YT YouTube2Text

YouTube2Text → Transcripts → Build & Sell Claude Code Operating Systems (2+ Hour Course)

Build & Sell Claude Code Operating Systems (2+ Hour Course) — Transcript

by Nate Herk | AI Automation · 36,834 words · 5,017 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00By the end of this video, you're going
0:01to know exactly how to build your own AI
0:03operating system, even if you've never
0:05even opened up Cloud Code before or some
0:07sort of tool like that. I'm even going
0:08to be giving you guys for completely
0:10free the entire setup guide. So,
0:11literally all you have to do is download
0:13that, follow along with what I do in the
0:14video, and you'll be up and running. If
0:16you guys don't know who I am, my name is
0:17Nate. I've been deep into the AI game
0:19for almost 2 years now. I scaled my last
0:21AI automation agency to over $100,000 a
0:23month, and then I sold it. I'm currently
0:25now running one of the largest AI
0:26communities in the world. We just hit
0:28350,000 members and this is a free
0:30community. I also run this YouTube
0:31channel and a lot of you guys have been
0:33asking me how I upload so many videos
0:34and how I stay consistent. It's because
0:36of my AI operating system and I want to
0:38give you guys access to the exact same
0:40thing that I use. So, let's not waste
0:41any time and just get straight into this
0:43one. I'm super pumped. Here we go. So,
0:45first of all, like what is an operating
0:47system? The first thing you might think
0:48of is like, okay, my Mac OS or my
0:50Windows iOS or my Apple iPhone iOS. It's
0:54basically the layer between you and your
0:56computer or your phone. It's where
0:57everything lives. It it has your files,
1:00your apps, your contacts, you know, it's
1:02the screen that you're looking at. And
1:03it's not really something that you think
1:04about too much because it just works and
1:06it lives there all the time. But now
1:08we're getting into this world where AI
1:09is taking over everything and it's
1:11seeping into everything. And now that I
1:13have an operating system inside of cloud
1:15code, it is my OS, but now it has agents
1:18on top of it. It has intelligence on top
1:19of it. So, I'm really not exaggerating
1:21when I say that I could spend an entire
1:23workday with just cloud code open and I
1:25could still do everything I need to and
1:27I would still be more productive than
1:28people that are clicking around in all
1:30of the different apps. So, when we add
1:31intelligence on top of our operating
1:33system, we have an AI that can see all
1:35of our files, all of our communication,
1:37everything important going on in the
1:38business. And not only can it see it,
1:40but it can also interact with it. And
1:41not only can it interact with it, but it
1:43can also remember it better than you can
1:45because you're a human and you use a
1:46brain and you forget things. But the AI
1:48not only has better memory, but it can
1:50pull things from the exact source and it
1:52can find it quicker than you can. Think
1:53about how much time you've spent on work
1:55about work, which is basically just the
1:57idea of searching for things. Your
1:59coworker drops you a file. You forget if
2:00that was in Slack or an email. You're
2:02looking for an old Google sheet that you
2:04knew you were working on last month, but
2:06now you can't find it because you worked
2:07on 15 more since then. But AI could find
2:09that instantly and grab it for you
2:11because it has all of the context, it
2:13has all of the connections, it has all
2:15of the capabilities, and it has the
2:16right cadence. So the combination of
2:18those four C's is everything that you
2:21need really to manage your business and
2:22you can get it to the point where it
2:24truly runs in the background for you.
2:26Now you guys have heard me say Cloud
2:27Code and obviously there's some other
2:28tools that we're going to be talking
2:29about today. But what's really important
2:31in this space is that you're building
2:32things to be tool agnostic because the
2:34tools change every 6 months. Those of
2:36you guys that have been with me for a
2:37long time on this channel, you know that
2:39last year I was all in on Naden and I've
2:42made the full pivot to cla code because
2:43I've realized I'm just way more
2:45productive with cloud code. I really
2:47don't think that I could have built this
2:48sort of AI operating system in nodn. I
2:51just don't think I could have the way
2:52that I can do it in cloud code. Models
2:54would be replaced. API endpoints might
2:56be deprecated. SDKs might be deprecated.
2:58Things will just happen. But we're going
2:59to be building this in a way that is
3:01very futurep proof because I'm going to
3:02change the way that you think. I'm going
3:04to change the way that you decide how to
3:06do things. I'm going to change the way
3:07that you build this AIOS so it can
3:09survive whatever happens. Just
3:11yesterday, I took my AIOS and I moved it
3:13over to Codeex just to make sure that
3:14everything was good. and it took Codex
3:16about 2 minutes to get adjusted. And now
3:18I have two different AIOS's, Cloud Code
3:20and Codeex. And if I need to put it in
3:21anti-gravity or if I need to throw it
3:23into the the other new kit on the block,
3:25I can do so. So I'm going to be teaching
3:26you the durable layer that sits
3:28underneath all these tools and all these
3:30different buzzwords. Okay. So there's
3:31two main frameworks that we're going to
3:33be talking about today. The first one is
3:34more of a highlevel framework around the
3:37way that I think about AI, and it's
3:38called the three M's of AI. We've got
3:40the mindset, the method, and the
3:42machine. All right. So, here is my three
3:44M's doc that I'm going to be reading
3:45over today. I'm not going to read every
3:47single word because that would take too
3:48long, but I will drop this doc in my
3:50free school community. The link is down
3:52in the description. I have all of the
3:53other resources that I've ever shared on
3:55YouTube in there as well for free. So,
3:57if you want to check it out in deeper,
3:58go there. But I wanted to spend some
3:59time here because the mindset and all of
4:02these frameworks I built out in a way
4:03that it's really important for you to
4:04think about not only when you're setting
4:06up your AIOS, but also as you use it
4:08every day and also as you scale it up
4:10and also as you start to, you know,
4:12bring it into other areas of your
4:14business because right now we're
4:15focusing on you personally, but now that
4:17you've got this set up, your business is
4:18set up better to be more AI ready and
4:20your employees and your co-workers are
4:22as well. So it's really important. So
4:24there's three habits I wanted to touch
4:25on. The first one is the default shift.
4:28basically the idea of you know before
4:30you do any task just ask yourself how
4:32could AI do this or at least how could
4:34AI do 30% of it because the truth is AI
4:38is not going to do everything 100% it
4:39might be able to do 75 or 50% but that's
4:42still a huge win and that is a
4:44productivity gain so here's a real
4:46example the other day I needed to update
4:48over 300 YouTube video description
4:50tracking links and the old me would have
4:53just been like man this sucks but I just
4:56have to do it and I would have manually
4:57clicked on each video, changed the link,
4:59and then clicked in the next video and
5:00just done that. Probably would have
5:01taken me an hour. But the new me, my
5:03default is, okay, I don't want to do
5:05that manually. Let me just brainstorm
5:06with cloud code and let me figure out
5:08how do I actually make that happen?
5:09Whether that's an API or whether that's
5:11an MCP or whether that is maybe even
5:13having to just do a browser automation
5:15script. But the point is, my default
5:18shift is so stubborn now where if
5:19anything sounds boring or repetitive,
5:21I'm not going to do it. I'm going to use
5:22my AIOS to do it. That's habit number
5:25one. Habit number two is the function
5:28breakdown. Understanding that your role
5:29is a set of functions. Everything you do
5:32is broken into a tiny task and you can
5:34just automate one piece of that at a
5:36time. So an example, okay, if I had to
5:38think about how can I automate an entire
5:39YouTube video. That's a very tall task.
5:42But break down the pieces that go into a
5:44YouTube video. Ideation is usually step
5:46one. And then you've got your scripting
5:47or you know your building or your your
5:49your slide creation. And then you've got
5:51packaging titles and thumbnails. And
5:53you've got descriptions. and you've got,
5:54you know, comment replies, that is a
5:57huge job. But when you break that into
5:59little things, oh, I can definitely
6:00automate ideiation or at least 95% of
6:03it. I can definitely automate scripting.
6:05I can definitely automate the slide deck
6:07creations for my videos. And when you
6:08have that mindset of just breaking it
6:10into little pieces, baby steps, it
6:12becomes way more achievable. So, in this
6:14example, if this is the full flow of
6:16automating a YouTube video, you just
6:17have all these little chunks. And as you
6:19slowly over time automate more and more
6:21chunks, you get to a place where you're,
6:23you know, now you've automated like 80%
6:25of a YouTube video. And each of these
6:27chunks is repeatable. So maybe this is
6:30like the, you know, slide deck chunk.
6:32Okay. What if we have a different
6:33process? Let's just say we're now doing
6:35like, you know, we're doing like meeting
6:37prep or something. Well, guess what? We
6:39can go ahead and grab this chunk and
6:41just move it over here and we can just
6:42fit this into this puzzle wherever it
6:44goes. So all of these little chunks and
6:46baby steps that you're taking are
6:48reusable and that mindset shift is very
6:52important to have. And then habit number
6:53three, we've got the curiosity rule.
6:55Never accept AI output without asking
6:57why. Treat AI as a mentor,
6:59[clears throat] not a vending machine.
7:01There's basically this idea of dark code
7:03because so many people nowadays are
7:04using AI to write code, which means they
7:06don't truly understand what each line or
7:09each block of code is doing. And not
7:11that you have to be able to read Python
7:12or whatever it is, but it is very
7:14important to at least understand why did
7:16you build this? And if X happened, what
7:19would happen to the code? You know what
7:21I mean? So the whole idea of being
7:22curious is asking those sorts of
7:24questions. Why did you design it this
7:25way? What happens if someone, you know,
7:27submits an invoice that's empty and this
7:29code is supposed to handle filled out
7:31invoices? And the other piece of
7:32curiosity is if you don't know if
7:34something's possible to be automated or
7:35automated with AI, then just ask. Okay,
7:38so moving on to method. This is really
7:40really important, but we're going to
7:41come back to this a little bit later
7:42once we've kind of, you know, got
7:43everything set up. But this basically
7:45just talks about, okay, how do you
7:46actually decide what's worth automating
7:48and how much of it you automate? So,
7:50there's some more information here about
7:51that kind of stuff. And then the last M
7:53stands for machine, and this once again
7:55gets a little bit more technical. All of
7:57this you guys will be able to dive into
7:58later, but we will revisit some of the
8:00stuff later on in this course. So, those
8:03are the three M's. Um, a couple of the
8:05key highlights that we already hit on.
8:06It's never binary. The question is never
8:08will AI do this for me. The question is
8:10to what extent can I leverage AI here?
8:1230% 60%, maybe it's zero, but it's very
8:15unlikely that it's zero. Every task on
8:17your plate has a leverage percentage.
8:19You just have to find it. And mindset
8:21isn't motivation. It's the lens that
8:22finds the percentage. The other one is
8:25treat AI as a mentor, not a vending
8:27machine. Vending machines take a coin
8:29and they give you something. But mentors
8:31ask you questions. They push you back.
8:32They make you sharper. They encourage
8:34you. They give you ideas. And that is
8:36the relationship that you want to have
8:37with AI, not a vending machine. And
8:39then, of course, we just talked about
8:40how your job is essentially a tree of
8:42tasks. And thinking about it like that
8:44makes big, vague, overwhelming ideas
8:48more realistic. So, another thing that's
8:50really important from the 3M is that
8:52productivity drops before it climbs. And
8:54I don't want you to get overwhelmed and
8:56just give up because you think it's too
8:57hard to build an AIOS. It is not. So,
9:00think about this. a solo operator and
9:03they are trying to build an AIOS and
9:05they're doing a 30-day sprint. This bar
9:08basically represents like productivity,
9:10right? So baseline, they are not using
9:13AI hardly at all. They're being maybe
9:1560% productive. Day three, they're
9:18slowing down. They're noticeably getting
9:20out less output because they're trying
9:23to adjust to this new way of living. And
9:25that's just how things work. Whenever
9:26you make a change, you typically are
9:28going to expect a 20% decrease in
9:30productivity. And let me actually just
9:32draw this out because I think it'll be
9:33easier to demonstrate. So if we just
9:35real quick go and draw a graph, then let
9:39me show you what this means.
9:42So let's say your business is operating
9:44like this. [clears throat] This is a
9:45steady line. Now if you decide to make a
9:47change, you have to basically be okay
9:48with the fact that this change might
9:50decrease your productivity by about 20%.
9:53So this is the, you know, sort of like
9:55negative 20% gap you're seeing.
9:59The question is, is this change worth
10:01it? Will this change ultimately result
10:04in more than a 20% dip? Because up here,
10:08what you might be getting with an AIOS
10:10is a 50% gain. So is the dip in 20%
10:13worth the 50% upside? Yes, because
10:16you're getting an additional 30%. Now,
10:18on top of that, you also have the idea
10:19of a learning curve. So, if I can just
10:22go ahead and draw another little graph
10:24real quick. And why am I being so
10:26boring? Let's actually use a color. So,
10:28the learning curve. People think that
10:30learning is linear like this. And it's
10:33just really not. Learning is
10:34exponential. But what does that mean? It
10:37means that when you start to learn
10:39something new, you basically are kind of
10:42below the linear line that you would
10:43have been if you were just doing it the
10:45old way. But because of the
10:47exponentialness,
10:49the exponential nature of learning, you
10:51will pass that mark. It's just that it's
10:53going to take you a few days to get
10:54there. Maybe this is day one, maybe this
10:56is day two, and maybe on day three, you
10:57break even. But by day four and five,
11:00you're way ahead of where you would have
11:01been if you just kept doing it the same
11:03way. So this gap is where people sort of
11:06tend to give up or get overwhelmed
11:08because they realize they're being less
11:10productive. So those two things are
11:12really important for you guys to
11:13understand. Don't give up. You get
11:15through it and then you start to see all
11:16the gains and now you are being 5x more
11:18productive, 10x more productive. I've
11:20seen that with myself. I've seen that
11:21with my team. It is an incredible thing
11:23to actually witness and by the time
11:25you're 2 weeks in, you won't even
11:26recognize how productive you're being
11:28and you can't even imagine if someone
11:29took away your AIOS and you had to go
11:31back to the old way. You would want to
11:33cry. Okay, so now let's get on to the
11:36second sort of framework here. And this
11:38one is called the four C's. This one is
11:40more specific to the way that we set up
11:42our operating system. And this is kind
11:44of the way that I'm calling the AISOS,
11:47the AI automation society operating
11:49system because we've basically just been
11:51getting everyone to do it in this way.
