as long as your agent isn't wasting your tokens, then I don't think there's a wrong answer.
09:47So anyways, that was just to hopefully give you guys a little sense of comfort. So that's kind of the idea of context, and the way that I like to think about what things to pull in and where to have my connections, which is the second c, is this is kind of like the first tier that I think about.
10:01I think about, okay. On a weekly basis, what apps might I open? What bookmarks do I have in Chrome?
10:07Where do I go to text people or talk to people, whether that's internal or external? And when you think about that tier one of like the main things you would reach for, those are probably great things to pull in first. You know, they're they're pretty high priority.
10:18And so if you're having trouble thinking through that, I would think about revenue, customers, your calendar, comms, tasks, you know, project management, meetings, knowledge.
10:27Where does that stuff live and where would you reach? And just start to write that down. So for me, just threw in some examples.
10:32Right? And you can see that some might live in multiple. So for me, school lives in revenue and customers.
10:38I also have customers on, you know, YouTube. I have revenue in Stripe and QuickBooks. My calendar's in Google Workspace.
10:43I come in Google Workspace, ClickUp, and Slack. So I just start to go through these things and and start to pull in context and start to make my connections. And my connections are typically through API keys and API endpoints, but I will get to that in the next section.
10:55But that is how it works, and I've got a full course, by the way. So this video is more of, like, my high level mindset, how mine is set up, how I think about using it, and how I get more out of it, and how I prompt it. But if you want a full step by step course, I've got one in my free school community.
11:09It's, like, a couple hours long. And like I said, it's completely free. So join my free school community, the link for that is down in the description, and you can do the build your own AIOS course.
11:17That's more of a like a step by step. I also in that course give you guys a free GitHub repo, which you can clone in, and then basically when you clone in that repo, it helps you walk you through all of that. The four c's, it gives you an audit, and it helps you build out the folder architecture from the beginning.
11:31So we go over the three m's and we go over the four c's. So all of that's available in my free school community. Anyways, let's keep on moving here.
11:37So the gut check is if you opened up your Claude code right now and you asked it about, you know, something about you, something about you and your business, would you get an answer that sounds like a stranger, or would you get an answer that sounds more like a teammate or a cofounder? That's a really good gut check. Just to show you guys real quick, two examples of something that Fable did for me and proving to you that it knows me so well.
11:57Look at this one. So this first one I said slash goal. Using hyperframes, I want you to look through this project.
12:02What do you know about me? What do you know about my YouTube channel? What do you know about our business?
12:05Create me a video that explains my journey, who I am, and what I'm up to. So let me pull up that video. Alright.
12:10So I'm gonna play this video on 1.5 speed. But who am I? Nate Herc, AI automation, Chicago founder, creator, dog dad.
12:17I teach everyday people to build with AI. 2024, hit record for the first time.
12:21Anyone can build AI agents. No code required. 41 videos in 2024.
12:25261 videos in 2025. The channel today.
12:30So here's one piece that I realized this is wrong. You know, right now I'm at, like, 800 k subscribers, but it got this wrong because I have a lot of static data inside of my AIOS.
12:42So last time it did a refresh, it found that I had 620,000 subscribers. Now if I right now prompted Cloud Code and said, many subscribers do I have And and pull that in live, it would pull that in live.
12:53It would update the static data, and we'd be fine. But anyways, just wanted to call out why that was outdated there. So anyways, those are channel stats.
12:59We move on to the free eight hour course, which got, you know, over 1,500,000 views. We've got the flywheel, YouTube, funnels down to AIS free, funnels down to AIS paid, which funnels into coaching high ticket, which is something that we don't have yet available.
13:13Our certification program we're working on, some of you guys might know, but more information to come about that soon. But, obviously, I've been talking to it about it, brainstorming, ideating, building automations. It knows what we're doing there.
13:24We go through the flywheel. Every video feeds the machine. It it talks about some revenue stuff, which is not completely accurate.
13:30There's a lot of different data sources that has to pull together. But either way, a 13 person team runs the machine, and that is Up dot ai. That's actually the core kind of, like, management company that owns everything that we're doing across our different ventures.
13:41So 13 person team is kind of the core team right now. My only job is to learn AI, make videos, delegate the rest. Obviously, it it has some bias because its job is to help me do that as much as possible, but we also recently made a big pivot this year from NNN to Cloud Code.
