Build Your CUSTOM Claude Code Agentic OS (3 Steps)

CChase AI
컴퓨터/소프트웨어창업/스타트업경영/리더십AI/미래기술

Transcript

00:00:00Most people use cloud code like a slot machine.
00:00:02They're just using random prompts on random tasks
00:00:05and ultimately getting random results.
00:00:07But if we instead use an agentic OS,
00:00:10we can create a system that we can optimize, track,
00:00:13and ultimately hand off to members of our team or clients.
00:00:16With a cloud code agentic OS,
00:00:19we turn your daily workflows into skills,
00:00:21skills into automations, and automations into architecture
00:00:25before we wrap the entire thing
00:00:27in a memory and observability layer.
00:00:29And today I'm gonna show you how to do all of this
00:00:32in three steps, and in the process,
00:00:34put you ahead of 99% of cloud code users.
00:00:37When we talk about an agentic OS,
00:00:38there's three things we're gonna cover
00:00:40and relates to the three steps for building it.
00:00:42First is the observability layer,
00:00:44and that's what you see here.
00:00:45This is sort of a dashboard visual setup
00:00:48that lets us interact with our OS from outside the terminal.
00:00:52Number two is the memory layer,
00:00:53and this is where the Carpathi Obsidian rag setup
00:00:56comes into play.
00:00:57And lastly, and I would say most importantly,
00:01:00is the architecture, the skill set up.
00:01:02When I talked about in the intro
00:01:04of taking your daily workflows,
00:01:05turning them into skills, skills into automations,
00:01:07and automations into architecture, I'm talking about this.
00:01:10This is the backbone, and this is what actually
00:01:14is the value add of an agentic OS.
00:01:17It's not the fancy dashboard, as cool as this thing is.
00:01:20This isn't the true value.
00:01:21This is the true value.
00:01:23So what is this architecture?
00:01:25Why should you care, and how do we set it up?
00:01:27Well, this right here is essentially a visual
00:01:29of everything Claude code should be doing for you,
00:01:32codifying it, and ultimately automating it.
00:01:35The idea is simple.
00:01:36We have you in Claude code, and if you're like most people,
00:01:39this is kind of where it ends.
00:01:40You have the terminal open.
00:01:42You ask it to do a random task.
00:01:43There is no system.
00:01:44There are no steps.
00:01:45Nothing is tracked.
00:01:46Nothing is optimized.
00:01:47In our agentic OS system, what we have done instead
00:01:50is we have broken down everything you do on a personal level,
00:01:54and more importantly, at a business level,
00:01:57and we've broken them up into domains.
00:01:59So for me, I have memory, productivity, research,
00:02:01content, community, on and on and on.
00:02:03What you would have specifically would be different
00:02:04depending on what you do, but the idea is
00:02:07we do a lot of different things,
00:02:08and they have specific domains.
00:02:10For example, let's take a look at the research domain.
00:02:12I do a lot of research, and under each domain
00:02:15are a number of discrete or individual tasks.
00:02:17For me, I'm often looking stuff up on YouTube.
00:02:20I need to do deep research into things.
00:02:22I need to deal with light rag.
00:02:23I want to have a morning report.
00:02:24I want to watch my competitors, on and on and on and on.
00:02:27So every domain's gonna have individual tasks.
00:02:29Each of these tasks, if this is something we do regularly,
00:02:32A, it should be in here, but B, we can turn it into a skill,
00:02:36and this can be things that are relatively simple,
00:02:38like this YouTube search.
00:02:39Hey, instead of me just going into YouTube
00:02:42and typing something in there,
00:02:43why don't I turn that into a skill
00:02:45and get a complete report every time?
00:02:47This can also get more complicated.
00:02:48Something like deep research isn't just me
00:02:51telling Claude, "Go do deep research."
00:02:53This is me looking at Twitter, GitHub, the web,
00:02:56YouTube still, but even more so,
00:02:58taking a look at previous entries in Obsidian
00:03:01to see what I've talked about in the past
00:03:02and consolidating all of it.
00:03:04The point is you do a lot of different things
00:03:06across a lot of different domains in your day-to-day life
00:03:09and in your business life.
00:03:10Have you codified it in this manner?
00:03:12Have you turned every task into a skill?
00:03:14Do you have a way of tracking all this and optimizing it?
00:03:17Chances are no, and if that's the case,
00:03:19even if you do nothing else with this whole Claude OS system,
00:03:22doing the memory, doing the dashboard, all that cool stuff,