11:53And it is the AI automation society way.
11:55So what goes into that? Four C's. We've
11:58got context. What AI knows about you,
12:00your team, your tools, your voice, your
12:03business, your money. And then we've got
12:04connections. So what data can it reach?
12:07Cloud code on its own can't really do
12:08anything besides search the web. But
12:10your business data, your especially the
12:12important business data you need isn't
12:14available publicly on the web. So you
12:16have to give it connections to all of
12:18your different tools and all of the
12:19different databases where your actual
12:20important data lives. And then we have
12:23capabilities. What can it produce? What
12:25can it do with that data? You know, it
12:27has to be useful in some way, not just a
12:29knowledge source. And then we have the
12:30final C, cadence, when it acts on its
12:33own while you sleep and while your
12:35laptop is closed. And all of these build
12:37on top of each other. You can't have
12:38cadence without connections. You can't
12:40have capability without context. And you
12:42have to go in this order. 1 2 3 4. And
12:44I'm going to show you guys your exact
12:46path to get from 1 to four today. The
12:49whole idea is that if you opened up a
12:50new Claude chat right now or a new Cloud
12:52Code session, how much would it know
12:54about you if you asked a question? So
12:56test it. Ask a question and see does
12:59this answer like a teammate or like an
13:01executive assistant or is this answering
13:02like a stranger who just met you 5
13:04seconds ago. So here is the four C's of
13:06an AIOS. And once again, the important
13:08part is that tools change every 6
13:10months, maybe even quicker. But the
13:11platform that we're doing right now and
13:13the foundation that we're building is
13:15going to be able to move across whatever
13:17happens. And I think you guys understand
13:19what we're getting at here. We're
13:20setting up our context, which is
13:21basically the brain, making sure it
13:22knows our business. We're setting up the
13:24connections to make sure that it can
13:25touch things, whether that is an MCP or
13:27an API or a CLI or, you know, whatever
13:29it is. We're setting up the connections.
13:31The capabilities is that it actually
13:32knows how to do things. You know, you
13:34have a bunch of SOPs probably already in
13:36your business. Can you now give those
13:38SOPs to your operating system so that it
13:40can do it just as well and just as
13:42consistently as one of your employees
13:44does it manually right now? And then of
13:46course the cadence where it gets really
13:48really cool and it starts to feel more
13:49like a 24/7 AI employee for you. And
13:52those are the four C's of our AI
13:55operating system. And we're going to be
13:56talking about how you can actually test
13:57if you have those things in place. And
14:00we have a skill inside of the um
14:02template I'm going to give you guys
14:03where you can run an audit and just make
14:05sure that you're good and you can find
14:06opportunities to add more connections or
14:08add more capabilities and things like
14:09that. But for example, if you have
14:11connections into your calendar and your
14:13tasks and your CRM and your inbox, it
14:15should be able to grab all that and help
14:17you out. You could ask like, "Hey,
14:18what's on my plate today?" or "Help me
14:19plan my day." And it would be able to
14:21read through everything, look at your
14:22calendar, and help you block off things
14:25based on your priorities and based on
14:26maybe previous conversations you've been
14:28having in that week. And with the
14:29capabilities instead of having to drop
14:32in like a full, you know, two paragraph
14:34brief for what you want, you could
14:36basically say, "Hey, we need a quarter
14:38three report and I need it by the end of
14:39the day and it will be able to go do
14:41that because it understands your SOP for
14:43that. It understands all the business
14:44data and there's going to be less
14:45follow-up, less back and forth, less of
14:47you micromanaging it." So, here's a
14:49quick snapshot of the pillars, what you
14:51might want to do to test, what you would
14:53see, and kind of be like, "Okay, this is
14:55good. This was a pass." and what you
14:57might see and be like, "Okay, this
14:59failed. I need to do some work on this
15:01pillar." So, hopefully this is all
15:02starting to click. I'm not trying to
15:03hammer this home too hard, but context,
15:05connections, capabilities, and cadence.
15:08Okay, so now you have to actually start
15:10to get this into cloud code somehow. So,
15:13what I want you guys to do is think
15:14about these different buckets, right?
15:16Um, there's actually kind of seven tier
15:18one buckets, but they kind of fit into
15:20these groups of ops, comms, data, and
15:23planning. So, what I want you to do
15:24right now is I want you to grab a piece
15:26of paper or a Google doc or come into an
15:28Excal or a mirrorboard or whatever you
15:31want to do and just kind of layer out
15:34these seven things. We've got revenue,
15:37customer, calendar, comms, tasks,
15:40meetings, and knowledge. And the cloud
15:43code template I'm going to give you guys
15:44is trained on these seven things. But
15:46before we get into that, I think it's
15:48really important to just kind of sketch
15:49this out on paper. you know, do it in
15:52your head first before you jump into the
15:54whole onboarding of the AIOS or you
15:56might forget some stuff. And if you do
15:58forget some stuff, it's absolutely no
15:59big deal. Like I said, I've been playing
16:01around with so many different AI
16:02operating systems over the past multiple
16:04months and I've changed it a lot. I've
16:06evolved it a lot. I've had to restart
16:08before and it's kind of funny because,
16:09you know, I think the term AI operating
16:11system is very new and it's kind of
16:14ambiguous and a lot of people probably
16:15have a different idea of what that
16:17means. Um, you know, I think first I
16:19built out an executive assistant and the
16:21executive assistant has kind of evolved
16:22into my AIOS, but a lot of people have
16:25also been throwing around the term like
16:26second brain. And I think a second brain
16:28is a very great start to building an
16:30AIOS. So there's a lot of these things
16:32that you might have heard and this is
16:34just kind of my take on an AIOS, which
16:36is why I'm calling it the AIS OS. So
16:39anyways, you sit here and these are the
16:43tier one things that you should be
16:44thinking about as far as on your
16:47month-to-month, what are you looking at?
16:49What are you tracking? We're tracking
16:51revenue. We're, you know, communicating
16:53with customers. We're making sure
16:54they're happy. We're obviously having
16:55things come through on our calendar.
16:57We're communicating with people both
16:58internally or on meetings down here or
17:01externally. And we have project
17:03management. We have tasks. We have
17:04things to prioritize. and we have
17:07general knowledge. So when you start to
17:10think of each of these layers, it helps
17:11you figure out the connections you may
17:13need and the tools you may need. So
17:15that's what this phase is for. That's
17:16what this section is about. So revenue,
17:19let me just show you guys what I wrote
17:20here. I wrote school, Stripe, and
17:23QuickBooks. These are the places that
17:24I'm actually checking pretty frequently
17:26to see how many members we have, to see,
17:28you know, how much revenue we're
17:29growing, to see our promotions, to see
17:31how those are working, to look at
17:33QuickBooks, to look at our P&L, to look
17:34at our expenses. These are the places
17:36where I'm frequently checking to make
17:38sure my business is doing well, that we
17:39have runway, and that we're growing
17:41healthily. And then you have customer,
17:43you know, what does that look like for
17:44you? Do you have a SAS product? Do you
17:45have like a brickandmortar? Where do you
17:46have your CRM? Where do you have
17:48information about your customer, and
17:50where do you provide them value? So, for
17:52me, I thought of the two biggest ones
17:54that I'm checking, which are school and
17:56YouTube. And so, as you start to think
17:57about these different sources, these are
17:59where your connections are going to come
18:01from. And you want to start to think
18:02about how can I make sure that cloud
18:04code can talk to school, Stripe, and
18:06Quickbooks or School and YouTube. Every
18:08single one of these tools will have some
18:10way to connect the data to an AI system.
18:13We're living in 2026. It's possible.
18:15Some might have MCP servers, some might
18:17not. Some have API documentation, some
18:20may not. But there is a way. I promise
18:22you. Even if you have to spin up a
18:23browser automation, which I've made a
18:24video about before, you can get through
18:27in some way. Okay, so let's keep going.
18:29Calendar. This one was pretty easy for
18:30me. Basically, just Google Workspace. My
18:33Google calendar, everything lives there.
18:35Even if people book in through
18:36Calendarly, even if someone, you know,
18:38does something on my ClickUp calendar,
18:39it all syncs to Google Workspace. So,
18:41all I need there is one simple tool. For
18:43comms, I have two main ones, I guess
18:45three, actually, that I thought of. So,
18:47obviously Google Workspace for my email.
18:50I've got ClickUp, which is where we do
18:51pretty much all of our internal
18:52communication. And then I've got Slack,
18:55which is where we communicate with a lot
18:56of our vendors because sometimes they
18:58throw us in their Slack space or, you
19:00know, whatever the case is. So I need to
19:02make sure that my AIOS has visibility
19:04into all of those sources so it can pull
19:06data from there if it needs to. And then
19:08we have tasks. So pretty much our
19:10project management mainly lives in
19:11ClickUp for our internal projects, our
19:13internal tasks, our deadlines, our
19:15prioritization. And then we also have
19:16notion because sometimes we're working
19:18with vendors on different projects and
19:19things like that and maybe they're using
19:21notion which is pretty common. And as
19:23you're going through and filling this
19:24out, you don't have to get it all
19:25perfect on the first try. You're
19:27probably going to forget a few things,
19:28but that's okay. You know, start small.
19:31Start with the most important core
19:33things that you use the most, and then
19:34you build up from there. So, meetings, I
19:37use Fireflies at the moment, and that's
19:38really important for my cloud code to be
19:40able to look at those transcripts and
19:41see what meetings I've had because it
19:43can help me, like I said, plan things or
19:45respond to an email better or make
19:47action items better. And then finally,
19:49we have knowledge. And this one might
19:51seem a little bit more ambiguous, but
19:52just think about where else is important
19:54information that doesn't live in these
19:56sources that you've already talked
19:57about. And so for me, I was thinking,
19:59okay, YouTube transcripts, because up
20:00here, I was mainly thinking of like, you
20:02know, my YouTube comments or my
20:04analytics, but right here, I'm thinking,
20:06okay, YouTube transcripts. I want all of
20:08my knowledge from YouTube videos in
20:09Cloud Code somehow. And then I thought
20:11about Google Workspace, of course,
20:13because I have tons of Google Docs, tons
20:15of Google Sheets, tons of just like
20:17things in my drive, even videos that I
20:19want Cloud Code to be able to access.
20:20And then local files, which is pretty
20:22easy. You don't really have to do
20:23anything to connect that, but I've got a
20:25ton of local files that Cloud Code has
20:26helped me organize and move around and
20:28sometimes even send to people. Now, as
20:30you start to look at these different
20:32tools, think about this. You know, I
20:34joke around with people that if someone
20:36wanted to get a hold of me or if someone
20:38had a question for me, they would
20:39actually be better off asking my
20:41executive assistant because my executive
20:43assistant is also linked to all of these
20:45sources of data. So, it has all the
20:47knowledge I have, but it has a perfect
20:50memory and it never sleeps. So, if you
20:53can look at these sources and be like,
20:54"Okay, cool. If someone was able to talk
20:56to an AI agent that had all of this data
20:58and it could answer the majority of
21:00questions that my team or that you know
21:02you know some a customer might ask me
21:04then we're probably in a good spot to go
21:06ahead and get started. And if there's
21:07other things that come to mind when you
21:08sort of do that you know mental test
21:11then just chuck it down into the right
21:12category. And once you get to a point
21:14where you feel good about what these
21:16tools look like now what we need to do
21:18is we're going to get started building
21:20our AIOS. So, the very first thing that
21:22I want you to do is go to my free school
21:24community. The link for that is down in
21:25the description and join this community.
21:28Your membership will get approved as
21:29quick as we can. And then once you're in
21:31here, you're going to go to the
21:32classroom and you're going to go to all
21:34YouTube resources and you're going to go
21:36in there and download the different
21:37resources for this video. You're going
21:39to need a GitHub repo and you're going
21:40to need, you know, those other docs that
21:42I talked about, but you need that to get
21:44started. And then we're going to be
21:45doing this inside of Claude Code. And
21:47the way that I like to use Claude Code
21:48is in VS Code. You could also do the
21:50desktop app if you choose, but pretty
21:53much in all my YouTube videos, I like to
21:54use VS Code and that's what we're going
21:56to be using in today's tutorial. It's
21:57completely free to download. It is just
21:59an IDE and then we're able to open up
22:01Claude Code inside of it. So, I'm using
22:03VS Code today. Google Visual Studio
22:05Code, download it, and then the next
22:07step is you need to be able to connect
22:09your Claude accounts. And you do have to
22:10have a paid Claude account in order to
22:12use Claude Code. You can do the 17 bucks
22:14a month plan or I guess it would be 20
22:16on monthly, but you can start here. And
22:18if you're hitting your limits really
22:19quick and you want to scale up, then
22:21scale up. It's not a big deal. All
22:22right. So once you've done those two
22:23things, you're going to open up Visual
22:25Studio Code. And it will look like this.
22:28Now, what you're going to do first is
22:30you're going to install the Cloud Code
22:32extension. So on this lefth hand side,
22:34go over to this button, click on
22:35extensions, and then you're going to
22:37type in Claude Code. And once that pops
22:40up right here, you're basically just
22:41going to install this extension. Once
22:44you install that, it's going to prompt
22:45you to log in. And that is where you'll
22:47log in with your paid Claude
22:49subscription. Once you get that set up,
22:51you're going to see this button in the
22:52top right, which looks like Anthropics
22:54logo. And you're going to go ahead and
22:55just click on that. And what that does
22:57is it opens up this little Claude code
22:59agent. And this is very similar to the
23:01way you would chat with Claude or Chatbt
23:02on the web. You know, you have a little
23:04chat box here. You can say hello. And
23:06this is going to call on Claude Opus 4.7
23:09and it's going to respond to you. It's
23:10super simple. Now, what we have to do is
23:12we have to open up a project to work in.
23:14So, open up your file explorer, go to
23:16your desktop or your documents or
23:17wherever you want to have your AIOS main
23:19folder and just create a new folder.
23:21Just call it AIOS or, you know, whatever
23:24you want and create that folder on your
23:27file explorer. And then you're going to
23:28go up to this top left hand explorer
23:31button and it says you have not yet
23:33opened a folder. What you have to do is
23:34just open up that folder. And that is
23:36basically our working directory or the
23:38project that we're working in right now.
