YT YouTube2Text

YouTube2Text → Transcripts → The Agentic OS Setup That Will 10x Claude Code

The Agentic OS Setup That Will 10x Claude Code — Transcript

by Chase AI · 6,596 words · 916 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00If you don't know how to build your own
0:01agentic OS, then you are falling behind.
0:04But not for the reason you think. It's
0:06not because you need some fancy
0:07dashboard or a Jarvis setup because that
0:10is not where the value of these systems
0:12lie. The value is everything under the
0:14hood. Everything you can't see. I'm
0:16talking about loop engineering, skill
0:18architecture, state management, creating
0:20a second brain, and being able to bundle
0:23it all together into a coherent
0:25customized product that works for you.
0:27That is where the value lies. And the
0:29skills required to build something like
0:31this can be applied to any project you
0:33work on with clawed code, which is why
0:36this is such a valuable thing to
0:38understand. So in this video, we are
0:39going to be breaking down the clawed
0:41agentic OS construct level by level so
0:43you can learn not only how to build one
0:45yourself, but why it's so important that
0:48you do. When we are looking at an
0:49agentic OS like this one or this
0:52Obsidianbased one, it's easy to get lost
0:55inside the visual spectacle of it all.
0:58all these buttons and moving parts and
0:59metrics, things you can't find inside
1:01the Claude Code terminal, jump out at
1:03you right away, and you kind of fall
1:05into one of two camps. The first camp is
1:07like, "Wow, this looks really cool. I
1:08want to get my hands on this. I love all
1:10these like sort of visual things I can't
1:12find elsewhere." And the other camp sees
1:14it for what it is, which is just the
1:16visual interface, and thinks, "This is
1:19just smoke and mirrors. There's nothing
1:20here that's actually moving the needle
1:22forward." And I think both camps miss
1:24something. And what they miss are the AI
1:26fundamentals that are working under the
1:28hood that turn an AIOS and just a fancy
1:31looking web app into a customized weapon
1:34you can use to attack any problem with
1:36clawed code or frankly any model. I'm
1:38going to talk about cloud code today but
1:40this can be run again with something
1:42like codeex or even a local model. And
1:44when we talk about these fundamentals we
1:46can kind of break it down into four
1:47sections when it comes to this AIOS. The
1:50first level is the backbone and that is
1:52the skills in the loop engineering in
1:54the idea that we have codified
1:56everything we do inside of cloud code
1:58and we've turned it either into a skill
2:00or some sort of automation. Level two is
2:03memory and state control. How can we
2:05make sure that our AIOS has some sort of
2:08database of information that it can draw
2:10upon when we ask it questions? And more
2:13importantly, can we use this state, this
2:16memory, whether that's obsidian or
2:17something else. You can use something
2:18like a standard database. How can we use
2:20this in combination with the skills and
2:23automation we've built to create
2:25actually properly built loop engineered
2:28constructs that are somewhat
2:29self-improving? If you watched my last
2:31video about loop engineering, you kind
2:33of know what I'm talking about. The idea
2:34that we need to be able to record things
2:36and set up loops that can see how we've
2:39done in past iterations to improve
2:41future runs. And these two levels, one
2:43and two, is where we make most of our
2:45money. This is 90% of the value of any
2:48sort of AIOS. And once you have that
2:49locked in, well, then you can move into
2:52some of the cool visual stuff. And
2:53that's sort of level three where we talk
2:54about the interface and the UI and sort
2:56of the customization aspect of this. If
2:59you want to get outside of the terminal
3:00and also sort of spread your wings
3:02versus some of these desktop
3:04applications because the claw desktop
3:05application is great, but there's only
3:07so much you can do inside of it or set
3:08it up in certain ways. And then lastly,
3:10we have level four, which is
3:11distribution. Because the cool thing
3:13about this is it doesn't have to be
3:14personal just to you. You could share
3:16your AIOS with members of your team or
3:18even clients. And it's a great way to
3:20raise the floor in your organization. A
3:22lot of what you do on level one and two
3:25can be turned into a literal button or
3:28voice command that you can implant
3:30inside of your AIOS user interface that
3:33anyone can use. They don't even need to
3:35run cloud code themselves. It can all be
3:36done for them. So those are the four
3:38levels we're going to be talking about
3:39today. I'm going to spend the vast
3:40majority of it here in levels one and
3:42two because the truth is you could still
3:44do all of this inside of your standard
3:47claude code terminal or the codec cli or
3:49the codeex desktop app if you really
3:51nail down these first two parts which is
3:53skills and loop engineering and memory
3:54and state. This applies to any and all
3:57problems not just AIOS. But before we
3:59dive into level one, a quick word from
4:01today's sponsor, me. So, I just released
4:04my cloud code master class and it is the
4:05number one way to go from zero to AI
4:07dev, especially if you don't come from a
4:09technical background. We focus on real
4:11use cases. It's updated every single
4:13week and it includes all of my personal
4:15build. So, everything you see in today's
4:17video that I'm using for my demos, you
4:19can find it here as well. So, if you
4:21want to get your hands on all that, you
4:22can find it inside of Chase AI Plus.