13:56Same mission, sharper tools, going AI native, agents in every quarter of the business. And then we also have AI's coaching, live events, and the book that I'm working on right here. So these are all things that aren't actually out yet, but you guys may know about all this.
14:09Obviously, it's helping me do that on the back end. So anyways, the point I was trying to make here is there's a video it did. In one prompt, I used the slash goal, and it created this video.
14:17Now one thing to throw out here is that, you know, Fable is not cheap, and that's one of the problems that people has been have been saying is that it is eating up their session limits. I when I first started playing around with this, I ate up my whole $200 a month max plan session. And, you know, the five hour session, I did that in it was, like, a little over an hour.
14:36Now, obviously, I was stress testing this thing, but it is noticeably eating it up more than Opus was, and that's obviously to be expected. So be careful with it. Now here's another one that I did, which is really cool.
14:45I did slash goal. I'm not gonna read this whole prompt. You guys can pause it if you want.
14:49But this was basically a one shot prompt. Right? This goal finished up in about twenty one minutes, and I basically asked it to put all of my transcripts and all of these, like, relationships into a super clean interface that anyone could go click through.
15:03Also, just realized that this window was a little bit too high, so sorry about that. Hopefully, that wasn't bothering you guys too much.
15:10But let's look at the output from this, which I think is amazing. So we have ideas. We have tools and harnesses, we have techniques.
15:17And when I click on them, they show the actual connection. So agentic workflows, that loops into context windows.
15:23And how does it? Cloud Code, Vercel, NNN, plan mode, value based pricing. And I can keep clicking around, and I can see these relationships.
15:30I can see the videos that those concepts are talked about in, and I think that this is just really cool. Cloud Code is kinda funny. It connects to literally everything, but we've got Codex here.
15:38We've got Open Claw, Hermes agent. We've got all these different things that we can click around. And basically, anything that I've talked about on YouTube, we now have this system where someone could come in here and and look at it.
15:47And this was once again a one shot prompt from Fable, and I also have up here. I'm new here.
15:53It can show you where to start. I build things. It shows you where to start, and I run a business.
15:57It can show you where to start. So this was something that I was pretty impressed by, honestly. If you think about what I asked Fable to do, this was pretty solid.
16:05Because keep in mind, like, what it had to go and do is it had to obviously build the front end and build these things, but it had to basically look through everything in here. It had to look through my actual YouTube transcripts and figure out the different relationships between things and how they actually come together in a practical way.
16:20And and it thought about the user journey there. So very, very impressive. So that was kind of starting to talk about connections.
16:27I kind of already hit on this. Right? Do they change constantly, or is it static?
16:31And how you think about bringing them in. And for most of the you know, for the most part, no database is gonna be needed when you start. Like you like you saw here, I'm basically just using markdown files.
16:40I'm using Carpathi's LLM Wiki with Obsidian, which I've got a video about, which I'll tag right up here. And that's kind of how I'm using all of my contacts and connections right now.
16:48Alright. So moving on to the third c, which is capabilities. This is basically now that you have, you know, contacts and connections, what can you actually do?
16:56What are the skills and what are the workflows and automations that you can build out of this thing? A big part of this once again is adoption. Now that you have everything in here, can you actually start using it instead?
17:06And not just using it for when you wanna brainstorm or when you have like maybe a script you wanna write, but using it to do things, using it to actually do your tasks. So the thing that I always try to challenge my students to do is rather than opening up that that Chrome tab when you were gonna send an email or when you were gonna go pull a report from, you know, some software, can you just default to opening up your Versus code, which is obviously where I like to use my cloud code.
17:29Right? Can you default to doing it inside of your AIOS rather than opening up that tab or opening up Perplexity or whatever you open up normally? Try to do it here.
17:39Try to figure out, okay. Does this tool have a p have an API endpoint? If yes, grab it.
17:43Does it have a CLI? Grab that. I tend to like to use CLIs and APIs more than MCP servers, but if you wanna use MCP servers, you can.
17:51The reason for that is I just feel like I have more control with CLIs and with APIs. They also seem to be cheaper, so I just like to use those instead.
17:59Anyways, skills don't have to be some crazy massive 10 step workflow. They can also just be a prompt. If you always are writing the same prompt or doing the same things on a Monday morning or a Friday evening, turn that into a skill.
18:11If you're brainstorming with clogged code about how to do something and then you actually end up doing it, then say, hey. What you just did was really good. I like this output.