23:39So, you're going to click on open folder
23:41and open up that one that you just
23:42created. So, I just opened up one called
23:44AIOS demo. Yours should look exactly
23:47like this because there's nothing yet in
23:49that folder. So, all I'm going to do now
23:51is I'm going to close out of everything.
23:52I'm going to double click, open up Cloud
23:54Code, and now what we have is on the
23:56right hand side, we have our agent, and
23:58on the lefth hand side, we have our
23:59files. So, right now, obviously, there
24:01are no files in here. Now, what you're
24:03going to do next is you're going to go
24:04to my free school community, and you're
24:05going to grab the link to the GitHub
24:07repo, and you're going to grab the URL
24:09from it. And then you're just going to
24:11message Claude and you're going to say,
24:12"Hey, can you clone this GitHub repo
24:14into this current project? It's going to
24:16go ahead and search the web for it. And
24:18then it is going to pull it in and
24:20you're going to see a bunch of folders
24:21and files pop up on the lefth hand side
24:23right here. It's going to ask you if
24:24it's okay to allow this command. And
24:26you're just going to go ahead and click
24:27yes, allow get clone." And what you can
24:29see is now that it has been cloned, we
24:31have a bunch of these different files
24:32over here. So you can see that we have
24:34acloud folder, we have a archives
24:36folder, we have a context folder,
24:37decisions folder, a references folder,
24:39and then we have all of these other
24:40files right down here. So let me real
24:42quick take a second to familiarize with
24:44what we're actually looking at. The
24:46dotcloud folder, this is kind of like
24:47your overall project
24:50guide, I guess, claude. What's in here
24:53right now is we only have a folder
24:55inside of this called skills, which is
24:56why we have this little breakdown. And
24:59this is basically where all of our
25:00skills are going to go in the future.
25:02Now, what is a skill? A skill is
25:04basically just something that gives you
25:06six arms. No, I'm just kidding. But it's
25:08basically a file inside of Cloud Code
25:11and it's like a recipe. So, let's say
25:13you had a skill for building a LinkedIn
25:14post or writing a LinkedIn post. What do
25:16you do every single time you want to
25:17write a LinkedIn post? Maybe you do
25:19research and then you generate a graphic
25:21and then you write the copy and then you
25:23review it and then you post it. So, how
25:24annoying would that be to explain every
25:26single time to Claude Code that you want
25:28to do those five steps when instead you
25:30could just package those into a skill
25:32and then say, "Hey, Claude Code, write
25:33me a LinkedIn post." It would read your
25:35skill document and then it would go do
25:36everything. So, it gives you more
25:38predictable outputs, higher quality. So
25:41thinking of it like a recipe is a really
25:43good analogy because also let's say you
25:45build this skill or recipe and it comes
25:47out not the way you want it and then you
25:49realize okay next time we do this I need
25:51to add like one more egg so that the you
25:53know brownies are cakeier or whatever
25:55then you just update the skill say hey
25:57instead of one egg do two eggs and then
25:59next time you say hey run the skill it's
26:01better so the skills will ever be
26:03evolving if you want them to and it's
26:05going to help out a lot so if we open up
26:07this folder you can see that I already
26:09have given you three skills in here we
26:11We have one called audit. We have one
26:12called level up. And we have one called
26:14onboard. So in just a sec, we're going
26:16to run the onboarding skill to get you
26:18fully onboarded into this AIOS. But
26:20let's take a look at what skills
26:22actually are. So inside of the level up
26:24folder, we have a skill.md. And this
26:26skill.md, as you can see, is a very
26:28simple markdown file that explains what
26:31the skill does. We have a name, we have
26:33a description, we have what the skill
26:34does, we have what the skill does not.
26:36And then we say like, okay, here's what
26:38you do every time the user actually runs
26:40the skill. Phase one is the interview.
26:43Phase two is the method interview. Phase
26:45three is blah blah blah. It's basically
26:47just an SOP. And what's really cool
26:49about that is, let's say you have an SOP
26:51for onboarding a client. You could give
26:53Claude Code that SOP and give it the
26:55right connections to onboard a client
26:57and say, "Hey, can you turn this SOP
26:59into a skill?" And then boom, you've
27:00essentially got an automation for
27:02onboarding a client. But anyways, I
27:04don't want to get ahead of myself or
27:05start to confuse you guys. Let's just
27:07actually keep going down real quick.
27:09What else we've got in this project? So,
27:11we've got an archives folder, which is
27:13where Claude will put old documents or
27:14things that you don't need. We've got a
27:16context folder, which is where Claude
27:17will put different context files about
27:19your business, about you personally,
27:20about maybe the way that you like to
27:22communicate. Everything that it needs to
27:24know about you and your business, it can
27:25start to organize inside of this context
27:27folder. We've got decisions. It's going
27:29to keep a log of important things that
27:30you guys have decided to do together.
27:32And then we've got references. And this
27:33one right here has a markdown file about
27:36the 3 M of AI. Now, this is a really
27:39good place to start. And I built this
27:41template so it's kind of hackable. Over
27:43time, what's going to happen is you're
27:44going to build more and more files.
27:45There's going to be more folders.
27:46There's going to be different things.
27:47You might have projects. You might have
27:49quarterly folders. That can evolve and
27:51that is completely fine. The important
27:52thing is that Claude understands what
27:55the folders mean and what type of stuff
27:57goes in them. And that kind of stuff
27:59gets defined in a file called the
28:01claw.md. The claw.mmd is basically the
28:04master prompt for this project. So right
28:07here it says, okay, blank's name
28:09operating system. You are Nate's
28:10personal AIOS. Your job is to be their
28:13thought partner. Help them think,
28:14decide, and ship faster on this
28:16priority. And there's a lot of
28:17placeholders here that are going to get
28:19filled out as you start to grow this
28:21project. Now you can see there's a
28:23section here called your skills which is
28:25important because it has to understand
28:26when to invoke what skill and then we
28:28also have where things live. So context
28:31references connections and as you add
28:34more folders and files on the lefth hand
28:35side you're going to just want to make
28:36sure that your claused file is getting
28:38updated as well so that claude actually
28:40understands where things live. So
28:42there's more information in the cloudmd
28:43file. I'm not going to read it all right
28:45now. And this is another one of those
28:46files that evolves every day. My clawmd
28:49file in my project has probably been
28:50updated two times a day for the past
28:52multiple months because I keep adding
28:54new things or I keep taking things away.
28:56So none of this is set in stone. Don't
28:58stress yourself out. As you learn more,
29:01your project is going to grow and learn
29:03more, too. And then what else we have
29:04down here are just a few other markdown
29:06files that are important. You know,
29:08we've got some connections here that are
29:09going to get filled out. We've got some
29:10expansions that can help you figure out,
29:12okay, once you finish Nate's tutorial
29:14that you're watching right now, what do
29:15you do from here? How do you make this
29:16better? So, those are just some other
29:18files that are in here, and I don't want
29:20to read all of them out, but hopefully
29:22that little overview gave you at least a
29:24bit more comfort of what you're looking
29:25at. So, now let's actually go ahead and
29:28get onboarded. So, I'm just going to say
29:30in natural language, all right, I just
29:33cloned this repo. I want to set up my AI
29:35operating system. My name is Nate. Can
29:37you help me get onboarded into this
29:39project? And if you say anything like
29:41that, it should understand to invoke
29:43that onboard skill. And the cool thing
29:45about cloud code is you can basically
29:46see what it's doing. So it said, "Okay,
29:48I'm going to read the project's
29:49onboarding doc to see how I can set you
29:52up." And we can see that it's reading
29:53things. You know, we can see the exact
29:55files that it's actually reading. And
29:56then it says, "Okay, welcome Nate. The
29:58repo comes with an onboard skill that
30:00does exactly this. It's a seven question
30:02interview that scaffolds your day one
30:04file set. Since the skill was just
30:06cloned, let me read it and then I'll
30:07walk you through it." So this is where
30:09it starts that onboard questionnaire.
30:11The first question is, "Who are you?
30:13What do you sell and who do you sell it
30:15to? So, keep in mind because I'm just
30:17doing this as a demo for you guys, my
30:18answers are going to be pretty short.
30:20They're going to be kind of vague. But
30:22for you, I would really spend some time
30:24here and give it context. You know,
30:26obviously it's not the end of the world
30:27because you can you can continuously
30:29iterate, but I would definitely spend
30:30some time here and give it more than
30:32just a couple sentences. Give it a few
30:33paragraphs. And what makes it even
30:35easier is if you use a voice dictation
30:37tool. So, the one I use is linked in the
30:39description. It's called Glido. You
30:40might see that right here. I'm actually
30:42using Whisper Flow, but I have fully
30:44switched over to Glido and I am an
30:45official member of the GLO team now. I
30:47just believe in that product much more.
30:48It's faster, it's private, it's way more
30:50agentic. So, if you guys are looking for
30:51a voice tool and you want to support,
30:53then definitely check out Glido. Windows
30:54support is coming very soon. All right,
30:56so my name is Nate. I run a AI
30:59automation YouTube channel. I've got a
31:01free community and a paid community and
31:02we're working on building out some other
31:04different types of programs. And um our
31:07offer is basically just we're teaching
31:09people how to learn AI automation
31:11regardless of their background. We're
31:13helping them figure out how they can get
31:14a career or we're helping them figure
31:16out how they can get clients. And that
31:18is pretty much the type of person that
31:20we're currently appealing to. Okay. So
31:21that is the short paragraph that I'm
31:23going with for now. Now it's going to
31:25save that to the intake file, which is
31:27great. And that is actually this one
31:28right here. So as you answer, it's going
31:31to save these answers right here. So as
31:33you can see, question one was who are
31:35you? and this is what I answered with
31:36pretty much right here. And then it's
31:37going to keep going down the list and
31:39asking us these questions and it's going
31:40to keep filling in our answers here so
31:42it can reference them later. So anyways,
31:44let's keep going. Question two is to
31:45paste one or two things that you've
31:47written recently. Don't edit them. Paste
31:49them verbatim. And the reason it's doing
31:51this is because it wants to understand
31:53the way that you speak, the way that you
31:54talk. So I'm going to go grab one
31:56LinkedIn post and one school post and
31:58I'm going to drop them right here in the
32:00chat. All right. So I'm dropping in
32:01those two posts. Now, think about as far
32:04as like the way that you would want this
32:06to understand your writing styles. If
32:07you wanted it to just be LinkedIn, then
32:09you should say, "Hey, these are both
32:10LinkedIn posts." Or maybe you wanted to
32:12give it, hey, these are two emails I
32:14sent to clients. These are two emails I
32:15sent to my team. I have different types
32:17of communication based on who I'm
32:18talking to. And that's completely fine.
32:20You can give it that information. The
32:22more the marrier. Anyways, question
32:24three. What are your two to three
32:25biggest priorities for the next 90 days?
32:28So, I'm just going to make up a few
32:29different priorities here. you would
32:31obviously give whatever you're really
32:33working on for Q3 or for Q2. What is
32:35your current sprint? What are some
32:37milestones that you're currently working
32:38towards? So, I don't think that you guys
32:40really want to watch me answer these
32:41next couple questions. So, go ahead and
32:43finish all seven and then I'll meet you
32:45back here when you're done with that.
32:48Okay. So, it now comes back and it says
32:49that day one is done. Your AIOS knows
32:52who you are, what you sell, what matters
32:53this quarter, and how you sound. So,
32:55today you could say, "What should I
32:56focus on this week?" Actually, let's
32:58just try what should I focus on this
33:02week and we'll see what it says.
33:04Tomorrow it says you can pick one tool
33:05from your connections and wire it up.
33:07And then on day seven, you can run your
33:08audit to see your four C's score. So, if
33:11you really want to take it slow, you can
33:13do that. But, we're going to keep moving
33:14forward and we're going to start setting
33:15up some other things as well today. But,
33:18as I ask it, what should I focus on this
33:20week? Obviously, it knows because we
33:22just had this conversation. But what's
33:23important is that it's pulling from
33:25these files. It's in the context folder
33:27right here. And we have these new three
33:29files that were added. We have about
33:31business, we have about me, and we have
33:33priorities. And it just read from those
33:35two together in order to see that we
33:37have AIS live on July 11th. We have our
33:40new program in development and team
33:42management and making sure that we can
33:45keep tasks on track and everything like
33:47that. And then it says, if I had to pick
33:49one thing for Monday, it would be the
33:50AIS live speaker lineup and where could
33:53the default shift apply here? to what
33:55extent could AI be leveraged on this
33:57task? So, as you guys see, like I said,
33:59I've trained this thing to be your
34:00mentor. Not just saying, okay, here's
34:02what you should do or here's what I
34:03found, but trying to train your mindset
34:06to actually shift. So, you're starting
34:08to think about this exact type of
34:10question and multiple other questions
34:11that come with using an AIOS first.
34:14Remember, your goal should be, how can I
34:17do everything I need to do right in
34:19here, right in VS Code. I know I have my
34:21Chrome tab open right here. So, if I
34:23needed to go in here and click around, I
34:25could. But how could I try to be as
34:26productive as possible just in this
34:28interface? That is the default shift.
34:30It's the whole cliche of, you know, like
34:32you only get out what you put in. So,
34:34um, yeah, if you don't really commit to
34:36using this, then you're not going to
34:38feel the ROI of the AIOS. Let's take a
34:41quick peek inside of these. These are
34:43all just markdown files. So like about
34:44business, if I open this up, we can see
34:46that it has access to our offer, our
34:49ICP, it has our services, and it has our
34:51revenue model. This is obviously very
34:53very vague at the moment. It's very
34:55minimal, but remember over time as you
34:58make decisions, as you launch new
34:59offers, as you, you know, make pivots,
35:02this will continue to evolve. Same thing
35:04with the about me doc. So right here, it
35:06just knows a little bit about me and
35:07about some of my top pains. But I could
35:09also give it some more actual context
35:11about me. I could tell it my age, where
35:12I went to college, you know, what I like
35:14to do in my free time. I can give it
35:15more data here. And then of course we
35:17have the priorities doc. And this one
35:18will always evolve. I think the best way
35:20to be doing this would be at the start
35:22of every quarter. You know, our
35:23business, we work in quarterly sprints.
35:25Not everyone does that. Maybe you work
35:27in two week sprints or maybe you work
35:28in, you know, yearly sprints, whatever
35:30it is. But find a cadence that you want
35:32to give to your AIOS so it knows the big
35:35milestones that you're working towards.