4:24There will be a link in the description.
4:26So, level one, skill architecture, loop
4:28engineering, all this stuff. What are we
4:30really talking about here? Well, there's
4:31sort of four subphases of level one. We
4:35have the workflow audit, we have the
4:37skill creation, we have automation, and
4:39then we have loop engineering. And so,
4:41step one is a workflow audit. Before we
4:44can create skills, we need to know what
4:47you actually need to create skills for.
4:49Because remember, why are skills
4:51arguably the most powerful thing in
4:52claude code? Well, because it allows us
4:54to get a specific output. We're telling
4:56Claude to do a specific thing in a
4:58specific way for a specific output. But
5:00what sort of specific outputs do you
5:02need constantly throughout your
5:04day-to-day and your week to week? Seems
5:06like an obvious question, yet most
5:07people cannot answer that. And even if
5:09they can answer it, they certainly
5:11haven't taken those workflows that they
5:13do manually inside of Cloud Code and
5:15turned it into skills or automated them.
5:17So this is the first thing you have to
5:19do. And this, if you do nothing else,
5:21will totally supercharge how you work
5:23with cloud code. So here's a visual
5:25representation of what I'm talking
5:26about. We have you and we have Claude
5:28code. And for most people, it stays just
5:31like this. And it's a purely manual back
5:33and forth. You open up the terminal, you
5:35open up claw desktop, and you tell it to
5:37do certain things. And inevitably,
5:38you're telling it to do the same things
5:40all the time. Well, what if instead we
5:43did an audit of everything you do
5:46dayto-day and week to week and all the
5:47ways you use Claude and we codify that
5:50into skills? You're doing the same tasks
5:52over and over. Why don't we actually
5:55consistent with the outputs and how they
5:56work? Now, if you have used skills in
5:58any capacity, the cell here is very
6:00obvious. And I'm just telling you to do
6:02what you do with skills already, but do
6:04that a hundred times more because
6:06undoubtedly the way you use Claude Code,
6:08whether it's as an individual or as a
6:10business, can be broken up into a number
6:12of different domains. For me, that
6:13includes things like research, content,
6:15my online community, my agency, my
6:17sales, on and on and on. And underneath
6:20each of those domains are different
6:22things that I do, specific tasks for
6:26content that include stuff like, hey, I
6:28need to do outlines for all my projects.
6:30I need to figure out the hooks for my
6:31videos. I need to repurpose my content.
6:33I need to create carousels. On and on
6:35and on and on. Why are those not skills?
6:38Truthfully, they should be. And yet, for
6:41most people, they probably don't have a
6:43large robust set of skills like I do
6:45right here. And this is the easiest way
6:47to improve cloud code. Now the practical
6:49question is how do you do this at scale?
6:51And there are a few different ways we
6:53can do this. Number one is we do this
6:56purely via a manual process. That means
7:00I go inside of clawed code. I already
7:02have an idea of what it is I do. I
7:04explain that task and then I turn it
7:08into
7:10can't type. We turn it into a skill. And
7:12obviously we're going to use the skill
7:15creator skill to do that, right? That's
7:18simple enough. The issue with doing
7:20this, especially if you haven't done it
7:21before, is you probably haven't
7:24validated how Claude Code should go
7:26about this. The ideal scenario is you've
7:28done this manually. You've confirmed it
7:29actually works, and then you tell Claude
7:31Code, hey, see how we just did that
7:33task? Now, I want you to turn it into a
7:36skill. And there's actually ways you can
7:37do that relatively quickly because
7:39remember, Cloud Code has access to
7:42essentially all your previous sessions.