18:19Turn that into a skill because next week when I have to do this again, I wanna do that again, and I want it to be faster. And the thing you have to think about is your skill will almost never be perfect on the first try. Every time you use your skill, that's data.
18:30It's data because you're able to say, here's what I liked what you just did. Here's what I didn't like. Update the skill.
18:35Let's run it again, and let's just make the skill better and better. And the thing is in my workflow or in my cloud code here, I've got a ton of skills. Right?
18:41This is maybe 20, maybe a little more than that. And then I've also got a bunch of global skills.
18:46Every single time I use a skill, I give it feedback and I say update the skill. So even though, you know, maybe this generate image skill I built four months ago, I'm still iterating upon it every single time I use it because my preferences change, the models change, maybe even the endpoints change.
19:02Who knows? Just have that mindset of like, there's no such thing as a finished product. And as you're building out more skills and as you're doing more work inside of Cloud Code, this is something that I've thought about even from the early n n n days, which is how can you have one AI doing one thing really, really well?
19:19Kind of thinking like an assembly line. You have one person that's making the wheels. One person is making the axles.
19:25One person's making the hood, and it finally all comes together at the end. So you're basically creating these little specialized AIs, and that's the way I like to think about my sessions. So if you did everything in one session, things might start to blur together.
19:37You might run into some context rot territory. But if you do, kind of like, hey. Spin up a Cloud Code agent and say, okay.
19:42Let's do some research on x, y, and z. And then you take that output, you do a slash clear, then and you bring that output into the next section, and now you you're you're working on the draft. And then you've taken that draft that is fully researched.
19:53You're doing you know, taking that output, putting into the next step, which is polishing. And that's obviously a very rough example, but I like to work in phases like that.
20:00I like to bring outputs and chain them together and have different specialized agents. Now skills obviously kind of help with that because skills lets you kind of bring in subject matter expertise into any session immediately, which is very nice. But for the most part, if you just have that mindset shift of like, my AI needs to do this one thing really, really well, and then I'm gonna bring that to a different AI.
20:20Now it's interesting because Claude Fable has been really impressive for me. It has been you know, I I try to tell you guys to not always feel like the benchmarks is the best way to judge if a model is is the right one or not.
20:34I think more importantly, it's the harness and the way it's the way that you use it. But I also think that there's a lot of feel. And I saw this text or I saw this tweet from, yeah, as if, like, Andre Carpathi is just texting me about how what he thinks about Claude Fable.
20:48I saw this tweet, and at this point, he said, you can give Fable a lot more ambitious tasks than what you're used to, and the model just gets it.
20:57And he put gets it in quotes. And it's interesting because a lot of times the model is all about feel, and I kind of have this feeling that when I'm talking to to Fable, it just gets things more often and just understands a little bit of what I'm talking about.
21:10And later in this video, I have some usage tips on how I feel like I'm able to get more on Fable, but one of those things is like giving it more context. And the more context you give it as far as like why you're doing something and what not to do, it just seems to get it with, you know, multi step reasoning and knowledge work.
21:27And this is Andre Kapathi. Yes. He did just join our anthropic, so maybe he has different motivations, but he's usually a pretty sound voice, and he said this is a super exciting release.
21:37He said, yes. The benchmarks are great and kind of state of the art, but this is a major version bump deserving step change forward. So far, I think it's a really impressive model.
21:46Some of the issues I've been seeing with it are on this the the pricing. Sometimes it's a little bit slow. Sometimes we run into those kind of guardrails because we know that it's being restricted, you know, cyber safeguards.
21:59And sometimes those are firing off frequently. Like, hey. I can't do that.
22:02Blah blah blah. Like it like it says here, the model still has quirks that people will run into, and the safeguards are configured to be a little bit too trigger happy for launch, which honestly makes perfect sense to me. But those will hopefully be tuned over time.
22:14So they will be monitoring this, and they will be tuning that stuff as more people use it and as the community gives more feedback on it.
22:22So that was context discipline. The other thing is about delegation. So not only do you kind of have your specialized, like, assembly line for chronological work, But for parallel work, you also wanna delegate, especially if you are using Fable, which is not cheap, then you can delegate to cheaper workers.
22:39Whether that be Sonnet or whether that be Haiku, you can delegate for parallel tasks and then get one clean summary back. And that's basically what the dynamic workflows do. I also just dropped a video on how to use Claude sub agents better.