35:37And it makes it really easy because for
35:38us inside of ClickUp is where we have
35:40all of our milestones for that quarter.
35:42So really the better way for me to tell
35:43my AIOS about my priorities would just
35:46be to say hey go to the ClickUp
35:48workspace called you know Q2 OTAAS and
35:52just read everything and those are our
35:54priorities for this quarter. And that's
35:55where you guys are starting to
35:56understand the value of the connections
35:58because right now this thing only knows
36:01what we've told it because you know
36:03that's the source of truth for it. But
36:05once we start to connect it to all of
36:06these things that we kind of sketched
36:08out earlier, once we connect it to
36:09these, it won't have to keep asking us
36:11questions because the default will be
36:13instead of asking Nate for answers, I
36:16will just go look in school or I will
36:18just go pull the data from QuickBooks or
36:19I will just use my GWS tool. That is
36:22what it's able to start to do. All
36:24right, so let's say that's where we're
36:25at now. I'm going to go ahead and do a
36:26slashclear which basically just clears
36:28the conversation and I'm going to
36:29pretend it's the next day. Okay. Hey,
36:31good morning. It's day two and I now
36:33want to start to actually connect you to
36:34things. You know, what are the most
36:36important things that I could start to
36:37connect you to so you can get more
36:39context about my business and we can
36:41just start to grow your database of data
36:44that you get to access. Now, if you
36:46already know, okay, like ClickUp,
36:48Google, Fireflies, those are the three
36:50most important things, then you can
36:51start wiring that up right now. But you
36:53can also have a conversation with it.
36:55You can be curious. You can ask it what
36:57it thinks. and you can start to, you
36:59know, just kind of layer on top of each
37:01other the different connections.
37:04Now, when you think about an actual
37:05connection, if you're not technical,
37:06don't worry about it because all that
37:08happens is you say, "Hey, I have
37:10ClickUp. Go do research for me and
37:12figure out how to connect ClickUp." And
37:14it just does it. The one thing you will
37:16probably have to do is you're going to
37:17be the one that has to go grab the API
37:20key and you're going to have to give
37:21Cloud Code the API key so it can
37:23actually access your stuff. It's
37:24basically just a password. So what this
37:26came back and said, "Good morning, day
37:28two. This is the exact right move. I
37:30looked at your connections doc and I can
37:31see your priority order." So here is the
37:34three things that you can do together.
37:36ClickUp for tasks, ownerships and
37:38deadlines and DMs and channels.
37:40Fireflies for all of your meetings and
37:42then Slack to get communication from
37:44other vendors. Now I actually disagree
37:47with this priority probably because I
37:48didn't give it enough insight. I think
37:50ClickUp is definitely number one for us
37:52and then for me number two would be like
37:54my Google workspace. But anyways, let's
37:56just start with ClickUp so I can show
37:57you guys how that works. It is a massive
38:00tool for us obviously. All right. So,
38:02one other thing that you can think about
38:03when it comes to ClickUp or you Gmail or
38:07Slack or whatever it is that you're
38:08giving your agent access to is that you
38:11don't have to give it access as you
38:13because your account probably has full
38:15permissions. It can see every space. It
38:17can download all the data. it can write
38:19whatever it wants. In some cases, you
38:22probably do want that, but in other
38:23cases, why not just create an account
38:26for your AIOS? What I did here is I
38:28created an account called Up AI. And now
38:32I give my uppai API key to cloud code
38:36rather than my own personal API key. And
38:38you can do this for all of your other
38:40tools as well. Maybe you want to get a
38:42separate API key inside of QuickBooks
38:44and you only want to give it read
38:46access. You know, you can do certain
38:48things like that to restrict the ability
38:50of the AI to make sure you don't have a
38:53situation like I forget which company it
38:54was, but I just saw some news article
38:56where an AI deleted like a really big
38:58database and you just want to make sure
39:00that that doesn't happen. So, per API
39:03key or per account, you can set
39:05different permissions. So, think about
39:06that. And what else that means is if you
39:09have a bunch of different AI agents,
39:10each of them can have their own API key.
39:12and if they're spending money on that
39:14platform, you know, whether that's a
39:15research platform or just an AI model,
39:18you can then see which automation is
39:20using how much money. So, think about
39:22the way that you separate out your
39:23permissions and your keys. But anyways,
39:25this is my uppi account of ClickUp. What
39:28I would do here is I would go to um up
39:30here in the settings. I would click on
39:32settings. I would go to the ClickUp API.
39:36And now I would just copy this token.
39:38And this is what I need to be able to
39:39give to Claude Code. And once it does
39:42that, it can access ClickUp. Now,
39:43obviously before that, it's going to
39:45say, "Hey, can we connect the ClickUp
39:46MCP or, you know, how do you want to do
39:48this?" A lot of these AI models are
39:50going to want to default to using an MCP
39:52server. As you can see here, how would
39:54you prefer to wire it? MCP server. It's
39:56the fastest if one exists for ClickUp. I
39:58don't like using MCP servers because
40:01basically what happens when you do that
40:03is it gives you access to every single
40:04function and every single endpoint
40:06that's possible, but you always don't
40:08need all of them. And also having a
40:10bunch of MCP servers loaded into your
40:12project actually eats more tokens. So
40:14what I prefer to do is API endpoints. So
40:17here's what I'm going to say and this is
40:18what you should probably say to all of
40:20the different integrations that you need
40:21to connect. I want to use ClickUp's API
40:24because it's more token efficient than
40:26having the MCP server. So what I want
40:28you to do is do research on ClickUp's
40:31documentation about all the different
40:33API endpoints. it would probably be
40:36helpful for you to set up a reference
40:38guide, a markdown file inside of this
40:40project that has all of the endpoints
40:42stored so that later if you need to use
40:44a different one, you don't have to go do
40:45research again. You can just reference
40:47that file. But anyways, go do research
40:49about that and then create me av file
40:52and I will give you my ClickUp API key
40:55in that file. So, let me just kind of
40:57like translate what I meant by that.
40:59It's going to look up the ClickUp API
41:00documentation. So, if I search ClickUp
41:03API documentation, it's basically going
41:05to find all of this and it's going to
41:07look through it because there's
41:08different endpoints to get a token or to
41:10create a task comment or to, you know,
41:13make a list or to remove things from a
41:15list. Every single function has a
41:17different endpoint. And every single
41:18time that Claude would need to do
41:20something in ClickUp, it would have to
41:21come research it. But instead, if we
41:23just let it research all of this once
41:25and then save it as a markdown file,
41:27markdown's really easy for AIs to read.
41:29It's also very cheap for them to read.
41:31So now we have our own sort of database
41:33of every single possible ClickUp
41:35function and we have that stored
41:36locally. So that is the research that
41:38Cloud Code is doing as we speak and it
41:40is going to save that for later. Now
41:43what is AENV file? The ENV file is
41:46basically just a secret file, right? So
41:50this is what it looks like. This is
41:51where you're going to store all your
41:52secrets. So we will input a ClickUp API
41:55token and a team ID. Now the reason why
41:58we have a dot is because it's aenv and
42:00it gets excluded from anytime we do a
42:03public push to a repo or anything like
42:05that. So it basically just means that we
42:07are protecting this API key. It's it's
42:09much more secure for you to paste in
42:11your API key into the ENV rather than
42:15you know giving it in the chat history.
42:17You know sometimes it might say hey you
42:18know just drop in your API key and I'll
42:20set it up. Don't do that. just tell it
42:22to create the env with a placeholder and
42:25then it's so simple for you to go into
42:26ClickUp, copy this, and then you would
42:28literally just paste it right here, hit
42:31save, and now Cloud Code can use it. And
42:33if you ever get confused about where you
42:36find your API key or how any of this
42:38works, just ask Cloud Code to do the
42:39research, and it will figure it out for
42:41you 99% of the time. Also, guys, I
42:43realized that I switched over this mode
42:46to bypass permissions. So, there's
42:48basically different modes in Cloud Code.
42:50What I like to do typically is start on
42:52plan mode if I'm going to build like a
42:53skill or I need help brainstorming. Um,
42:57but if you're on edit automatically and
42:58you're trying to do something like
42:59research like you just saw, it might
43:01stop constantly and say, "Hey, can I do
43:03this web fetch? Can I do this? Can I do
43:04this?" And if you get annoyed of that,
43:06you can switch to auto mode or you can
43:08also switch to bypass permissions. So
43:10auto mode uses a little bit more tokens
43:13because it basically uses an AI to
43:14analyze what am I about to do? Is this
43:17safe? If it's safe, I'm just going to go
43:18ahead and do it. But if I'm doing
43:19something like a delete or if I'm doing
43:21something like a push or anything that
43:23might be a little bit risky, then I'm
43:25going to stop and ask Nate or I'm going
43:27to stop and ask the user. Bypass
43:28permissions just says, "Okay, you can go
43:30do everything." I have never had an
43:33issue with bypass permissions. It's it's
43:35deleted things, sure, but only when I've
43:37asked it to. But you do run that risk of
43:40full autonomy. So, just wanted to at
43:42least let you guys know about that. The
43:45way that you can enable bypass
43:46permissions is if you go to the settings
43:48down here and you click on settings and
43:50you type in claude and then you can see
43:52right here allow dangerously skip
43:54permissions. If you turn that on, you
43:56should now be able to see this option.
43:57But you could always just use auto if
43:58you're scared and then eventually if you
44:00feel more confident you can switch to
44:01bypass. But that's why you might have
44:03seen that down there. Okay. So it says
44:05that ClickUp is wired. It says that we
44:07have a full V2 reference. So if I come
44:09into the references, we have a ClickUp
44:11API markdown file. And if I look in
44:14here, this is a very comprehensive doc.
44:16It is pretty long. So if cloud code does
44:19use this, it's going to have to read all
44:20of these lines, but it has all the
44:23endpoints. And now it has basically full
44:24understanding of how to use ClickUp. So
44:26let's go ahead and check it. It says, do
44:29you want me to run a test per assenee
44:30workload snapshot right now? It will see
44:33who's holding what across all 17 people.
44:35I'm just going to go ahead and say sure.
44:37and we're going to see if it's able to
44:39use my API key that I gave it, use the
44:42team ID that I gave it, and use the
44:44right endpoints to find that
44:45information. And the cool thing is as
44:47you're testing out these different
44:48endpoints or these different skills, if
44:50it hits a mistake, it's going to fix
44:51itself. But what else is good about that
44:53is it found an edge case because every
44:55time that it fails, it learns and it can
44:58update something. So if this fails, I'm
45:00going to say, okay, update the API doc
45:01so that next time you do this, it never
45:03happens again. Or if you were if this
45:04was running a skill right now, like
45:06let's say we ran a skill called team
45:08check-in, I would say, okay, can you go
45:10update the skill with the issue that you
45:12just ran into, how you fixed it, and to
45:15make sure that you never ever run into
45:16that issue ever again. So honestly,
45:18think of a failure as a good thing
45:20because it gave you more data about what
45:22not to ever do again. Another one of
45:24those super healthy Nate mindset shifts.
45:27Okay, so I'm not going to show you
45:30exactly what's on the screen right now.
45:31I'm going to have like a blur over this
45:33or a box over this because I don't want
45:34you guys to know what I'm up to. But all
45:37of this is legit. This is exactly what's
45:40going on in our ClickUp. And at the end
45:43of this, you guys can't see. I'll show
45:44you this line, but it says the three
45:46things worth doing. It gives me a number
45:48one, which is definitely seems like I
45:49should probably check up with this
45:52person right now. We have number two and
45:54number three, which all make sense in
45:56the context of my business right now.
45:58So, that is proof that that's working.
46:00Let's do one more quick thing. I'm going
46:02to go ahead and just do a /cle.
46:03Actually, I'm just going to open up a
46:04new tab and I'm going to say shoot Nate
46:07a message in ClickUp. Actually, I don't
46:09know why I'm not speaking. ClickUp and
46:12just say, "Hey, YouTube. Making sure
46:14that this is working." Okay, so this
46:16will be interesting to see if it decides
46:17to use a channel or if it decides to
46:20just go ahead and DM me directly.
46:24But what is it doing? It's reading the
46:26connections doc and then it's reading
46:27the ClickUp reference dock and then it
46:30is going to shoot me off a message. Now,
46:32there's one thing that I would say here
46:33that you might want to consider is if
46:35you have like a couple main API
46:38endpoints that you hit, you probably
46:40want to put those somewhere that makes a
46:41little bit more sense. So, for example,
46:44we don't want it to read every single
46:46time that entire doc because that could
46:48just waste tokens. But what's
46:49interesting is, you know, let's take the
46:51example of the team check-in. Whenever
46:52we run the team check-in, we could say,
46:54"Okay, when you ran that, what endpoints
46:56did you use?" And it would say, "I used
46:57endpoints X, Y, and Z." Okay, put
47:00endpoints X, Y, and Z in the skill file.
47:02So, you don't have to look at the full
47:04API documentation whenever you want to
47:06do that specific skill. So, I hope you
47:08guys are starting to understand here why
47:10it's a little bit difficult to give you
47:12like a step-by-step playbook for setting
47:14up your own AIOS because everyone runs
47:16their business differently. Everyone has
47:18different connections. But I'm trying to
47:20give you guys the way that I think about
47:22the, you know, the mindset of setting
47:24this kind of stuff up. So anyways, it
47:27decided to create me a task. So it
47:28didn't completely understand what I
47:30wanted. So in my ClickUp, my real Nate
47:33ClickUp, not the UPAI version, I got a
47:36task here and the task says, "Hey
47:38YouTube, making sure this is working."
47:39And it assigned the task to me. So it
47:42misunderstood my request a little bit. I
47:43was hoping it would send me an actual
47:45DM, but we can see here that it was able
47:47to create a task and assign it to me.
47:49So, at least it understands that. And
47:50really, what I was trying to prove to
47:51you guys there is that it was able to
47:53use a different endpoint as well. That
47:55is ClickUp. Now, before we set up
47:57another one and before we talk about
47:58some other stuff, I wanted to show you
48:00guys these skills that I have built in.
48:02So, the first one is called audit. So,
48:04I'm just going to go ahead and do /
48:06audit to show you guys what this skill
48:08actually does.