7:43It can see the tool calls. It can see
7:45what you told it to do. It has the full
7:46back and forth. We can really get a head
7:49start on creating this sort of like
7:51skill repository by telling Claude Code,
7:54hey, I want you to take a look at our
7:56last three sessions, five sessions, 10
7:59sessions, 20 sessions. I want you to
8:01pull out everything we've done and give
8:03me a list of things that we can turn
8:05into skills that I do all the time. So
8:08that becomes option number two, which is
8:10we have it look at previous sessions and
8:12sort of pull the work out of us. This
8:15way, it's actually going off real data.
8:17You know, this isn't a guessing game of
8:18what you think you should do in Cloud
8:20Code. It's actually going to take a look
8:21at what you've done. And so, that prompt
8:24can just sound something like this.
8:26Hey, can you go through our last 10
8:30sessions that I've had back and forth
8:32with you? And I want you to sort of pull
8:33out some repeated task or things we've
8:35done over and over again that aren't
8:37skills yet, but that I want to turn into
8:40skills and create some sort of chart
8:41showing what that task is, what the
8:43output should be, and the proposed
8:45skill. So, that's pretty much it.
8:47Doesn't have to be fancy. You can talk
8:48to it in plain language. And like you
8:50can see right here, what is it doing?
8:52It's going to find our session files
8:54first. Now there's actually a third
8:56option for how we can approach this.
8:59Option number three is we have it do an
9:02interview right we have Claude code
9:05interview us and we say hey I'm just
9:08going to give you a stream of
9:09consciousness about what I do daytoday
9:11week to week and I want you to ask any
9:13questions if there's any blind spots and
9:15then I want Claude code to pull out of
9:17that conversation these sort of tasks
9:19that can be turned into skills. So, same
9:22sort of idea. It'd be the same sort of
9:25simple prompt. We'll just pull it up
9:26right now. And then I would say
9:28something like,
9:31I'm trying to turn all my daily and
9:34weekly tasks into skills when they make
9:36sense. So, I'm going to start off by
9:38giving you a stream of consciousness of
9:40sort of like what I do each day. And
9:42then I want you to sort of turn it into
9:44an interview and call out any blind
9:47spots because at the end of this I want
9:48you to have as much context as possible
9:50about what I do and sort of the outcomes
9:52I'm looking for and I want you to pull
9:54out specific tasks and I want to be able
9:56to turn these tasks into skills and
9:58eventually automations.
10:02That's pretty much it. Doesn't have to
10:04be fancier than that. The whole idea is
10:07that we are just throwing as much
10:09context as possible about our work,
10:11about our day, about our week into cloud
10:13code and codifying it. We kind of want
10:16to just turn what we do into a
10:17checklist. This is the kind of the
10:19saying, the way you should approach this
10:20is like you brought in some person to be
10:22your personal assistant. You want to
10:24offload as many tasks onto them as
10:26possible. Well, how would you do that?
10:27Well, it's pretty obvious. You would
10:29tell them what you do and then give them
10:31step-by-step instructions to do that.
10:33That's all we're doing. We're just doing
10:35it with clawed code. And yet most people
10:38don't do this. And because these are
10:40skills, these are now tangible workflows
10:41that we can look at and edit as needed
10:43until we really dial in those outputs.
10:46And right here you can see one of the
10:47repeated tasks that I pulled out from
10:49previous sessions. Things like checking
10:51for tool and repo updates for videos.
10:53And it continues to go on and on with
10:54these tasks it found. Now creating these
10:57skills is just layer number one of this,
11:00right? We've codified it. And oftentimes
11:03these are things are going to get
11:04repeated over and over. Well, if they're
11:05going to get repeated over and over, is
11:07there any reason why we don't simply
11:09turn these into automations,
11:12right? Again, we want to move away from
11:13this manual approach to everything.
11:15Well, I had this task I did manually.
11:18Now, it's a skill. It's been codified.
11:19Well, now let's just set it up into an
11:21automation if it makes sense. And this
11:23is so easy to do inside of Cloud Code.
11:25We can quite literally just prompt it,
11:27can we turn that skill into an
11:28automation? And if you want a more
11:30visual approach, you can set this up
11:31very easily inside of Claw Desktop. We
11:33just go to routines. We give it a name.
11:36So that just be like auto one. The
11:39instructions would just be run this
11:42skill. Insert the name of the skill. And
11:44then we would set it on a specific
11:46schedule. That's pretty much it. And the
11:48third level here would be setting up
11:50some level of loop engineering. Now, I'm
11:53not going to go ultra deep into loop
11:54engineering because the video I put out
11:56yesterday certainly does, but we have
11:58sort of the foundation for creating
12:00strong loops. We have the skill. We've
12:03turned it into an automation. Now, it's
12:05simply a question of well, are we taking
12:08a certain automation and are we trying
12:09to add some sort of like
12:11self-improvement loop to it? And this
12:13will tie into memory and state as well.