00:03:25if you just stopped here, you would get a ton of value.
00:03:27So we've broken up our life and our business
00:03:29into different domains.
00:03:30We've broken up the domains into tasks.
00:03:32Those tasks become skills.
00:03:34Now, the next step is to take those skills
00:03:36and turn them into automations.
00:03:38Now, not everything needs to be automated, but some things do.
00:03:41Take a look at the morning trend scanned.
00:03:43This is obviously something I would want every single morning
00:03:46that populates inside my Obsidian database saying,
00:03:50here's the scan of what's going on in AI
00:03:52and your competitors on YouTube, on GitHub, et cetera, et cetera.
00:03:55That's an easy win in terms of an automation.
00:03:57Now, automations can really come in two flavors
00:04:00when we talk about automations.
00:04:01They can either be local automations,
00:04:03or they can be remote automations.
00:04:05Luckily for us, we don't even need
00:04:06to know which ones they should be,
00:04:08because you know who's good at figuring it out?
00:04:09Claude code.
00:04:10And if I tell Claude code I want to create a local automation
00:04:12or a remote automation, it will be able to figure it out.
00:04:14But what you care about isn't me going
00:04:16into the minutia of local versus remote automations.
00:04:19What you care about is, Chase, how can I create this?
00:04:22How can I figure this out?
00:04:24Well, luckily for you, it's not too difficult.
00:04:27All you need is this prompt.
00:04:29But before we go into that, a quick word
00:04:30from today's sponsor, me.
00:04:33So I just released my Claude code masterclass,
00:04:35and it is the number one way to go from zero to AI dev,
00:04:38especially if you don't come from a technical background.
00:04:41Everything you see in today's video from the prompts
00:04:44to my actual agentic OS system that I use myself
00:04:48can be found inside of here.
00:04:50So if you want to get your hands on that,
00:04:52there's a link to it in the pinned comment.
00:04:54Hope to see you there.
00:04:55So like I just said, this exact prompt that we're going to use
00:04:57can be found inside of my community.
00:04:59And the idea is this is going to kick off
00:05:02a conversation between you and Claude code.
00:05:04So you can build something like this.
00:05:07In general, the way it's going to start
00:05:08is you and Claude code are just going
00:05:10to have an open conversation.
00:05:11I suggest opening up the terminal,
00:05:13turning on your microphone, and just doing a stream
00:05:15of consciousness, sort of explaining what you do day to day
00:05:19and what your discrete, your specific tasks are.
00:05:22From there, it's going to continue
00:05:24to have a back and forth with you.
00:05:26And then it's going to be like, OK, you're doing X, Y, and Z.
00:05:29Can we turn X, Y, and Z into a skill?
00:05:32If we can turn it into a skill, does it make sense to then turn
00:05:35it into an automation?
00:05:37Like I said, not everything needs to be an automation.
00:05:39Something like the morning trend scan makes total sense.
00:05:42Deep research, not so much.
00:05:44But it's going to go through each and every task,
00:05:46create a skill for you.
00:05:49So you can execute that task the same way every single time.
00:05:51And it's going to use the skill creator skill.
00:05:53And then it's going to figure out,
00:05:55does it need to be an automation?
00:05:56And if it's an automation, does it need to be local?
00:05:58Or does it need to be remote?
00:06:00It's going to continue that process for each and every
00:06:02domain you spell out.
00:06:04So it's not going to be just what you see here.
00:06:06If you don't do research, you don't do content,
00:06:08you don't run an AI agency, that's fine.
00:06:10But whatever it is you do, we're going
00:06:13to create a domain for it, create skills,
00:06:16create automations.
00:06:17And in the process, you create this backbone
00:06:21of a Claude code powered agentic OS.
00:06:23You are codifying behaviors in a way that we can track
00:06:26and we can optimize.
00:06:28That way, when you show up to Claude code
00:06:29and you use the system, you're not just
00:06:31guessing every single time and hoping
00:06:33that Claude code does the same thing it did yesterday.
00:06:35And the power of that goes beyond just you
00:06:38as the individual using this Claude code system.
00:06:41If you're someone who works on a team
00:06:43or you're someone who works with clients, this is massive.
00:06:46Because what does this mean?
00:06:48If I have pretty much codified everything I do into a skill,