48:10This is basically going to judge your
48:12AIOS. It's going to look at the four
48:14C's. So, context, connections,
48:17capabilities, and cadence. And it's
48:20going to tell you what you might be
48:21missing, and it's going to tell you what
48:22you need to set up. So, right here, it
48:23says, "Okay, I'm running the four C's
48:24audit. Let me scan the project
48:26structure." And it's going to come back
48:27with a grade. And this is really great
48:28because if you're getting overwhelmed
48:30with what you should be doing, just run/
48:32audit and see where the gaps are and
48:34kind of help plan out your next steps
48:36from there. Okay, so April 30th audit,
48:39we got 54.5 out of 100 and we're
48:43currently, you know, we literally just
48:45got this set up. So, honestly, I think
48:47this is being pretty generous with
48:48context. I mean, 18 out of 25, it could
48:50definitely get a lot better.
48:52Connections, we're at 16 out of 25. It
48:54could get a ton better. So, maybe we
48:56need to tweak this actual audit a little
48:57bit because we only have one connection
48:59and that's ClickUp. But the point being,
49:02it's going to start to go over your
49:03strengths and weaknesses. So, ClickUp
49:04got wired the same day. Okay, maybe it's
49:06doing relative because it's like, okay,
49:07on day one that's not too bad. The
49:10connections.mmd is comprehensive. We
49:12have ClickUp, Fireflies, Gmail, Slack,
49:14and revenue. But here are the top three
49:17gaps and they're ranked by leverage. So
49:19tier one domain is thin, meaning only
49:21one of seven tier one domains are
49:22reachable. So if we think back to this
49:24where you worked up, we had seven of
49:26these, right? Revenue, customer,
49:27calendar, comms, tasks, meetings,
49:29knowledge, and we've only hooked up
49:30ClickUp. So, it's a little bit of comms,
49:32a little bit of tasks, but still not
49:34anywhere where we want to be. So, it
49:36says, "Why are fireflies next?" Then we
49:39have no recurring trigger or cadence,
49:41which is the fourth C. We'll get into
49:42that later. And zero useruilt skills and
49:46zero agents. So, that's really where we
49:47get a lot of the leverage is when you
49:49come into this AIOS and you build your
49:51own custom skills on top of everything.
49:54Now, speaking of custom skills, let's go
49:56ahead and also run this other one. This
49:58is called level up. So, I'm going to go
50:00ahead and do a slash and I'm going to do
50:01the level up skill. And I actually
50:03didn't realize that it suggested that it
50:04said, "Hey, these are structural gaps to
50:07explore capability gaps, which is what
50:09your AOS could do that it can't do yet.
50:11Run level up after this audit." And
50:14also, you can see it says that do you
50:15want me to save this? So, every single
50:17one of your audits can get saved so that
50:20you can see how you're actually
50:21improving your AIOS over time. This
50:24level up is basically going to analyze
50:26your business, your connections,
50:28everything that you're doing. And it
50:29might even ask you questions, but it's
50:31going to show you some areas where you
50:32could just level up your stuff. So, it
50:35says pulling from your stack. You have a
50:36top pain of managing your team. You have
50:38quarter 2 priorities. You have things
50:39that are reachable now, and you have
50:41things that are reachable soon. So,
50:42here's what you should do. Answer these
50:44questions however feels natural. One
50:46sentence is fine. Walk me through this
50:47past week. What did you do three plus
50:49times? Drudgery. Is there anything that
50:51felt manual, boring, or copy and paste?
50:53Smart intern test. Anything where you
50:55thought a smart intern could absolutely
50:57handle this, but you ended up just doing
50:58it yourself because explaining it would
51:00take longer. Constraint. If 500 new
51:02community members showed up on Monday,
51:04what would break first? And growth
51:06lever. What would give you 500 more
51:08clients tomorrow if it were running on
51:10autopilot? So, if you answer these five
51:13questions, there is absolutely no way
51:15that you're stuck. There's absolutely no
51:17way that you think, I don't know what to
51:18add to my AIOS. I don't know how to make
51:20this better. Every time that you answer
51:21these five questions, it will find an
51:23opportunity for you to automate
51:25something, to build a skill out of
51:26something, to connect something else.
51:28And that is why I built these two skills
51:29to go together, audit and level up, to
51:31make sure that you're constantly
51:32improving and that you don't ever feel
51:34alone and you don't ever feel stuck with
51:36your AIOS. So, I'm not going to type
51:37that right now. I'm not going to answer
51:38that cuz we have a lot of work to do on
51:40our own. But, I want to show you guys
51:42something I did recently with Claude
51:43Code and Google Workspace. So, I live in
51:47Google Workspace, right? Um, everything
51:49that I do is pretty much inside of there
51:51besides ClickUp. And what happened was I
51:53realized that across my company,
51:56everyone had spun up so many different
51:58Google Docs and so many different Google
51:59Sheets and we had tracker sheets all
52:01over. And Google search is horrible.
52:04It's absolutely horrible. So, I was able
52:06to connect Cloud Code to the GWS CLI. It
52:08is one tool and it gives you access to
52:10everything in the GWS environment. So,
52:12I'm going to show you guys exactly how
52:14to set that up in just a sec here, but I
52:15was able to do some really cool things.
52:17So, let me just show you guys a few of
52:18these cool things that I was able to do.
52:20If I go to my sheets, I have this video
52:22database that has all of my videos. It
52:25has the thumbnail, the title, a link to
52:27them, and a summary, and all of the
52:29resources that are associated with that
52:31video. And this is now my video
52:33database. And I don't do this manually.
52:35Every time I post a video, the video
52:36gets pulled in, the resources get pulled
52:38in, and then a row gets updated on here.
52:40And I built this because it was able to
52:42search through my entire drive. It could
52:44find all my YouTube videos. That was a
52:46different connection. That was a YouTube
52:47data API. But anyways, it could find all
52:49that and then it could find all my
52:51resources and then it could just put
52:52everything in here. Super super
52:54powerful. Another example, we have this
52:56pretty massive doc about our AIS live
52:59event and there's speaking slots and
53:01there's, you know, different timings and
53:02there's different outreach things and
53:04there's a lot of data here. And I
53:06basically said, "Hey, Claude Code, go
53:07read this document." I gave it the link
53:09to this. It opened up the doc in Google
53:11Drive. It read all of it and then I
53:12said, "Create me a database, a tracker
53:16sheet." And it created this Google sheet
53:18because it analyzed everything that was
53:19going on in that doc. And it was able to
53:21label everyone. It it made all of these
53:23have drop downs with color coding. It
53:25assigned these tasks to everybody. It
53:27was able to just do everything for me.
53:29And this would have taken so long for me
53:30to do manually. Another thing I was able
53:32to do is I have like all my YouTube
53:34videos stored locally and it was able to
53:36search through that, organize
53:37everything, date everything, and then
53:39upload all of that to Drive. And then it
53:41deleted everything locally. And now what
53:43I could do is I could have it organize
53:44all of these into different folders
53:46based on the date. So it can just help
53:48you organize things, move things around,
53:50create things. And all of this living
53:52inside of my Google Drive, inside of a
53:54shared company, Google Drive is awesome
53:56because it can find things instantly.
53:59And everyone else on my team, I was
54:00like, "Hey guys, can you use the GW CLI
54:03real quick and can you look through
54:04everything you've ever built for the
54:06company and can you put it into the
54:08shared drive?" And now all of our AI
54:10agents are able to touch all of the
54:11things so much better. So the GWS CLI is
54:15a huge unlock. Unfortunately, if you use
54:17like Microsoft and that is your main
54:19environment, the GW CLI probably won't
54:21be too useful for you. I'm sure there's
54:23something different for Microsoft's
54:24environment that you should use. But I'm
54:26going to play a quick clip for you guys
54:27about how to set up the GWS CLI and a
54:30little more information about it. So if
54:31that doesn't appeal to you or not
54:32interesting to you or you already have
54:33it set up, then go ahead and skip past
54:35this next part. But I'm going to play
54:36this clip real quick. Google just
54:38dropped what some are already calling
54:39the most powerful workspace CLI on the
54:42internet. So if you've got a ton of
54:43stuff that lives in the Google
54:44environment just like I do, then you're
54:46going to love this because now any of my
54:48cloud code projects can access
54:49everything. And all I had to do was
54:51install one simple thing. So here you
54:54can see I said what can you do with GWS
54:56which is Google Workspace CLI. So it can
54:59search, list, upload, download, move,
55:01copy, share anything in my Google Drive.
55:04It can do anything in my Gmail. It can
55:06do anything in my calendar. It can do
55:08anything with Google Docs. Same thing
55:10with Sheets. Same thing with Slides. And
55:12it also has multi-step workflow recipes.
55:14So these are basically like skills.
55:16These are chain command patterns for
55:18common tasks like creating docs from
55:20templates, reading sheet data, and
55:21creating a report doc, finding free
55:23time, and scheduling a meeting. And
55:24there are over a hundred of these that
55:25it actually has. So out of the box, when
55:27you give Claude Code the GWS CLI, you
55:30can do anything across any of the tools.
55:31And you also have access to over a 100
55:33skills. So, I don't know how many times
55:35you guys have tried to use something
55:36like Claude or Naden to build you a
55:38Google doc. And you do this over API and
55:40it ends up just looking like something
55:42like this. It literally just looks like
55:43raw markdown and it's obviously
55:45horrible. And sometimes to go along with
55:47a YouTube video, I make resource guides
55:48that look like this, but obviously they
55:50have to be formatted. I've got like a
55:51header up here and I've got links and
55:53different things in this format. But now
55:56I can literally just take the link to a
55:57YouTube video. I can drop that into
55:59Cloud Code and say, "Create me a YouTube
56:01resource guide." It's going to go ahead
56:02and download that transcript from the
56:04video. And now what it's doing is it's
56:05creating the Google doc, not via API
56:08call, not via MCP, but via a bash
56:10command, meaning it's literally running
56:12a terminal command in order to talk to
56:14Google and make this. So, it just
56:16actually created the doc. Here's the ID.
56:18And now it's going to populate it with
56:19what I need. And now it finished this
56:20up. It gave me the link. I'll click into
56:22this. And we can see, boom, we have an
56:24actual resource guide. It's got the
56:26image inserted up here as a header. It's
56:27got a link that goes right back to my
56:29YouTube channel. It breaks down the
56:30market traditional automation. It goes
56:32through all this stuff and then even has
56:33my CTA at the bottom as you can see
56:35after all these horizontal lines to join
56:37the plus group. So that was obviously
56:39just one quick example, but there's so
56:41many different benefits here using this
56:42workspace CLI. The first one is that you
56:44have one interface. So basically, like I
56:46said, it was one GWS CLI that cloud code
56:48now has access to and it can access my
56:50Gmail, my drive, docs, sheets, calendar,
56:52admin, and more. It's also JSON first
56:54with structured responses. So our AI
56:56agent is really, really good at working
56:57with it. We have auto discovery, meaning
56:59the CLI is pretty much always going to
57:00stay up to date automatically. Pretty
57:02much zero maintenance because we
57:03authenticate and then we're going to be
57:04good to go. It has built-in skills for
57:06triage, for prep, for generations. Like
57:08I said, there's a hundred others. And
57:10it's not much overhead because it's
57:12basically just one tool. It's not the
57:14same as like having all these different
57:15API endpoints or all of these different
57:17MCP configs and tools that would take up
57:19more context. But I know you're probably
57:21wondering, what is a CLI? It stands for
57:22command line interface. And what we're
57:24typically used to is a GUI or a
57:26graphical user interface where we can
57:28see buttons, we can see form fields, and
57:30we can click on things and that's how we
57:32navigate, but computers are more
57:34navigating by text and by commands and
57:36by typing. So that's really all that a
57:38CLI is. And this is an open- source
57:40Google Workspace product, and obviously
57:42it's completely free. So I'll leave a
57:44link to this GitHub repository down in
57:45the description if you want to check it
57:46out, read more about it. But I'm also
57:47going to walk through some of the key
57:48details right here. The first thing that
57:50I wanted to show you is if you go down
57:51here to the skills, this is where we can
57:53actually see all of the different kind
57:55of recipes they call them for pre-made
57:57multi-step workflows that it has. As you
57:59can see, creating events from sheets,
58:01creating presentations, creating meat
58:03space, label and archiving emails.
58:04There's so many different patterns that
58:06you might use from this pre-built
58:08library. Now, if we keep scrolling down,
58:10what you'll also notice is that right
58:11here it says this is not an officially
58:13supported Google product. Now, that
58:15doesn't mean that it's unsafe. This is
58:16an actual Google product, but the reason
58:18why it's not officially supported is
58:20because right now it's more of like an
58:21open- source beta. It's kind of a
58:23developer playground rather than like an
58:25enterprisebacked software. And you can
58:27see right here that it also says, "This
58:29project is under active development.
58:30Expect breaking changes as we march
58:32towards V 1.0." So, this thing is
58:33already really good out of the box and
58:35it's only going to get better. And you
58:36can see, like I said, when Google
58:38Workspace adds an API endpoints or
58:39method, GWS picks it up automatically.
58:42So, you might as well chuck it into
58:43cloud code right now and start getting
58:44used to it. Okay, so I just uninstalled
58:46this so I can walk you guys through step
58:47by step how this actually works. It's
58:49super easy. What I do is I basically
58:51copy the link to this GitHub repository
58:53as you can see. And I'm going to
58:55basically just give it to Cloud Code and
58:57say, "Hey, I want to install this GWS
59:00CLI, read through the documentation, and
59:02help me install everything that I need
59:04to install, and then we're going to get
59:05set up." So, this is basically going to
59:07do all the research for me, and then all
59:09I have to do is follow its instructions.
59:11So, it read the docs. It's looking at
59:12what we already have installed. It
59:14basically saw that I already had some of
59:15the prerequisites. So if you don't have
59:17those, you'll have to install those. And
59:19then it told me that we needed to
59:20install the CLI. So it did that. And now
59:22we have two options. So the first one is
59:24to install G-Cloud CLI so that we have
59:25automatic setup and off. Or we could do
59:27it manually by creating our own project
59:30and whatnot. So let's just go ahead and
59:31try option A. Okay. I thought this was
59:33going to be just like a simple command
59:35that it ran and then we were good. But
59:36it's actually like some other thing to
59:38install. So let's actually go back and
59:39try manual and I'll just show you guys I
59:41guess the harder way. Okay. So I'm going
59:42to go to this link. go to our Google
59:44Cloud Console and make sure you're
59:45signed in with the right account up in
59:47the top right. And I'm just going to go
59:48ahead and create a new project just to
59:50show you guys what this would look like.