12:15But that's all I'm really going to
12:16mention there. Just understand if you're
12:18someone who's trying to dive into loop
12:20engineering and that whole side of the
12:22equation, you're set up very well to do
12:24this given we laid the foundation with
12:26our skills and our automations. But
12:28zooming out here for a second, this is
12:31the backbone of everything. It's
12:34codifying your life and making it so we
12:37can get consistent outputs from clawed
12:39code for the tasks we actually care
12:41about. And as you can imagine, that kind
12:43of has nothing to do with these fancy
12:46dashboards and all this other stuff that
12:48comes along within a OS. I love the
12:50fancy dashboards. I think they're really
12:51cool, but this is the power. And you can
12:53imagine applying that to any problem and
12:56any sort of use case with cloud code is
12:58is very simple and straightforward. It
13:00doesn't require any of the other levels,
13:01which is why I think it's important to
13:03sort of be able to create your own AIOS
13:05because as we stack these levels on top
13:07of one another, you can see how sort of
13:09modular and flexible this is. So
13:12workflow audit talked about the three
13:13different ways to do that, right? We can
13:15do it manually, we can have it look at
13:16our previous sessions, or we can have it
13:17run an interview. Once we've done that,
13:19we have it create the skills. Then we
13:21ask ourselves, hm, can any of these
13:23skills be automated? And then lastly, we
13:25kind of have that discussion about does
13:26loop engineering make sense for this
13:28particular use case? And speaking of
13:29loop engineering, that brings us into
13:32things like state and memory, which is
13:35level two. Now, a lot of what I'm going
13:37to talk about today will be in the
13:39context of Obsidian, but understand it
13:41doesn't have to be Obsidian. Everyone
13:43likes talking about Obsidian because
13:44it's free and it's relatively easy to
13:47understand. Everything that works with
13:49Obsidian can also be done inside of a
13:51traditional database, right? Doesn't
13:53have to be Obsidian. Obsidian is just
13:55simple to use. And really, even more so
13:57than Obsidian, it's the idea of file
14:00structures and setting up cloud code in
14:03a coherent manner. Truthfully, you could
14:06probably have no database and you could
14:08have no Obsidian. And if you just set up
14:10Cloud Code with a file structure that is
14:12coherent and makes sense, you're like
14:1499% of the way there. Obsidian just
14:17makes it easy. And databases obviously
14:19have their own power that goes along
14:20with them. So, how should we set up this
14:23part of the equation? Well, we're going
14:24to answer that question through the lens
14:26of Obsidian and what the file structure
14:28should be doing for us. Now, Obsidian,
14:30like I mentioned before, is completely
14:31free. The idea with Obsidian is that you
14:34simply download it and then you're going
14:36to designate a folder on your computer
14:39as the vault. It can be any folder or it
14:41can be a brand new folder. Now, when
14:43you're installing Obsidian, you'll see a
14:44popup like this where again you can
14:46create a new vault or you can open a
14:48folder as a vault. Now, when we're
14:50talking about a vault, again, it's
14:51literally just a file. So, you need to
14:53ask yourself for this Agentic OS, what
14:56file should it live in? What file is
14:58going to have all the information I
14:59wanted to know about? So, if we're
15:01treating it like a personal assistant,
15:03perhaps one of the domains you're going
15:05to have it help you out with is like
15:06sales. So, you have a lot of sales data.
15:08Well, whatever file I designate as the
15:10vault, I'm going to put want to put at
15:12least a copy of all my sales data in
15:14there. That's sort of the idea. Now,
15:15once we've designated a folder as the
15:17vault, in order to connect Cloud Code to
15:19Obsidian, it's as simple as just opening
15:21Cloud Code inside Obsidian. So, I
15:24navigate inside my terminal to whatever
15:27I called that folder. In this case, I
15:29literally called my vault the vault. So,
15:32I'm now inside the vault. Next, I just
15:34open up Claude Code
15:38and boom, Claude Code for all intents
15:40and purposes is now connected to my
15:42vault. So, I now have my vault and
15:45Claude code is open and connected to it.