00:06:51well, then I can give this system to someone else in my team
00:06:55who should be using Claude code but never will,
00:06:57and now they can use it.
00:06:59Same thing with clients.
00:07:00You can set the same exact system up for other people,
00:07:03package it, sell it, force them to use it
00:07:06if you're on your team.
00:07:07But they don't even have to use the terminal.
00:07:09Because when we eventually go into the dashboard section
00:07:12and we look at something like this,
00:07:13what we're eventually going to do
00:07:14is we're going to turn all these skills and automations
00:07:16literally into a button you have to click,
00:07:18and anybody can do that.
00:07:20So that's step one of three of creating the agentic OS.
00:07:23It is the architecture, and it's the most important.
00:07:27And if you do nothing else,
00:07:28you'll get a ton of value out of this.
00:07:31Now, step two is the memory layer,
00:07:33and we are going to be using Obsidian for this
00:07:34because it doesn't do us any good
00:07:36to have all this stuff running in an operating system,
00:07:40yet we can't go ahead and look at what we've done
00:07:42in the past or store information.
00:07:44And Obsidian gives us a very simple way to do that.
00:07:47Now, the great thing about Obsidian is it's free.
00:07:49I put out a bunch of content on Obsidian and how to set up,
00:07:52so definitely check that out
00:07:53if you want to do a deep dive for it.
00:07:54But the thing with Obsidian is if we get very reductive,
00:07:58all Obsidian really is is it's a nice layer,
00:08:00a nice interface for us to be able to interact
00:08:03with Markdown files.
00:08:04If you just download Obsidian
00:08:05and you run Cloud Code inside Obsidian,
00:08:07it's not going to do much for you.
00:08:08It's how we set up the sort of file structure.
00:08:11Within Obsidian itself, that's important.
00:08:14That's how we actually derive value from this piece.
00:08:16By now you've probably heard of the Carpathi Obsidian
00:08:20quote unquote air quotes here, rag system.
00:08:23And that's kind of the structure we're looking at here.
00:08:25And again, I've done content on this as well.
00:08:26And this is a great place to start
00:08:28when we talk about a memory layer for our system.
00:08:32So the way Obsidian works is when you download Obsidian,
00:08:35you designate a single folder as the vault.
00:08:38It doesn't have to be called the vault,
00:08:40but in this case it is.
00:08:41It's literally called the vault.
00:08:42The vault is where your Cloud Code agentic OS
00:08:45is going to live.
00:08:46So if you want to use the OS
00:08:48when you start up your terminal,
00:08:49you're going to need to be in the vault.
00:08:51Now, how you set this up is ultimately up to you.
00:08:54The great thing about everything we covered today
00:08:56is it's customizable.
00:08:58You don't have to do it exactly like this,
00:08:59but it's a great template to start with
00:09:01and you can tweak it as you see fit.
00:09:03But the way Andre Carpathi lays it out
00:09:05is we should really have three sub folders
00:09:07in the vault system.
00:09:08We have the raw, we have the wiki, and we have the output.
00:09:12Big picture, why does this work?
00:09:15Well, we sort of have like one folder
00:09:17is sort of the dumping ground.
00:09:18Whether it's us just talking to Cloud Code
00:09:21or researching random stuff, this is like the staging area.
00:09:24We then have sub folder number two,
00:09:26which is the wiki section.
00:09:28And the wiki section is sort of this intermerry,
00:09:31intermediary piece where we take stuff from the raw
00:09:35and we then codify it into wiki type articles.
00:09:38So we don't just have a bunch of random information
00:09:41sitting inside of our agentic OS.
00:09:43Well, now we have a series of like wiki articles.
00:09:46So let's say I did a bunch of research about rag systems.
00:09:50Well, all that research would go into the raw.
00:09:53And then Cloud Code would create articles
00:09:55that are actually detailed reports
00:09:57about everything it researched.
00:09:59That would go into the rag system wiki.
00:10:03Then let's say we wanted to take those reports
00:10:05and turn it into a slide deck.
00:10:06Well, that goes into section number three,
00:10:08which is the outputs, right?
00:10:09So maybe we have a slide decks sub folder,
00:10:13which now has information about rag systems.
00:10:15You kind of get what's going on here.
00:10:18The thing with memory is A, we use Obsidian
00:10:20to kind of control it all.
00:10:21But B, the real value is how are you gonna set it up, right?