59:51So, new project. I'm going to call this
59:53one Claude Code GWS. And we're just
59:58going to go ahead and create this
59:59project. So, this is spinning up right
1:00:01now as you can see. And now that it has
1:00:03been created, I'm going to select it so
1:00:04we're inside of it. And then I'm going
1:00:05to go up here and type in APIs and
1:00:08services. Click on that. And we have to
1:00:09set up our OOTH consent screen. So, I'll
1:00:11click on this. and it's going to say get
1:00:13started. Click on that. We have to give
1:00:16our app a name. And then we have to
1:00:18choose an audience. So, I'm just going
1:00:19to do internal because I only need this
1:00:21right now for my own organization. If
1:00:23you want to do external, it'll basically
1:00:25have you do testing or published. And if
1:00:27you do testing, just make sure that you
1:00:29add your email as a test user. And then
1:00:32all you have to do after you put in your
1:00:33contact information is hit I agree. And
1:00:36then you go ahead and create that. Now,
1:00:37once that has been done, you're going to
1:00:39go to create a client ID. So, I'm going
1:00:41to go back into APIs and services. I'm
1:00:43going to go to credentials and then I'm
1:00:45going to go ahead and do a create
1:00:46credential oath client ID. Now, in here,
1:00:48we're going to choose a desktop app. I'm
1:00:50going to just call this GWS and go ahead
1:00:53and hit create. And now, we have our
1:00:55client ID and our client secret. And so,
1:00:57what you're going to do is download this
1:00:58as a JSON file. Now, you can see here
1:01:00that it says to download that file and
1:01:02save it to your global.config/GWS.
1:01:06So, basically, if you can't find this,
1:01:08just say, "Hey, can you give that to me
1:01:09in a full path?" And then you can paste
1:01:10that into your finder or your file
1:01:12explorer and it will take you there. It
1:01:14will probably look something like this.
1:01:16And then you just drag in that
1:01:17credential thing. I called mine client
1:01:19secret. And cloud code will be able to
1:01:21look at this globally now. And so what
1:01:23you'll notice is that we didn't in this
1:01:24project yet enable these APIs. So let me
1:01:27just show you what happens without that.
1:01:28So it says the last step is to run GWS
1:01:31off login. So I just said, hey, I
1:01:33finished option B. The credentials are
1:01:34called client secret. And then I told it
1:01:36to run the O login. So that should
1:01:39basically open up a tab for you, but if
1:01:41it doesn't, then you can ask for it to
1:01:43give you that URL so that you can
1:01:44actually authenticate in. So you would
1:01:46basically choose your account that you
1:01:47want to use. And then you just have to
1:01:48basically confirm that it can access all
1:01:50of these different things as you can
1:01:51see. And then when you hit allow, you
1:01:53should be properly authenticated. After
1:01:55that, it's going to come back and say,
1:01:56"Okay, cool. Let me see if everything
1:01:58works." Now, this hasn't been perfect on
1:02:00the first try every time, but if you
1:02:02just go back and forth a little bit,
1:02:03say, "Hey, that didn't work. Hey, this
1:02:04is what I'm seeing." It will be able to
1:02:06get you there. It's going to be your
1:02:07best friend for something like this
1:02:09because remember it can read all of the
1:02:10actual documentation. And now it says
1:02:12that the O is working, but we have to
1:02:14enable these APIs in our Google Cloud
1:02:16projects. So basically just clicking
1:02:17open these one at a time and all you
1:02:19have to do is hit enable. So it's super
1:02:21simple. You just have to do this like I
1:02:23said for all of these different services
1:02:24that you actually want to be able to
1:02:26use. So that's why I did this on a new
1:02:27project cuz I just wanted you guys to
1:02:28see that. But if you already have one
1:02:30that has all these enabled, then you can
1:02:31just use that project and generate that
1:02:33OOTH client ID. So there you go. You can
1:02:35see that this works. I said, "Can you
1:02:37find my Google doc that I made in April
1:02:38of 2025?" And I went ahead and pulled
1:02:40links to all five of these because
1:02:42obviously that was a very vague request.
1:02:44And now we could take action pretty much
1:02:45anywhere in Google Workspace super
1:02:47simply with this CLI. But like I said, I
1:02:50just got this set up today and I've been
1:02:51playing around with it a ton in my
1:02:52executive assistant project and it's
1:02:54been awesome. It can literally do
1:02:55anything. So here I'm asking it to grab
1:02:57my unread emails from today and based on
1:02:59what it knows about my business and my
1:03:00priorities, give them a score and if the
1:03:03priority score is below five, just mark
1:03:04it as unread automatically. All right.
1:03:05So, here you can see it said, "Got 30
1:03:07unrated emails. Here's my priority score
1:03:09based on your business context." And as
1:03:11I scroll down, you can see that it's
1:03:12getting different ratings. And based on
1:03:14what I'm seeing right now, this actually
1:03:15looks pretty good. So, then I started
1:03:17playing around with Google Slides
1:03:18because I use Gamma right now. But at
1:03:20some point, I could imagine that if this
1:03:22gets good enough, then I wouldn't need
1:03:23Gamma anymore. And this is a free option
1:03:25compared to Gamma subscription. So, I
1:03:27had it create me a slide deck and it was
1:03:28okay. I threw in my brand guidelines. I
1:03:31threw in my logo and I said, "Hey, can
1:03:32you see this? you created this using the
1:03:34Google slides and it's okay but there's
1:03:36some weird things that I need you to
1:03:37fix. So then it came back and said I
1:03:39cannot see the slides I just know how to
1:03:41build them programmatically so that's
1:03:42why there may be some errors with
1:03:43spacing and stuff. So then I basically
1:03:45just gave it access to ChromeDev tools
1:03:47so that it could open the page
1:03:48screenshot it look at it and then we
1:03:50built a plan to add visual validation to
1:03:52this Google slide creator skill. So now
1:03:54you can see as it's going through it
1:03:56actually takes screenshots and then it
1:03:57can make fixes based on that. So then
1:03:59after it fixes everything, it says,
1:04:00"Okay, cool. Updated the skill. Take a
1:04:02look at it now." So I'll open up this
1:04:04link. Brings me to Google Slides where I
1:04:06have this slide deck. It has kind of my
1:04:08brand colors. It's got the logo up top
1:04:10right. And then as we go through, we can
1:04:12also see that the spacing is a little
1:04:13bit better. It's still not perfect
1:04:14obviously, but we have custom images
1:04:16here that were generated with Nano
1:04:18Banana 2. And even the images are kind
1:04:20of on brand with the sort of orange and
1:04:22blue color scheme. As you can see, we've
1:04:24got this one with the WT framework.
1:04:26We've got this slide. And it even ends
1:04:27with a CTA for the free school
1:04:30community. So just to see what else
1:04:31happens, I'm going to say take a look at
1:04:32the slide deck and do another audit. How
1:04:34could you improve the skill in the
1:04:35future? So it's going to go ahead open
1:04:38up a tab as you guys just saw. It's
1:04:40going to take images. It's going to
1:04:41flick through the different slides and
1:04:42capture them. And as you can see over
1:04:44here, it now says take screenshot. And
1:04:47now it's reading that screenshot right
1:04:49there.
1:04:50Now it just moved on to the next slide.
1:04:52And it's going to go through and look at
1:04:53every single slide. And then it's going
1:04:54to come back with a plan. And we could
1:04:56probably do a similar visual and
1:04:57validate flow with creating Google Docs
1:05:00as well. So now you can see it's almost
1:05:02on to that last slide. And I hope it
1:05:03fixes this last slide because what you
1:05:05can see here is that the spacing is
1:05:07really off down here. So you can see it
1:05:09came back with an audit. It came back
1:05:10with some future improvements. And one
1:05:12thing that I did notice is that because
1:05:14I made the window smaller, its
1:05:15screenshots were probably worse quality.
1:05:17So it said presentation mode screenshots
1:05:19would probably be better. But anyways, I
1:05:21just wanted to give you guys a little
1:05:22taste of how you can use the GWS CLI.
1:05:25but also use it with other tools to make
1:05:27the functionality even more powerful.
1:05:28So, just remember that this is very new.
1:05:30There's a lot of people out there on
1:05:31Twitter right now saying that this is
1:05:33insanely overpowered. There's also a lot
1:05:34of people that are saying that it just
1:05:36feels kind of finicky. So far, for me,
1:05:37it's been pretty great. Everything that
1:05:38I've asked it to do or find or schedule,
1:05:41whatever it is, it's been doing that
1:05:42pretty much perfectly. But, there are
1:05:44some people saying that it's asking them
1:05:45to reauthenticate multiple times. So, if
1:05:47that's a little frustrating, I guess
1:05:49just keep in mind that it will only get
1:05:50better. And we're not even to version
1:05:52one yet. So I would definitely recommend
1:05:53that you come to this GitHub, read about
1:05:54it, but more importantly get this thing
1:05:56installed in your cloud code setup and
1:05:58just start using it, using it, using it.
1:06:00Okay, so we've got the foundation of our
1:06:02AOS set up. We've pretty much linked in
1:06:05our context. We've pretty much linked in
1:06:07connections. At least we understand now
1:06:09that we've, you know, looked at all
1:06:10these things. We understand how we could
1:06:12go look at the documentation. We
1:06:14understand how we could grab our API
1:06:15keys and stuff. And we understand how to
1:06:17hook everything up. And now we are at
1:06:19this place where we want to keep
1:06:20leveling up and keep moving up the chain
1:06:22of the four C's. So next thing to think
1:06:24about is capabilities. Now when I say
1:06:27capabilities, what do I mean for the
1:06:28most part? I pretty much primarily mean
1:06:31building skills. Skills are those
1:06:33reusable recipes we talked about that
1:06:34help you do things more efficiently. And
1:06:37so everything that I do on some sort of
1:06:39cadence, you know, every time that I
1:06:40need to pull in my YouTube analytics,
1:06:43every time that I need to check in with
1:06:44the team, every time that I need to
1:06:45build a slide deck, all of those things
1:06:47I've built skills for in my AIOS. If I
1:06:50need to create images, if I need to
1:06:51create a LinkedIn post, all of these
1:06:53things are skills that I've built over
1:06:54time because I've challenged myself to
1:06:57do everything to do as much as possible
1:06:59inside of my cloud code. So building
1:07:01skills, building custom skills for your
1:07:03specific SOPs, your specific workflows,
1:07:05your specific tasks. Remember how we
1:07:07talked about this thing earlier? Every
1:07:08single task that you have can be broken
1:07:09up into chunks. And all of these chunks
1:07:12and this overall larger thing can all be
1:07:14turned into a skill. So, what I want you
1:07:16to do right now is I want you to grab a
1:07:18piece of paper and I want you to think
1:07:20about it's a Monday morning. You wake
1:07:22up, you sit down at your desk, you're
1:07:24ready to do work. What do you do? Write
1:07:26down what you do on a Monday. And then,
1:07:29unfortunately, pretend it's Tuesday. You
1:07:31go to your desk. Write down what you do
1:07:33on Tuesday. And as you do this
1:07:35throughout the week, as you do this at
1:07:36the end of each day, look at what you're
1:07:38doing. Circle things that are repeated.
1:07:41Circle things that you see pop up
1:07:43consistently. Circle things that you
1:07:44hate doing. And all of those things you
1:07:46circled are things that you should come
1:07:47into Cloud Code with and say, "Hey, I
1:07:49have this process. You know, every
1:07:51Monday I have to organize all of the
1:07:53different tickets that we got over the
1:07:54weekend, and I have to tag them and
1:07:56classify them, and then I have to route
1:07:58them to the right team, and then I have
1:07:59to follow up with them throughout the
1:08:00week. I hate doing that. Can you help me
1:08:03turn this into a skill? Sure. Let me ask
1:08:05you five or six questions to make sure I
1:08:07understand what you want and let's build
1:08:09a skill together. And that is exactly
1:08:10how you start adding more and more
1:08:11capabilities into your agent. So, in
1:08:15order to get a better understanding of
1:08:16skills, I'm going to go ahead and play a
1:08:18video where I break down a bunch of the
1:08:20ones that I use, how they work, and how
1:08:22you think about building them and
1:08:24improving them over time. It is super
1:08:25cool. It is the biggest unlock inside of
1:08:27an AIOS. So, I'll see you guys in that
1:08:30video. I have genuinely never been as
1:08:32productive as I am right now because of
1:08:34Claude's skills. This image really does
1:08:35sum it up. I feel like I'm working on
1:08:37multiple different computers and
1:08:38multiple different tasks at the same
1:08:39time without sacrificing quality. It
1:08:42just comes down to one simple word and
1:08:43that is leverage. With Claude skills or
1:08:45any agent skills for that matter, you
1:08:47have way more leverage than if you were
1:08:48doing this by yourself. So, in this
1:08:50video, I'm going to be breaking down
1:08:51what skills are, how they work, and how
1:08:53you can build really, really good ones,
1:08:54even if you've never heard of the
1:08:56concept or built a single skill ever
1:08:57before. So, let's hop into a live demo
1:08:59real quick and get going. All right, so
1:09:02here we are in Cloud Code in my Herk 2
1:09:04project, which is kind of just like my
1:09:05personal assistant. Now, if this stuff
1:09:07over here looks overwhelming, don't
1:09:08worry about that right now. Just worry
1:09:09on what I'm actually asking the agent to
1:09:11do. So, I've got the skill that I called
1:09:13morning coffee that helps me plan my day
1:09:14every morning. So, it's not morning
1:09:16right now, but I'm going to run this so
1:09:17you can see how it works. So, as Cloud
1:09:19Code's figuring that out, what I'm going
1:09:20to do is open up another agent. And in
1:09:23this one, I'm asking it to run a pulse
1:09:24check on all my projects and commitments
1:09:25to see how things are going. I'm opening
1:09:27up another one to create me an Excal
1:09:29diagram of the difference between local
1:09:30AI models and closed source models. And
1:09:33let's just do one more that's going to
1:09:34scrape the comments from my recent
1:09:35YouTube videos and give me an analysis
1:09:36on what I need to improve. So what
1:09:38you're looking at right now is four
1:09:39different agents running in parallel
1:09:41doing things for me. And that took me
1:09:42probably 30 seconds to ask them to do
1:09:44that. And because I built all of these
1:09:46skills, all of these agents have all of
1:09:47the context about my business, what's
1:09:49going on with our projects, my YouTube
1:09:51channel. It has literally everything it
1:09:52needs. And now all of those agents are
1:09:54done. Here was my February 26th morning
1:09:56coffee. I had three things on the
1:09:57calendar. So, what it's going to do is
1:09:59it's going to look at my ClickUp. It's
1:10:00going to see what else I've got this
1:10:01week and look at my tasks and then help
1:10:03me plan the rest of my day. So, this is
1:10:05the plan that it suggested. All I'd have
1:10:06to do is say, "Yep," and it would block
1:10:08off everything for me. And for me,
1:10:09that's huge because I don't have
1:10:10decision fatigue anymore of what I have
1:10:12to work on. The second agent came back
1:10:13with a pulse check. 2 days until the end
1:10:15of the month. Here's where everything
1:10:16stands. So, obviously I'm going to blur
1:10:18all of this out, but it's basically
1:10:19catching me up on all of the different
1:10:20main initiatives that we're doing this
1:10:22month and this quarter and making sure
1:10:23that everything is on track. And right
1:10:25here, you can see there's a few things
1:10:26that I need to follow up on manually.