15:47Now, the question becomes, how do I
15:50actually set up my file structure inside
15:51the vault? This is not a trivial
15:53question. This is very important because
15:56the whole value ad for something like
15:57Obsidian is the idea that clawed code
16:00can be connected and opened inside of
16:03some folder that has a bajillion
16:07subfolders and a bajillion files inside
16:09of those folders. And you, the human
16:11being, can ask Claude Code a question
16:13about any of the files and folders
16:15inside of here and give you an accurate
16:18and quick answer. How do we do that?
16:21Well, it's going to all come down to how
16:23we sort of set this up, right? If we
16:25just have one folder with 10 million
16:27files in it and there's like no back
16:29links and nothing is connected and
16:31there's no sort of hierarchy, Cloud's
16:33going to struggle to quickly find you an
16:35answer. And when I mean it's slow, I'm
16:37also meaning it's going to use more
16:39tokens and ultimately it's going to cost
16:40you more money. So, we have to set this
16:42up in a clear manner. Your mental model
16:44here is setting up a map for Claude
16:47code. Right? We're looking at the
16:48knowledge graph of my Obsidian vault.
16:50These are all the files inside my vault
16:52and how they connect to one another. The
16:54ideal scenario is that when I tell
16:57Claude Code or ask it a question about
16:59something inside this giant morass of
17:01files, it has a very clear path to
17:04finding that file and therefore finding
17:06my answer. So think of Obsidian as
17:08essentially your filing cabinet for
17:11everything. Now there's become somewhat
17:13of a common way to set up your file
17:16structure inside of Obsidian when we're
17:18talking about cloud code. And this comes
17:20from Carpathy. This tweet from
17:22Cararpathy got over 20 million views and
17:24it was all about how he sets up his
17:27knowledge base using Obsidian so that
17:29large language models like Claude can
17:32access information quickly and
17:34effectively. And it goes something like
17:36this. We have the primary vault which
17:38you are now familiar with. And then we
17:40sort of have three subfolders. We have
17:43one subfolder which we call the slash
17:47raw folder. R A. This folder is where
17:50all unstructured data goes. Beneath that
17:54we have another subfolder called the
17:57wiki folder. The wiki folder is where we
18:00have taken the unstructured data. Think
18:02hey we just did all this research about
18:05I don't know AI agents right a bunch of
18:08unstructured data includes a bunch of
18:10articles and all that sort of stuff and
18:11we've turned it into structured data.
18:13We've now turned that into like a
18:16Wikipedia style article about AI agents.
18:20So the data and like what going on here
18:22and everything we want to know about it
18:23is very clear, right? This is instead of
18:26trying to go through, you know, 20
18:27different like source documents, we now
18:29have everything nice and neat under this
18:30wiki file. The third subfolder in the
18:33vault is essentially for outputs.
18:36So let's say I did my research about AI
18:40agents which is now in the raw folder
18:42folder number one. I then turn that into
18:45structured data into a Wikipedia article
18:47which is in folder number two. Now I
18:49want to turn that into say and let me
18:51move this over here. Let's say I now
18:52want to turn that into a I don't know a
18:54slide deck. Okay, some sort of clear
18:57deliverable. Right? We're not just
18:59talking about data here. We've actually
19:00turned it into something useful. Well
19:01that would be folder three. Okay. And so
19:04the idea is for most of the data we play
19:06around with, we can sort of set it up
19:08like this. Unstructured, structured, and
19:12then outputs. That is the Carpathy
19:15obsidian rag. And I say rag in air
19:17quotes because this is a new rag system.
19:20But that's one way to do it. Now the
19:22beauty of this Carpathy system isn't
19:24necessarily that we split it up into
19:27unstructured, structured, and outputs.
19:28The real beauty is that at every level
19:31of this, we have an index
19:35MD file, right? This is just a text
19:37document, a markdown file that is
19:40telling claude code at every level we go
19:43down what it's looking at. So, for
19:45example, if I'm talking to Claude Code,
19:48I'm talking to my AIOS and I say, "Hey,
19:49I want you to give me all the
19:51information about AI agents. Remember,
19:53we created a wiki article about AI
19:56agents."
19:57Well, first thing it's going to do is
19:59it's going to look in the vault and it's
20:01going to hit this index.md. And this
20:03index.md is going to say, hey, inside at
20:06this level of our system, we have a raw
20:08file for unstructured data, a wiki for
20:10structured data, and outputs for things
20:12like slide decks and that sort of thing.