00:10:24What makes sense for you?
00:10:25This is just one way to do it.
00:10:27All you need is you need something that makes sense.
00:10:30So if we go back to sort of the agentic architecture here,
00:10:34you could do something where every single domain
00:10:37is a sub folder.
00:10:39Everything about research goes into research.
00:10:41Everything to do with my AI agency goes into my AI agency.
00:10:44Everything to do with sales goes into sales sub folder.
00:10:47It really doesn't matter, right?
00:10:49It really doesn't matter.
00:10:50There's no right or wrong answer here except to say,
00:10:54you just need something that makes sense
00:10:55and you wanna use Obsidian
00:10:57because it's a great middle ground
00:10:59between a full fledged rag system.
00:11:02For 99.9% of people, you don't need
00:11:06even something as lightweight as light rag.
00:11:08You don't need a vector database.
00:11:09It's too much.
00:11:11And if you're just using Markdown files,
00:11:13Claude code can handle something like this just fine
00:11:15within the confines of Obsidian.
00:11:17Now, the one thing you do need to do
00:11:20when it comes to Obsidian and Claude code in this OS
00:11:22is create a proper Claude.md file.
00:11:25Right here, I have a template you can use.
00:11:29And what this is going to do is it's A,
00:11:32it's gonna tell Claude code
00:11:33what the heck is going on here, right?
00:11:35What is my purpose?
00:11:36How should I be functioning?
00:11:38What do you do?
00:11:39What should Claude code care about
00:11:41when we give it any prompt at all?
00:11:42Because the Claude.md file for all intents and purposes
00:11:45is pretty much appended to every single prompt you give it.
00:11:48Secondly, what the Claude.md file is going to do
00:11:50is it is going to spell out for our agentic OS system
00:11:55how its memory is actually structured.
00:11:58And if we tell it how the memory is structured,
00:12:00well then it's actually going to adhere to it.
00:12:02And it's gonna be able to find what it needs to find
00:12:04with less tokens and ultimately give you a more efficient,
00:12:07less costly system that not only can Claude code
00:12:11actually navigate its way through,
00:12:12but you can navigate its way through.
00:12:14Right here, you can see my structures.
00:12:16It's not too complicated.
00:12:17I have an archive, content, ops,
00:12:19personal projects, raw and wiki.
00:12:22So it's somewhat of a spinoff of the Carpathi rag structure.
00:12:26The point is it makes sense to me
00:12:28and it's clear enough to Claude code
00:12:31in terms of how I want it to be structured
00:12:33and where I want things to go
00:12:34that it makes sense and that it works.
00:12:36That's all you really need, but you do need it.
00:12:38You can't skip the section.
00:12:39The sort of memory system also gives us the ability
00:12:42to track things and therefore optimize them.
00:12:45Because if everything's done in a vacuum,
00:12:47we never know what really is working.
00:12:49So again, everything is tied together.
00:12:52We need to nail the memory piece.
00:12:53Now it's time for the sexy part,
00:12:55which is the agentic OS dashboard
00:12:57and the whole observability system.
00:12:59What we are really doing here
00:13:01is we are simply taking all of this,
00:13:04taking the skills, taking the automations,
00:13:06and we're gonna take each one that we care about and we use,
00:13:10and we're gonna put it here inside the OS.
00:13:13That's kind of what I've done here.
00:13:15Each of these buttons is either an automation
00:13:17I can trigger with a click
00:13:18or a skill I can trigger with a single click.
00:13:21So if I hit something like deep research,
00:13:23you see it populates the prompt right here.
00:13:26And I just need to put inside an input.
00:13:28And it's the same thing as if I took this exact prompt
00:13:32and put it into Claude code.
00:13:33So if I put here, so if I write in here,
00:13:37Claude code skills and hit run,
00:13:40what's happening is it's now starting another instance
00:13:44of Claude code, but it's headless.
00:13:46It's like an invisible version of Claude code.
00:13:49So it's using, it's just using the dash P flag to do that.
00:13:53And then here I'll get a hole right up
00:13:55and just as if I wouldn't set the terminal.
00:13:57That whole system of turning skills into buttons,
00:14:00the real value play for there is if you're doing this,
00:14:02again, with team members or clients,
00:14:07because the truth is if you are someone who is adept
00:14:10at Claude code and using the terminal,
00:14:11or maybe even just using inside of something like VS code