22:51I'll tag that one right up here if you guys wanna check that out. And then moving on to number four, we have cadence. So this is basically the idea, like I said, that you can have things running while you sleep.
23:01And this is the last step, and you have to earn this because you have to prove that your skills are battle tested. You have to figure out the right use cases to actually automate. Because there's two things I wanna call out here.
23:09As you start to add more AI into things, the cost goes up and the risk goes up, and also the maintenance goes up. Because just because you put something into production and you automate something doesn't mean you can just forget about it forever. You still need visibility.
23:22You still need to check-in on it, and you still need to make sure that it's actually moving the needle. And more autonomy doesn't mean you forget about it. You still have to be the owner of that automation at the end of the day, or or someone has to have ownership over it.
23:32So from here, you're basically figuring out, I have all these skills. I have all this context. I have all these, you know, capabilities.
23:38What's the trigger? If it's manual and I just ask for it, then maybe it stays in your AIOS, and maybe you do If it's an event, meaning whenever you get a new email or whenever a new customer books a call or whatever it is, then you have a process happen. Or maybe it's a schedule.
23:52Maybe it's every Monday you do this, or maybe it's every Sunday night you do this. There's different ways that you can trigger automations. There's different ways to deploy them too.
24:00You can have them be Cloud Code routines. You can have them be loops. You can have them be deterministic scripts that you chuck on something like modal or TypeScript.
24:09There's a lot of different ways to have automations. You can even have your Cloud Code OS build you out an NNN automation, and then you can just push that into NNN. So there's so many different ways, nine ways to skin a cat, but figure out what works best for you and for your use case.
24:23I also have a video going over different ways to deploy your cloud automations, which I will tag right up here if you wanna check that one out. And the important part here about operating in your iOS and giving it so much reach and giving it so much context is that you have to earn the trust. And it's not just a feeling.
24:39It's not a vibe. It's a legit thing that you can actually feel safe about. So a permission layer.
24:45A prompt is never a permission layer. You basically have to have the assumption that if it can, it will. So if it could send an email, it might.
24:52If it could read that database, it will. I'm sure you guys have heard the story that I've told, which is we accidentally sent out an agent sent out an email to a 150,000, 200,000 people on our list with a discount code that was not supposed to go out.
25:06And what happened was the agent was proactively picking up tasks from a list, and it it interpreted the task wrong and thought that I need to write this discount code and send it out, and it did. And so that was obviously a big problem. We had to issue an apology.
25:18And so that was just a really great lesson for the entire team to realize that we have to understand exactly what our agents could possibly touch. What do they have the keys to? Because if they don't have the key to get into the room where they can send an email, then there's no way that they can actually go send that email, and that is what you need to have.
25:33You need to have keys, not prompts, and that's how you have an permission layer that you can actually trust. And from there, just think about every single time that you use your AIOS as, you know, more data.
25:45It compounds. If it slips up, you get data. It's not a failure.
25:49I wasn't mad at the person who built this agent where this happened. We treated this as an opportunity to write a case study and for the whole team to understand the risks of what we're doing and how we do things safely. But that was great data for us.
26:00Right? Like, in a way, we almost needed that experience to learn. So fix the instruction.
26:04It never happens again. You go in this loop. Like I said, I've been building my AOS for months now, and every day I make it better.
26:11And that is what's so cool about this is the data helps us make it better. Okay. So let's move into some usage tips here.
26:18So the first one is to treat this thing as a thought partner. This thing is really smart, and involving it in your planning and your ideation is a really good idea, but you wanna take that with a grain of salt.
26:29Don't just use it as a thought partner and say, okay. How would I do this? It gives you an answer and you say, okay, go.
26:35Let's do it. Use it as a thought partner, meaning brainstorm with it and make it play devil's advocate and you play devil's advocate. So take it with a grain of salt and maybe have it spin up different sub agents to give you different perspectives.
26:45I love doing that. I love having sub agents or maybe even agent teams debate with each other, give me different perspectives, and then I'm able to use all of that research and those different, you know, personas and perspectives to make a more informed decision based on my gut.
26:59But I love using my Claude as a thought partner. These models do have a tendency to be sycophants, so they will just basically please you and say yes. I do think that's getting better, but because that's kind of a known issue with all AI models in general, you should keep that in mind.