20:13So, Claude Code knows right away, hey,
20:16he wanted information about AI agents.
20:18So, we're going to go to this wiki
20:19article. Now, inside of the wiki
20:21article, inside of the wiki folder
20:23rather, guess what's inside here? There
20:25is also an index.n MD. Now, we have one
20:29article in here. So, does it really need
20:32essentially a table of contents for a
20:34folder with one thing in it? No, of
20:36course not. But what if you've been
20:37using this for a year or two or five and
20:39you have not one or two or 10, but you
20:41have thousands of documents inside of
20:43here and potentially subfolders as well?
20:45Well, an index.md is going to make it
20:47way easier for cloud code to hit that
20:49folder and understand what it's looking
20:51at and where it needs to go. Because
20:52remember, what is the purpose of all
20:54this? is to give Claude code a map. And
20:57if it every single new room it enters,
20:59there's a clear spot it can go to and
21:02figure out what it's looking at, it's
21:03going to be faster and it's going to be
21:05cheaper. And it's important to
21:06understand that the power comes from
21:08that, not necessarily the somewhat
21:10arbitrary folders we created. It just
21:13needs to know where it's going. You
21:15don't have to do raw. You don't have to
21:16do outputs. You don't have to do any of
21:18this carpathy stuff. You just need a map
21:20for cloud code that makes sense. And
21:21it's probably going to be unique to you
21:23because your structures of your data and
21:25what you're trying to do are always
21:26going to be unique. So you have a better
21:30answer than anyone for how you should
21:31structure this. Now if you don't have
21:32that answer, guess who can help you?
21:34Claude Code. Simply tell to look at your
21:36vault and say, "Hey, what sort of
21:37structure makes sense? Oh, use
21:39Carpathy's obsidian rag setup for
21:41inspiration." That will kind of do
21:44everything you need it to do. The only
21:46other thing I would mention would be,
21:47hey, let's create a claw.md file that
21:50talks about this. In my claw.md file
21:52inside of my vault, it talks about my
21:54vault conventions, specifically the
21:56vault structure, aka what sort of files
21:58and folders it's looking at. You can see
21:59over here on the left, I don't just have
22:01three. I have several. I have content,
22:02notes, runs, inbox, ops, projects, etc.,
22:05etc. Furthermore, I have a whole thing
22:07about the navigation pattern,
22:08essentially saying, "Hey, when you're
22:09trying to find something, here's the
22:11path you should follow." And I think
22:13using that sort of template is very
22:14flexible. you can apply it to any sort
22:16of structure or any sort of data you're
22:18working with and you can come up with
22:20something that will be effective for
22:22you. So when we talk about level two
22:25that's what we're talking about. We're
22:26giving cloud code a map and it needs to
22:28make sense. Now this also plays into you
22:30know loop engineering because and skills
22:33and automations because all these
22:34outputs need to go somewhere and they
22:36need to be logged and they should be
22:37logged in a way which you know what I'm
22:39going to say makes sense for loop
22:40engineering specifically when we talk
22:42about self-improving
22:44skills and automations. Well, for that
22:47to work, we need somewhere where claude
22:51code can see what the past runs have
22:52done or really what the loop should be
22:54able to see what past runs it's done. So
22:55then it can make future improvements.
22:57And this should all be tied together in
22:59the same place. And if you master those
23:01two levels, you have 90% of the power of
23:05an AIOS already at your fingertips. And
23:07all this can be done through the
23:08terminal or the desktop app or anything
23:11because at this point you've kind of
23:12codified your workflows and now you have
23:14a way to actually see what's going on,
23:16record what's going on, sort of create
23:18your second brain and allow cloud code
23:20to pull out insights you otherwise
23:22wouldn't have in an efficient manner.
23:24Now level three is where we put a custom
23:27visual wrap around everything we've done
23:28up to this point. And we have a few
23:30options with how we do this. Obviously,
23:32it can be purely customized, but we can
23:34make this something that is purely web
23:36app based or something that is obsidian
23:38based. Now, when we talk about web
23:39appbased, we're talking about something
23:41like this. Again, this is cloud code
23:43under the hood. It's connected with
23:45Obsidian, but I now have a bunch of
23:47custom metrics that I have tuned for
23:49what I need to see. So, for me, on the
23:51left, it shows stuff related to my
23:53content creation, right? What's my
23:55YouTube subscriber count, Instagram, my
23:57latest video, my clawed 5 hour window? I
23:59have directives that are pulled from my
24:01Google calendar. I can look at documents
24:03that it's created. Over here on the
24:04right, I've taken a bunch of my
24:06automations and skills and I've turned
24:08them into just single buttons. So, if I
24:11just click on something like inbox
24:12brief, you can now see that it's cued.