00:14:14or the desktop app, hey, the idea of taking these automations
00:14:17and taking these skills and turning them in the buttons
00:14:19sounds great, doesn't really do anything for you, right?
00:14:21Because you're good enough at this point.
00:14:22Like I can just get those going.
00:14:24I don't need that system.
00:14:25But if you do any AI agency work, huge.
00:14:29If you work with a team and they're just like,
00:14:31they're not gonna do the terminal,
00:14:33but you're trying to give them the power of Claude code.
00:14:35'Cause think about that.
00:14:36You are giving them the power of Claude code.
00:14:37I could take anybody, put them in this chair right now,
00:14:40put them in front of the agentic OS and say, do X, Y, and Z.
00:14:43Here's the skills.
00:14:44They can do it.
00:14:45Like that's, there's real value there.
00:14:46But the second piece of this whole
00:14:48like non-terminal visual thing is the observability side.
00:14:51And again, this is extremely customizable.
00:14:53So right here, I have stuff related to usage,
00:14:55like my five hour window, my weekly window,
00:14:57the amount of routines I've used for the day.
00:15:00I also have stuff over here on the right-hand side
00:15:02related to recent changes to my vaults and forecasts
00:15:06and things of that nature.
00:15:07But this can be whatever you want.
00:15:08In an ideal world, it's tied to, you know,
00:15:11your sort of skills and things like that.
00:15:13Like what do you wish you could actually see
00:15:16inside the terminal?
00:15:16As good as the terminal is,
00:15:18the terminal does have some limitations.
00:15:20If we move to a system like this,
00:15:22we can get around those limitations.
00:15:24Because if there are things we want to track,
00:15:26we can put that here.
00:15:27And it is literally one prompt inside of Claude code
00:15:30to execute that.
00:15:31So here's a look at that deep research output.
00:15:33Gives me sort of the overview,
00:15:35gives me a link to all of its sources,
00:15:37and also includes a link to where it exists inside Obsidian.
00:15:40Now, in terms of creating this dashboard,
00:15:41I have a whole prompt for that as well.
00:15:44It's going to look slightly different than mine
00:15:45when you run it.
00:15:46And that's because at first,
00:15:47it's just going to be a lot of placeholders
00:15:48because it's going to start a conversation
00:15:50between you and Claude code where you figure out,
00:15:52okay, which skills do you actually want tied
00:15:55to this dashboard?
00:15:56Furthermore, what do you want in terms of observability?
00:15:59Do you want the usage limits?
00:16:00Do you want routines?
00:16:01Do you want sort of forecasts and update
00:16:02of what's going on in their vault like mine?
00:16:04Maybe not.
00:16:06Doesn't really matter what you want
00:16:07because you can customize it to be anything.
00:16:09And I think that customization piece is so big,
00:16:11especially if you do any level of client work.
00:16:14But that's really it in terms of the three steps.
00:16:17Step number one is architecture.
00:16:20What do you do? Can we break it into domains?
00:16:23Can we take the domains and the tasks,
00:16:25tasks and the skills, skills and automations.
00:16:28Step number two is the memory piece.
00:16:31How are we going to set up our obsidian vault
00:16:33so that not only does Claude code have a clear path
00:16:35for where data needs to flow,
00:16:38but so you have a clear idea of where data needs to flow
00:16:41and where it is,
00:16:42because it's not enough for just Claude
00:16:43to figure out everything you need to be able
00:16:45to actually see what's going on.
00:16:47And speaking of being able to see what's going on,
00:16:49that is step number three,
00:16:50which is observability,
00:16:51which is one piece, you know,
00:16:53being able to do things we can't do inside the terminal,
00:16:56but a second piece is empowering members of your team
00:16:58or clients by giving them the ability
00:17:00to execute these skills and automations
00:17:02with literally the press of a button
00:17:04and they never have to touch a terminal.
00:17:06You put all that together
00:17:07and you get a Claude code powered agentic OS
00:17:10that you can customize to your heart's desire.
00:17:14So that's where I'm going to leave you today.
00:17:16As always, let me know what you thought in the comments.
00:17:18Make sure to check out Chase AI+
00:17:20if you want to get your hands on the Claude code masterclass,
00:17:22as well as my exact Claude code agentic OS.
00:17:27And besides that, I'll see you around.