27:13Right? The next thing is have it interview you. So I've got a skill called grill me, which I will drop in my free school community along with everything else that I told you guys you can get in there.
27:22And the grill me skill was originally from Matt Pocock, so shout out to Matt. But I changed it up a little bit so that it has, like, brainstorm docs. So if I come in here and I go to my brainstorm folder, which is right here, which will automatically get created if you use my skill, I've got different sessions where it's grilled me relentlessly for I'm talking like 15 questions, 25 questions, 30 questions, and it's extracted so much more knowledge out of my head into my AIOS.
27:49So that's a really great way to start off too. You could literally set up your AIOS and say, use the grill me skill to figure out everything about my business, and it will just ask you questions until it has enough info. I actually did a grill me here, as you can see, literally today, to help figure out the way that I use my AIOS, the way that I've set it up, the way that I think about it, and this helps me plan out this entire video.
28:08So it's a very, very helpful skill. And then the final thing, which I think is the most important is to to verify its own work. You'll notice if I go into Claude and I go to this session where I prompted it to build this, you know, this relationship map thing, At the end of my prompt here, I said I gave it some context.
28:24I said this is a demonstration for YouTube, so don't feel the need to make this production ready, but it should be easy to understand so don't make it confusing. That's not even what I meant to show you guys. I meant to show you guys this.
28:33And then once you have built that, use a dynamic workflow to verify that everything is accurate and works as expected. It's really important that you are visually checking your work and testing that different personas would be able to click through this and understand it, meaning a beginner, a software engineer, a business owner, etcetera.
28:50So verifying its own work is super important, whether that is visually, whether that is by opening up a Playwright browser and clicking around. However you, as a human, would verify the work that, you know, maybe an intern gave you or an analyst would give you. However you would verify, just give Claude code the ability to do that instead.
29:08So that way when it verifies its work, instead of it giving you something on the first pass that was maybe like 70% of the way there and you have to iterate a little bit, maybe it can give you something that's 92% of the way there. And you still iterate a little bit, but it's giving you things that are better, and you're able to trust those outputs more.
29:22And one more thing, another big mindset shift is that the model, the harness, all of this is just basically the engine, and what you're building here is a system. You're building a second brain.
29:32You're building capabilities. You're building context and connections.
29:36You're building something that looks like this, and you're building something that looks like this. And at the end of the day, all this is is folders and files. So who cares if you switch over to codecs tomorrow?
29:44Who cares if you switch back to Sonnet 4.5? I don't care. It's folders and files, and every coding agent can use this stuff.
29:52That's why you see here, I've got my dot Claude, but I've also got my dot codex, and I've also got my dot agents, and I've got my Claude. M d, but I've also got my agents. M d.
30:01And I want this thing to be as tool agnostic as possible so that I can switch in different models. Like, I'm still using this thing on a daily basis with Codex because I'm trying it out with different stuff. I'm seeing what's better.
30:10And that is the mindset you should have with, oh, these new models come out. Right?
30:14Like, I don't have to do I have to rebuild this? Oh, I'm building this wrong. I'm I'm learning codecs.
30:18I'm not learning code. You're learning everything at once because you're building your own repeatable, essentially, IP of your business, your life, your capabilities.
30:28Folders and markdown files, skills and routing logic, logs and wikis. That's what you're truly building in your second brain and in your AIOS. You're not building a cloud code AIOS.
30:37You're building your own personal operating system. And it's really, really cool to think about that because it should hopefully remove some of that overwhelm or stress that you have about trying to stay up to date with the latest drops and stuff. You're building your own system, and whatever you plug in, whatever AI intelligence or harness you plug in, that's just what you're using right now.
30:54Because it'll probably get to the point one day where maybe you're using your own harness that you've built with an open source model, and you're not technical at all, but it just happened. Like, I I could see that definitely happening for a lot of you guys. And finally, I thought I'd end off with a little bit of a lightning round with other questions that I might feel like I I hear frequently.
31:11So what does this cost to run all day? Depends on the day. I mean, I'm generally just doing knowledge work in here.
31:17I'm not doing heavy, heavy coding, but I'm on the $200 a month plan. And I rarely hit my five hour limit every once in a while, and every once in while, I hit my session limit for the week.
31:29So that's what I'm on. Where does my data actually go?
31:32Well, if you're using Claude models, then that's going to Anthropic. It is a closed source model. So if you are dealing with tons of sensitive data, then you might not wanna use closed source models.