24:15It's popped up right over here. And
24:17under the hood, Claude is running, going
24:19through my inbox, creating drafts, and
24:21it's going to let me know what it thinks
24:22is important. And the cool thing about
24:24this is again you can change whatever
24:27metrics are shown here to be whatever
24:29you want. The whole idea of this being a
24:32floor raising mechanism. The idea being
24:34I can now give some of the power to
24:35claude code to non-technical team
24:37members and clients. We'll talk about
24:38that in level four has a lot to do with
24:41what you see over here on the right
24:42where we've turned automations and
24:43skills into a button you can literally
24:46press. So instead of telling them, hey,
24:48learn how to use cloud code, here's a
24:49skill you need to install. Here's how
24:50you run the skill. Here's how you
24:51automate it. No, I'm just going to go
24:53ahead and I'm going to set up this AIOS
24:56for them. And now they can just click a
24:58button and it will do all that for them
25:00and it'll either dump it inside their
25:01own Obsidian or Teams Obsidian.
25:04>> Inbox brief is done. Inbox brief ready.
25:0832 threads triaged with a Gear Up
25:11contract and the Open AI merge campaign
25:14flagged urgent.
25:15>> All right, that's enough out of you. But
25:17yeah, you can hear in this case I also
25:18have a voice model attached to it. So I
25:20can talk to it. It can talk back. And
25:22that voice model is completely local, by
25:24the way. Right? That's running on my
25:26actual computer that isn't going out to
25:2811 Labs. So it's free. If I click on the
25:30inbox brief,
25:33it brings up the entire write up. And
25:35this too is something that I can open
25:37inside of Obsidian. And speaking of
25:39Obsidian, you can also create a command
25:41center, a visual layer inside of
25:43Obsidian itself. And that's what we see
25:45right here. So the metrics are a little
25:47bit different, but they're similar. I
25:49can see my token burn. I have, you know,
25:51the same sort of setup where I click a
25:53button and it runs skills or
25:54automations. I have different tabs. So
25:57for me, I want more insight into sort of
25:59like audience metrics that's going on
26:01the content side as well as some
26:02research stuff. So you can make this
26:05extremely customized. And again, the
26:07cell here at this point isn't the fancy
26:10visuals. It's that I can have a one-stop
26:12shop for a lot of different things that
26:14are a little harder to have visibility
26:16into when I'm strictly inside the
26:18terminal. Now, in terms of how you would
26:20create something like this, the web app
26:22is just like creating any sort of web
26:25app using cloud code. You're going to
26:26give it some sort of visual idea. I
26:28mean, obviously my exact setups you can
26:30find inside of Chase AI Plus, but you
26:32should just find some sort of, you know,
26:34website or setup you like. You can take
26:36a screenshot of this. You would dump it
26:38into clawed code and you essentially say
26:39like, hey, here's all the skills I
26:42already use. I want this connected to
26:44the vault. here are the metrics that are
26:46important to me that I want to see in
26:48one place. Let's go ahead and create
26:49this. Let's put a visual wrapper over
26:51all this. And the same deal goes for
26:53Obsidian. Obsidian runs in a plug-in
26:55system. So, you're pretty much creating
26:57an app, but it's specifically for
27:00Obsidian. And if you just tell Claude
27:03Code to hey say, hey, can you sort of
27:05take the web app we just created and
27:07create an Obsidian plug-in version of
27:09that? Again, it will give you something
27:10like this. And you just install it and
27:12run it from Obsidian. Now, a quick note
27:14about what's going on under the hood
27:16here. Now, under the hood, if I click
27:18one of these buttons, which again are
27:19related to skills. Let's say I click the
27:21morning brief skill. What's actually
27:22occurring? Well, this is essentially
27:24calling on a headless version of clawed
27:27code. So, it's just like as if I opened
27:29up my terminal and I have a version of
27:32clawed code running and it's now going
27:34to run, you know, forward slashming
27:37brief. The difference is when I click
27:40that button, it's doing a headless
27:41version of it. So, this terminal doesn't
27:43literally pop up on your computer. It's
27:45headless. It's invisible. And it uses a
27:48command called claude-p.