1:10:28And this might have slipped through the
1:10:29cracks because I'm so busy making
1:10:30YouTube videos if I didn't have this
1:10:32personal assistant to check up on me
1:10:34using the skill. The third agent came
1:10:35back and has finished the Excal diagram
1:10:37and I pasted it in and it looks like
1:10:39this. So, if I needed to make a video
1:10:41about this, I wouldn't have had to take
1:10:42my own time to create this
1:10:43visualization. So, for the comment
1:10:45analysis, that came back and we can see
1:10:46all of the comments, all the views, and
1:10:48things that I need to address either in
1:10:50future videos or in the comments. We got
1:10:52some confusion. We've got some cost
1:10:53things that I need to cover. We have to
1:10:55stop demoing toy examples for tool
1:10:57videos. Seems like you guys really want
1:10:58to see some anti-gravity stuff. I
1:11:00promise that'll come soon. And then
1:11:01we've got top three priorities. So, I've
1:11:03been recording now this video for about
1:11:056 minutes. Just think about if I would
1:11:07have done all four of those things
1:11:08myself, how much context switching I
1:11:10would have done and how long that would
1:11:11have taken me. Okay, so now that you
1:11:12guys have seen an actual demo and
1:11:14hopefully you're a little bit excited to
1:11:15learn about cloud skills if you haven't
1:11:16used them yet, what actually are they?
1:11:19So skills are reusable instructions. You
1:11:21write them once, you save them as a
1:11:23skill. You can trigger them anytime and
1:11:25you're going to get way more consistent
1:11:26results because it's going through that
1:11:27same process every time. So this visual
1:11:29right here was actually an AI generated
1:11:31image that I used with a skill. It was
1:11:33this one right here in my cloud code
1:11:35called Excalaw visuals. But sometimes
1:11:37with AI generated images, they don't
1:11:39spell words right. As you can see here,
1:11:40this is all messed up. This stuff is
1:11:42messed up. So, I also have one, as you
1:11:44saw in the demo, to create Excal
1:11:45diagrams. And this one creates the
1:11:48actual Excal that I could move and edit.
1:11:50And the words are always perfect because
1:11:51it's actually just typing. And those two
1:11:53skills alone have saved me so much time.
1:11:54And what I'm going to do, by the way,
1:11:56for all you guys, is in my free school
1:11:57community, the link for this is down in
1:11:58the description, I've added a new
1:12:00classroom section called agent skills
1:12:01where I'm just going to be dropping a
1:12:03ton of these skills that you guys can go
1:12:04grab for completely free. So, before we
1:12:06dive into this today, just real quick,
1:12:08why should you care? And there's three
1:12:09big reasons. You can be way more
1:12:11productive as a person because you can
1:12:12automate things like you just saw me do.
1:12:14And you can legitimately build a
1:12:16personal assistant that can do almost
1:12:17anything for you. The second one is team
1:12:19leverage. So you can turn existing SOPs
1:12:21into automations really, really easy.
1:12:23And if you build something new, not just
1:12:25you can use it, but your entire
1:12:26organization can. So everyone as a group
1:12:29is getting way more productive, which
1:12:30will almost undoubtedly result in growth
1:12:32of the business. And also monetization.
1:12:35We're entering this new world where
1:12:36skills are having a big moment and
1:12:38you're able to capitalize on a lot of
1:12:39this stuff. Now, I'm not saying this is
1:12:41going to be a viable business model for
1:12:42a long, long time. So, you shouldn't
1:12:44bank on it, but it is something to be
1:12:45aware of just like when people were
1:12:47selling end workflow templates and
1:12:48things like that. But once again, it
1:12:50just comes down to one word, which is
1:12:51leverage. This isn't just theory. This
1:12:53is something that we're seeing with
1:12:54clients. This is something that we're
1:12:55seeing internally in my own business.
1:12:57This speed of work that we're able to
1:12:58achieve right now feels insane, but that
1:13:01is going to become normal. And if you
1:13:02can't do that, you instantly become way
1:13:04too slow and way too expensive for the
1:13:06business and they might not keep you
1:13:07around. We're actually making it a
1:13:09priority to make sure all of our
1:13:10employees are using cloud code. Because
1:13:12now I have all these different skills
1:13:13that I can just run with a simple slash
1:13:15command or a simple natural language
1:13:16prompt and get in one day a week's worth
1:13:19of output. Because once again, one
1:13:20person can figure out the best way to do
1:13:22something and turn it into a skill that
1:13:23the entire team can use. But they don't
1:13:25just generate text. They're basically
1:13:27automations. They can run scripts, they
1:13:29can call APIs, they can create things,
1:13:30they can have sub agents, and they can
1:13:32be called on from agents as well. So,
1:13:34this is truly AI automation. And just to
1:13:36really hammer it home, they're basically
1:13:37SOPs for your AI agents. The same way
1:13:40where you would train a human employee
1:13:41by letting them read through an SOP to
1:13:43understand the process, and then they'd
1:13:44be able to do it. You just train an
1:13:46agent on it. You give them the skill,
1:13:47they read it, and then they do it. And
1:13:49the coolest part about it is the more
1:13:50you use the skill, the better and better
1:13:52it gets. So, we've talked about a lot of
1:13:53these benefits, right? But what actually
1:13:55is a skill? Well, it's just a folder and
1:13:58it lives somewhere in your project. The
1:14:00most common example is probably going to
1:14:01be in your.claude/skills/skll
1:14:04name and then you've got like a skillmd
1:14:06or a markdown file. So right here in my
1:14:08herk 2 project, you can see up top we've
1:14:11gotclaude. I open up.claude, we have
1:14:13agents, rules, and skills. Right now,
1:14:15we're just talking about skills. If I
1:14:17open that up, we can see all of the
1:14:18different skills that I've created in.
1:14:19So let's say for example, the excalagra
1:14:21diagram skill. If I click into that, we
1:14:23have a skill.md. And when I open that
1:14:26up, you can see we have the name of the
1:14:27skill, the description, and then we have
1:14:29the actual workflow. Step one,
1:14:30understand the concept. Step two, plan
1:14:32the layout. Step three, generate
1:14:34elements. And this entire skill
1:14:35basically teaches my agent how to build
1:14:37these Excal diagrams for me. So, like I
1:14:39said, that is the anatomy of the skill.
1:14:41We've got front matter, which is kind of
1:14:43between these two dotted lines, and that
1:14:46is in something called YAML, which you
1:14:48don't need to worry about what that
1:14:49means. This is just the way that it's
1:14:51kind of indicated, sort of like your
1:14:53markdown or your JSON or Python. Now, up
1:14:56here, we'll have a name and a
1:14:57description, which tells Claude Code
1:14:59what the skill is called and what the
1:15:00skill does. So, as you can see, this one
1:15:02is called Excalraw- Diagram. And there's
1:15:04a brief description about what it
1:15:05actually does or when to use it. And
1:15:07then we have the step-by-step rules,
1:15:09which are basically the instructions.
1:15:10And this is what Claude actually does
1:15:13once it decides that this is the right
1:15:14skill for the job. Now the interesting
1:15:16thing about skills is that sometimes you
1:15:18need way more data. So for example,
1:15:20let's say we're writing a LinkedIn post.
1:15:22We have a skill for that, right? But
1:15:24what needs to go in the skill is other
1:15:26information sometimes like a company
1:15:28tone of voice or maybe your LinkedIn,
1:15:30you know, tone of voice, a target
1:15:31avatar, current priorities, a logo,
1:15:34maybe there are other things that you
1:15:35want to put into a skill besides just
1:15:37like the step-by-step instructions that
1:15:39will make it better. So the question is,
1:15:40where do these things go? Well, there's
1:15:42typically two options, but essentially,
1:15:44as long as you're pointing to the right
1:15:46path in the skill.md, you're fine. So,
1:15:49let me explain what I mean by that, and
1:15:50then I'll show you what I mean by that.
1:15:51So, first option is to have it
1:15:53self-contained. So, in
1:15:54yourcloud/skills/skillame,
1:15:58you can have the skill.md, you can have
1:16:00your scripts right there, and you can
1:16:01have your references right there. Or
1:16:03option B is that they're not directly
1:16:05nested right under that skill. So here
1:16:07we have do.claude/skill/infographic
1:16:10and the skillmd and we still have our
1:16:12scripts and our references in the same
1:16:13project but it's just not nested
1:16:15directly under that skill. And so I know
1:16:17that might have made no sense. So let me
1:16:19show you exactly what I mean by that.
1:16:20Okay. So here we have a skill called
1:16:22idea mining. And so what happens in here
1:16:24is basically used when someone asks for
1:16:26content ideas, video ideas, what to make
1:16:28next or to run idea mining. And so in
1:16:30here I gave it some context, right? My
1:16:32channel has this many subs. It's about
1:16:34AI automation. My content pillars are
1:16:36naden, rag agents, cloud code, voice AI.
1:16:39And then what I gave it is a bunch of
1:16:40references. I gave it channel data which
1:16:42is YouTube channel.md. I also gave it
1:16:45the raw data which is a JSON file. I
1:16:47gave it a competitor list for me. And I
1:16:49gave it an actual script to run analysis
1:16:52on my YouTube channel. And so in this
1:16:53case, what you can see is that I went
1:16:55for option B where I'm storing those
1:16:58reference files and those scripts, not
1:16:59directly nested in this skill. So
1:17:02basically what it could look like is
1:17:04within the skill itself, we could have a
1:17:06folder called, you know, references. We
1:17:09could also have a folder in here called
1:17:11scripts. And then within both of these
1:17:13subfolders, we could have more things
1:17:14like, you know, in the references, I
1:17:16could have channel data. And in the
1:17:18scripts, I could have
1:17:19YouTube-analysis.js
1:17:22or py, whatever it is. Basically, the
1:17:24idea is it doesn't matter where those
1:17:26actual reference files or scripts live
1:17:28as long as you point to the right spot
1:17:30in the MD file. So in my case, where
1:17:32these actually live is in a different
1:17:34folder. So here for the channel data, I
1:17:35would basically just go all the way down
1:17:37to references and then I could go down
1:17:39to right here, YouTube channel.md. So
1:17:41cloud code reads the skill and then it's
1:17:43able to find this if it needs it. Same
1:17:45thing for the scripts. It would go down
1:17:46here to scripts and then it would find
1:17:48analyze YouTube. py and it would just
1:17:51pull this in if it needed it. So
1:17:53hopefully you guys are with me. However
1:17:54you want to set it up works. And I think
1:17:56that's the most overwhelming thing about
1:17:57cloud code right now is that everyone
1:17:59uses different kind of folder
1:18:01architecture. But don't worry guys, I'm
1:18:02totally on top of this. I have a skill
1:18:04that I built out called skilluer. And
1:18:06this one I'll be giving away for free
1:18:08once again in my free school community
1:18:10right here. And all you'd have to do is
1:18:11load in the skill builder and then it'll
1:18:13help you build out everything you need.
1:18:15And I'll be showing a live demo of that
1:18:16in a few minutes here. So the skill.md
1:18:18is the actual brain itself and the
1:18:20supporting files are the tools that it
1:18:22can use. That doesn't mean every single
1:18:24time the skill is invoked that those
1:18:26reference files will all be called. And
1:18:27just in case you guys were wondering if
1:18:28you've watched some of my previous Cloud
1:18:30Code videos where we've used the WAT
1:18:32framework to build automations, this is
1:18:34very very very similar. In that
1:18:35framework, the W the workflows were the
1:18:37markdown file SOPs. That's basically the
1:18:40skill. The tools were the actual Python
1:18:42scripts and that's basically just the
1:18:44scripts that you might write or the
1:18:45references that you would add in. So if
1:18:46you've already been building some WAT
1:18:48stuff, you will pick up skills super
1:18:49super quickly. The cool thing about
1:18:51skills is that you don't have to build
1:18:52all of them. Obviously, as you're
1:18:53working with cloud code and you're
1:18:55finding that you're doing things
1:18:56repetitively, you can go ahead and build
1:18:57a skill for it. But there's an official
1:18:59library from Anthropic of Skills.
1:19:01There's a community of everyone that's
1:19:03open sourcing their skills and giving
1:19:04them out. And there's a marketplace
1:19:06where you can share and sell or you know
1:19:08download skills from people. And then
1:19:10you would take that skill or that
1:19:11essentially a prompt and you would add
1:19:13your own flavor to it. The one thing I
1:19:15would say is just be careful and make
1:19:16sure that no one's trying to, you know,
1:19:18give you a skill that has any malicious
1:19:19intent in there. And all these skills
1:19:21can work across different products. So
1:19:24cursor, anti-gravity, codeex, because
1:19:27it's so based in markdown, and it's
1:19:28essentially just a prompt, tons of
1:19:30different AI models can use them. Okay,
1:19:32so how does Claude know when to use a
1:19:33skill? Well, there are two ways to
1:19:35actually trigger them. The first one is
1:19:36you can be explicit, which basically
1:19:38means you can do a slash command and say
1:19:40the skill name, and it will just
1:19:42directly fire off that skill. Or it
1:19:44could just be natural language. So if I
1:19:46had a school post skill, I could say
1:19:47slashschool-post.