Key Takeaway

Building a Claude Code Agentic OS replaces unpredictable prompting with a three-step framework of codified skills, an Obsidian-based memory vault, and a visual dashboard to enable scalable team collaboration.

Highlights

  • A Claude Code Agentic OS transforms random prompts into a codified system of skills, automations, and architecture.

  • The system operates through three distinct layers: architecture for workflows, Obsidian for memory, and a dashboard for observability.

  • Individual tasks within domains like research or content are converted into reusable skills using the skill creator skill.

  • The memory layer utilizes a three-folder structure consisting of raw staging, wiki articles, and final outputs.

  • A custom Claude.md file serves as the system instruction that is appended to every prompt to ensure structural adherence and token efficiency.

  • The observability dashboard allows team members or clients to trigger complex terminal commands via simple button clicks.

Timeline

The transition from random prompting to agentic systems

  • Unstructured use of Claude Code leads to inconsistent results similar to a slot machine.
  • An agentic OS converts daily workflows into a sequence of skills, automations, and architecture.
  • A memory and observability layer surrounds the core architecture to track and optimize performance.

Standard terminal usage often lacks a repeatable system, resulting in untracked and unoptimized tasks. By building an OS, users can prepare their workflows for handoff to team members or clients. This process begins by viewing every action as part of a larger structural framework rather than a one-off command.

Building the core architecture and skill domains

  • Architecture is the most important component of the OS, functioning as its backbone.
  • Workflows are categorized into specific domains such as research, content, community, and productivity.
  • Tasks within these domains are codified into skills that execute the same way every time.
  • Automations within the OS function in two types: local and remote.

The value of an agentic OS lies in codifying behavior so it can be tracked and improved. For example, a research domain includes discrete tasks like YouTube searches or deep web research which are then turned into executable skills. This allows someone with zero technical background to run complex operations by simply calling a pre-defined skill or automation.

Implementing the memory layer with Obsidian

  • Obsidian provides a free interface to manage the Markdown files that store the system's memory.
  • The folder structure uses three sub-folders: raw for data dumping, wiki for codified reports, and output for final assets.
  • A dedicated Claude.md file defines the system's purpose and memory structure for the AI.
  • Markdown-based memory is more efficient and cost-effective than complex vector databases for most users.

The memory layer ensures that information is not lost between sessions. Using the Carpathi RAG structure, raw data gathered during research moves into the wiki section as detailed articles before becoming an output like a slide deck. This setup allows Claude Code to navigate the file system with fewer tokens by following the map laid out in the Claude.md file.

Observability and the agentic dashboard

  • The dashboard turns terminal skills and automations into clickable buttons for non-technical users.
  • Clicking a dashboard button triggers a headless instance of Claude Code using the -p flag.
  • Observability widgets track usage limits, vault changes, and routine forecasts outside the terminal environment.
  • Customization of the dashboard enables AI agency work by providing clients with a simplified interface.

The dashboard serves as the visual front-end for the underlying architecture. It bypasses the limitations of the terminal by displaying real-time data like five-hour usage windows and recent file changes. This layer is specifically designed to empower team members who may be reluctant to use a command-line interface but need the power of Claude Code.

Community Posts

View all posts