27:51Now, there was some drama with claude
27:53dashp not too long ago because Anthropic
27:55came out and said, "Hey, if you use
27:57claw-p, it's not going to pull from your
28:00cla subscription. It's going to pull
28:02from this $200 credit which is tied to
28:04API costs." That was kind of a problem.
28:06Although, they've sort of walked back
28:08from that and that isn't something that
28:10has occurred yet. So for now, this is
28:12still pulling from your max plan. So
28:13it's the same as if you opened up your
28:15terminal and ran it. And that's how
28:16we're able to create these sort of
28:17structures that still call on cloud
28:20code. We get all the power out of cloud
28:21code, but it's done invisibly behind the
28:23scenes. So when we talk about level
28:25three, what is this bias? Well, bias
28:26customization. We aren't locked into the
28:28terminal or the desktop app. We can have
28:30it show whatever we want. We also have
28:32the ability to give this to members of
28:35our team. This is what we talk about in
28:37level four, which is distribution. I
28:39mentioned a little bit earlier. If I
28:40hand someone this web app and it's tied
28:42to all these different skills, they're
28:44just one click away from getting a lot
28:45of power out of cloud code. Because
28:46remember, all the power comes from those
28:48skills and automations. If I make it
28:50super easy for someone to use those,
28:51well, it's kind of like spinning them up
28:53on cloud code without actually spinning
28:55them up on cloud code. The obvious
28:57question then becomes, well, how would
28:58you actually distribute to them? And
29:00there's a few options. When we're
29:01talking about something that's web-
29:02based, it's actually much easier. So
29:04giving something this somebody this
29:06versus giving them the Obsidian version
29:08is much simpler because since it's web-
29:10based I can put it on GitHub. I can, you
29:12know, create a whole zip folder. I it's
29:14very easy for me to transfer it to them
29:16and have it get get it up and running
29:18versus something like Obsidian.
29:21Obsidian is a little harder to work that
29:22way. So Obsidian would require a little
29:24more hands-on work from you. So if
29:26you're like, "Hey, I really like the
29:27sort of like Obsidian Command Center
29:29deal. How would I bring that to a member
29:30of a team?" Well, you would kind of have
29:31to set it up for them. It's not as
29:33straightforward. It's not much more
29:35difficult, but it isn't as simple as
29:36like, hey, there's a GitHub repo. Go
29:38ahead and clone it and point Cloud code
29:40edit. But again, the customization piece
29:42is a huge cell here, especially for
29:44those of you who are doing any sort of
29:46client work. I can't tell you the amount
29:49of people who want to use AI and want to
29:50use Claude, but are totally turned off
29:53or frankly just really scared of the
29:55terminal and even the desktop app. Like,
29:56it is a bridge too far for most people.
29:58If you're watching this video, you
29:59probably scoff at that. But I'm telling
30:01you, you live in a bubble that 99% of
30:03people just like won't go there no
30:04matter what you do. And being able to
30:06say, "Hey, instead I'm just going to
30:07throw you this and I'm going to set it
30:09up for you and you just either talk to
30:11it with a voice or you press a few
30:12buttons." It goes a long way. The sort
30:14of dashboard effect with a non-technical
30:17population like genuinely needs to be
30:19studied because it changes how people
30:20interpret these like technical tools.
30:22And so zooming out now that we've gone
30:24over levels three and four, you can see
30:26they're really just the cherry on top of
30:28the AIOS.
30:30Almost all the power, almost all of your
30:32time should be invested in these first
30:35two levels. The skills, the loop
30:37engineering is the automation, the
30:39codification, the memory, and the state,
30:41right? Can you do the same things with
30:43cloud code every time? Can you be
30:44consistent? and can we log those things
30:46and essentially create the second brain
30:48for cloud code that it can not only
30:50reference but use to improve upon what
30:52it already does. If you can figure out
30:55that then you're going to be well ahead
30:58of the general population when it comes
30:59to using this tool. So that is where I'm
31:01going to leave you for today's video. I
31:03hope breaking down AIOS and sort of
31:05these four levels made it a bit more
31:06clear about how these things work, where
31:08the value is, and how you can create
31:10something like this yourself. Like I
31:12mentioned earlier, if you want my exact
31:14setups, that's all inside of Chase AI
31:16Plus.

About this transcript

This page contains the full transcript of The Agentic OS Setup That Will 10x Claude Code by Chase AI, generated from the public captions YouTube serves with the video. The transcript has 6,596 words across 916 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.