31:44Do I need to know how to code? No. Day one, empty folder.
31:47What's the first thing I type? I would use the AIS OS GitHub repo that I am dropping in my free school community along with the course. It tells you exactly how to do that.
31:55Follow those in there, and you will be able to get up and running in one day. What happens when it confidently gets something wrong? Well, as soon as it makes a mistake, you should be checking the work, obviously.
32:04But as soon as it makes a mistake, update it. Say, update your cloud dot m d so this never happens again. Update the skills so this never happens again.
32:11Always be verifying the exact source and show me the exact source where you're getting this data from. Saying things like that is how you make the system better and better over time because it will confidently tell you something, and it will be wrong. So it's on you to find that out and then to fix it.
32:23How do the live connections actually work? APIs or CLIs, and really all you have to do is search like, let's say, for example, you wanted to connect to Fireflies.
32:32You would search Fireflies API documentation, give that research to Cloud Code, or even just have Cloud Code do that research in the first place, and then say, here are the exact endpoints I need to connect to. What do you need?
32:42How does that work? And it will walk you through pretty much everything. If you really wanna get into it, you can start to get scoped API keys, which basically means, okay.
32:49This API key is for Cloud Code, and it can only read all my meeting transcripts, and it can't actually edit them or can't delete them or can't, you know, do anything with my team. All it can do is pull in transcripts, and that's how we talked about the permission layer. What is your cloud code able to actually physically do, not just prompts?
33:06What if I ignore it for a few weeks? Probably nothing too bad. You'll just have to, like, sync in your data or pull in some data.
33:11What about my team? Does everyone build their own? Ah, yes.
33:14That's a good one to end on. I say yes, but I think that you have to learn it first. You have to learn how it works first, and you have to be able to teach them at first because then you can help them set it up.
33:23You can help them maybe you've built a few skills that are team wide that you can give to your team. And then when it comes to shared knowledge, you should be thinking about where is our team's knowledge?
33:32Does that live in ClickUp or Slack or Notion or maybe both or Google Drive? Where is the team knowledge that everyone should have read only access to? Because the worst case scenario really is that everyone's duplicating knowledge, duplicating skills, duplicating work.
33:46And the biggest issue there is adoption. Actually getting people to use the shared knowledge and actually getting certain stakeholders and process owners to update that shared knowledge. Adoption is the biggest issue, which is why you have to learn the tech first and be able to explain it and communicate it and make sure that your team adopts it.
34:02So that is going to do it for this one. I tried to put a ton of value in here, and I tried to make this one practical for you guys. So I hope that it was insightful, and I hope that it helps you guys out.
34:13If it did, please give it a like. It helps me out a ton. And as always, I really appreciate you guys making it to the end of the video, and I will see you all in the next one.
34:19Thanks, guys.
The Hook
The bait, then the rug-pull.
An Obsidian knowledge graph fills the screen - hundreds of nodes, each one a piece of a creator's business. The claim is audacious: a single AI agent understands this better than its owner does. What follows is not a product demo but an architecture tour, and the lesson is that the model is the least interesting variable.
Named ideas worth stealing.
02:47model
The Four Cs
- Context
- Connections
- Capabilities
- Cadence
A build-order framework for an AI operating system. Context = who you are and your routing tree. Connections = live data access. Capabilities = skills and automations. Cadence = scheduled/event-triggered runs. The order is not optional - each layer depends on the previous.
Steal forExplaining AIOS architecture to a team or client before building
04:55concept
The Routing Tree
CLAUDE.md acts as a router, not a knowledge store. It holds pointers to where content lives. Pulse check: if you cannot manually drill to any file, the agent cannot either.
Steal forAuditing an existing CLAUDE.md that has become bloated with inline content instead of references
19:16concept
Assembly-Line Sessions
Each Claude Code session should do one thing well and hand a clean output to the next session. Parallel tasks delegate to cheaper sub-agents with one summary returned to the orchestrator. Prevents context rot.
Steal forMulti-step research-to-draft-to-polish workflows
24:33concept
Keys Not Prompts
A prompt is never a permission layer. If an agent has a credential, assume it will use it. Real permission layers are scoped API keys that limit physical capability: archive instead of delete, draft instead of send, read instead of edit.
Steal forAny production automation where the agent has write access to external systems
How they asked for the click.
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.