1:19:49Or I could natural language just say,
1:19:50"Hey, help me write a school post about
1:19:52X." Cloud would find that skill and then
1:19:54invoke it. So when you ask Cloud to do
1:19:56something, it will first read through
1:19:58the cloud. MMD file. It will analyze
1:20:00your request and it will search through
1:20:01the skills and see which one do I have
1:20:03that helps with this query. If it finds
1:20:05one, it will invoke it. But if it can't
1:20:07find anything, then it will basically
1:20:08just use its general knowledge. So not
1:20:10every single request that you give to
1:20:11cloud code will invoke a skill. Now a
1:20:14really important part of that is
1:20:15understanding how skills stay
1:20:16lightweight. Because if you've been
1:20:18using cloud code, you know that context
1:20:20management is a huge deal. And if you
1:20:22had all of these skills to look through
1:20:23and all of these skills are, I don't
1:20:25know, hundreds and hundreds of lines,
1:20:26then if cloud code was searching through
1:20:28all of these every single time, that
1:20:30would surely eat up a ton of your
1:20:31tokens. So what's used is something
1:20:33called progressive context loading,
1:20:35which basically means we have three
1:20:36levels. Level one is the initial search
1:20:39where cloud code only looks for the name
1:20:41and the description. So right here you
1:20:43can see let's say we ask for an Excal
1:20:45diagram. It would basically search
1:20:46through all the skills but it would only
1:20:48read the YAML front matter. So it would
1:20:50read the name and the description. And
1:20:51typically this front matter is only
1:20:53going to be you know maybe roughly 100
1:20:55tokens. So it stays very lightweight.
1:20:56And then moving down to level two let's
1:20:58say it identifies okay cool this is the
1:21:00right skill for the job. Then it would
1:21:03run the full skill.md and it would read
1:21:05through everything. And so that's when
1:21:06it would start to actually understand
1:21:08what goes on in the skill. And that
1:21:10might be anywhere from a thousand to a
1:21:12couple thousand tokens. And then level
1:21:14three is once again a decision. Only
1:21:16load in the extra files when needed. So
1:21:18if I need to look at any scripts or
1:21:20references or templates or I need to
1:21:21pull in some brand assets or more
1:21:23context, I'm only going to do that if
1:21:25the specific request requires it. And so
1:21:28hopefully now you're starting to
1:21:29understand a little bit more about under
1:21:31the hood what's actually going on when
1:21:32you ask Claude Code to do something for
1:21:34you. And you can always go to Cloud Code
1:21:36docs and go to the skills section and
1:21:38just read about how this stuff works.
1:21:40It's really, really simple. On the doc
1:21:42itself, it will tell you just make sure
1:21:43to keep the skill.md under 500 lines.
1:21:46Move detailed reference material to
1:21:48separate files. And so, I know this may
1:21:50seem like it's just a lot of information
1:21:51being thrown at you. So, let me just
1:21:53kind of contextualize this and slow it
1:21:54down and reassure you guys. You're never
1:21:56ever ever going to write a perfect skill
1:21:59the first try. The way that I build my
1:22:01skills is I have Claude Code do
1:22:02something with me. I walk it through the
1:22:04steps, you know, each time. And then
1:22:06when we're done, if we've went from
1:22:08point A to point B, I say, "Cool. This
1:22:10is something I do once a day. Let's turn
1:22:12this into a skill. Ask me more questions
1:22:14so we can make sure you have all the
1:22:15information you need." And once again,
1:22:17I'm going to show you guys opening up a
1:22:18brand new project and setting up a skill
1:22:20from scratch so you understand the full
1:22:22process, but I just had to give you guys
1:22:23some context first. Now, we have this
1:22:25thing called the feedback cycle, which
1:22:27basically means you invoke the skill,
1:22:28you actually watch the agent work, you
1:22:30give feedback, and then it fixes the
1:22:32skill, and then you do it again. And so,
1:22:34the first couple times you run a skill,
1:22:35you may feel like, eh, this feels very
1:22:37AI generated. But by the time you've run
1:22:39that skill 10, 20, 30 times, every
1:22:41single time it gets better. And so,
1:22:43that's why it's actually important to
1:22:44watch the agent work the first couple of
1:22:46times because that's how you're able to
1:22:47identify opportunities to speed it up
1:22:49and save tokens by doing things like
1:22:51this. So here's an example of the pulse
1:22:53check skill that we actually ran
1:22:54earlier. Now this skill gets invoked
1:22:56when I ask for a pulse check or checking
1:22:57in on commitments. And what it does is
1:22:59it reads through some context of how
1:23:01OTAAS work, which is important for it to
1:23:03understand every single time it reads
1:23:04the skill, which is why I put it here
1:23:06rather than a reference file. And what
1:23:07it has to do is it has to do a live
1:23:09lookup on my ClickUp to understand
1:23:11what's going on. So what I did is I
1:23:12hardcoded in these list ids because when
1:23:14I was watching it, I realized every
1:23:16single time it was doing this, it was
1:23:18calling the ClickUp MCP and it was
1:23:20gathering all these lists and it was
1:23:22searching and parsing the results and
1:23:24then it would extract the ID and that
1:23:25just was taking so long and it was
1:23:27costing me a ton of tokens. So I
1:23:29realized that's always going to be the
1:23:30same. Why don't I just give it in the
1:23:32skill document the list IDs and now it
1:23:35knows how to do that instantly every
1:23:37time and it doesn't waste all those
1:23:38tokens. And on top of that, I know that
1:23:40searching through ClickUp can consume a
1:23:42lot of time and tokens. So, I built a
1:23:43specialized sub agent that in this
1:23:45skill, I say, "Hey, delegate to the
1:23:47ClickUp searcher agent with this query
1:23:49in order to do all of this searching so
1:23:50that you don't blow your own context
1:23:52window." All of that's handled over
1:23:53there and then you only get the
1:23:55information that you need. So, there's a
1:23:56lot of advanced things that you can do
1:23:57to manage your context. I'm not going to
1:23:59dive into all of that right now. We're
1:24:00just focusing on skills, but just wanted
1:24:02to give you a little taste of what's
1:24:03possible in the skill.md files. So,
1:24:06another good example of needing a
1:24:07reference doc like that is in my skill
1:24:09builder skill. I obviously use this when
1:24:12I'm creating new skills, optimizing
1:24:13skills, auditing skill quality, things
1:24:15like that. And a lot of the inspiration
1:24:17I got from this was of course straight
1:24:19from Claude Code Docs itself about how
1:24:22to actually use and build and optimize
1:24:24skills. And so, when I was building this
1:24:25out, I I was watching the agent, you
1:24:27know, run the skill and I realized it's
1:24:29searching every single time. It's doing
1:24:31a web search and it's crawling the
1:24:32entire document, even if I just need a
1:24:34little piece of information. So, what I
1:24:36decided to do was I told it to basically
1:24:37scrape that whole thing. And then I gave
1:24:39it a reference.md, which is basically
1:24:42the documentation. So, I've got my
1:24:43skill.md. And what it does is it
1:24:45references that full file if it needs
1:24:47it. But really, the main idea that I'm
1:24:48trying to drive home here is that
1:24:50processing markdown files for your agent
1:24:52is so much quicker and cheaper than
1:24:55actually making API calls or HTTP
1:24:57requests, you know, executing functions
1:24:59and reading tons and tons of tokens. So,
1:25:01the goal is your skills will get to a
1:25:03place where you can invoke them, focus
1:25:05on something else for 10, 15 minutes or
1:25:07whatever, and then come back and have a
1:25:08finished result that is really, really
1:25:09good. But the first couple times that
1:25:11you are testing out a skill, I think
1:25:13it's a really good idea to just sit
1:25:14there and watch it and see what it's
1:25:16doing. And a lot of people have asked me
1:25:17like when do you know when to build a
1:25:19skill? Well, basically just go about
1:25:21your work and if you ever realize that
1:25:23you've done something already or you've
1:25:24instructed something differently like I
1:25:27tell my claude to not use m dashes.
1:25:29Okay. Well, that's probably a good idea
1:25:30to put that in the prompt, right? So, if
1:25:32you ever find yourself doing a process
1:25:34or repeating prompts, then that's
1:25:35probably a good use case to build a
1:25:37skill around it because skills don't
1:25:39have to be complex. They could literally
1:25:40just be a 50line markdown file. All
1:25:43right, so we're about to hop into a live
1:25:45build of a skill from scratch. But what
1:25:47I wanted to do real quick was go over
1:25:48the six-step skill building framework.
1:25:50So number one is the name and the
1:25:52trigger. What is it called and the
1:25:54natural language that would basically
1:25:56fire it off? Number two is the goal. So
1:25:58in one sentence, what will this skill
1:26:00accomplish by the end? What will be the
1:26:01output? Number three is the actual meat
1:26:03of it. That's the step-by-step process.
1:26:06If you had to do something manually,
1:26:07exactly what do you do in what order?
1:26:09What do you look at? And what decisions
1:26:11do you make? Number four is the
1:26:12reference files. What context do you
1:26:14need? Do you need images? Do you need
1:26:15understanding of current projects,
1:26:17current priorities? Do you need style
1:26:19guides? What do you need to do the job
1:26:21well? Number five is the rules. Think
1:26:23about what could go wrong and then the
1:26:24agent can help you building guard rails
1:26:26and constraints around that. And then
1:26:27number six is kind of like after you've
1:26:29built it, it's just the self-improvement
1:26:30loop. And after the live build, I'm
1:26:32going to talk about actually testing and
1:26:34iterating and what you need to do to
1:26:36make them really, really good. But for
1:26:37now, that's the six-step skill building
1:26:39framework. Let's hop into a live build.
1:26:41Okay, so here we are in Visual Studio
1:26:42Code, which is where I like to use Cloud
1:26:44Code. If you don't have Visual Studio
1:26:46Code, just go ahead to a browser, type
1:26:47in VS Code, and then go ahead and
1:26:49download this. This is what it will look
1:26:50like. If it's your first time using
1:26:52Cloud Code in here, you just have to go
1:26:53to extensions on this lefth hand side,
1:26:55type in Cloud Code, and then install
1:26:57this, and then log in with your paid
1:26:59Anthropic subscription. Now, after that,
1:27:01you're going to click on this top left
1:27:02button, and it's going to pull up this
1:27:04little thing that says you have not yet
1:27:05opened a folder. What you need to do is
1:27:07open up a project to work in. So you
1:27:09could either open up one that you're
1:27:10already working on or you could go ahead
1:27:11and create a new folder and then open
1:27:13that one up. For the sake of the demo, I
1:27:14just opened up a new blank folder called
1:27:16a bunch of skills. And I'm going to show
1:27:18you exactly what to do. So the first
1:27:19step is to go to my free school
1:27:20community link in the description. Go to
1:27:22the agent skills classroom and download
1:27:24the skill builder folder. Once you've
1:27:26got those files ready to go, first thing
1:27:28we want to do is just set up this
1:27:30workspace real quick. Initialize this
1:27:32project with a simplecloud/skills
1:27:35structure. Cool. So, as you can see,
1:27:38that got set up. We have aclaude. We
1:27:40have a skills folder. And what I'm going
1:27:41to do is in this skill folder, I'm going
1:27:43to create a new folder called
1:27:45skill-builder
1:27:47and hit enter. And then I'm going to
1:27:49take those two files for my school
1:27:50community, the reference and the
1:27:52markdown, and I'm going to put that
1:27:54right in here. So now we have this skill
1:27:56builder set up with the reference file
1:27:58and the actual skill markdown. I'm
1:28:00asking it if it can see that new skill
1:28:01that I just added. It says, "Yes, I can
1:28:03see it." And I'm basically just going to
1:28:04say, "Cool. Let's run that skill to
1:28:06build a new one together. So now you can
1:28:07see what it did is it basically is
1:28:09reading the skill right now. This is the
1:28:11instructions that we saw right in here.
1:28:13As you guys know, since that's how
1:28:15skills work, it starts to read this. So
1:28:17here we go. I built this skill to
1:28:18actually ask you questions so that it's
1:28:20way easier for you to communicate what
1:28:22you want. So the first thing is what
1:28:23problem are you trying to solve? What we
1:28:25want to do is content creation because
1:28:26in this skill, what I want to do is
1:28:28building branded infographics. What kind
1:28:30of content does the skill create? What's
1:28:32the specific use case or workflow? And
1:28:33I'm actually just going to choose other
1:28:34for this. And I'm going to say
1:28:36educational infographics. Now it's
1:28:38asking how we should trigger this skill.
1:28:39So does it want to be natural language
1:28:41or do we want to just use slash
1:28:42commands? And I'm just going to say both
1:28:44is fine. And now we're moving on to the
1:28:46step-by-step process, which is really
1:28:47important because at this point we
1:28:49haven't told it what text stack we
1:28:50actually want to use or anything else
1:28:52about our business. So walk me through
1:28:54what should happen from trigger to
1:28:55output. And it has some good guesses,
1:28:56but what I'm going to do is do other and
1:28:58explain this the way that I want it
1:29:00built. I will tell you what I want an
1:29:02infographic about. You will create a
1:29:05concept. You will make a request to
1:29:08key.ai to use nanobanana to generate the
1:29:10outline or sorry to generate the image.
1:29:12And you will also look at the brand
1:29:14guidelines that I give you so that
1:29:16everything that is created follows my
1:29:18brand colors and typography and stuff
1:29:21like that. The output format that I
1:29:22actually want is a PNG, not any of this
1:29:24stuff. Does this need to be
1:29:26conversational or fire and forget? I'm
1:29:28just going to go fire and forget. All
1:29:29right. All right. So, how does the key
1:29:30AI nanobanana integration work? Is it an
1:29:32API call? Yep. We're just going to go
1:29:34with an API call. And in these options,
1:29:36you could literally say, I don't know.
1:29:37Let's try different things. You know,
1:29:38help me figure out what's best. It's
1:29:40asking where those brand assets live, so
1:29:41I'll put them in a folder. And where
1:29:43should the generated PNG infographic be
1:29:45saved? Yeah, sure. Let's start a new
1:29:46folder called projects, and we'll throw
1:29:48all of them there. So, it's going to
1:29:49keep our project organized as well. So,
1:29:52now it's asking about brand guidelines.
1:29:53I created this folder, and I put in our
1:29:55kind of color scheme as well as the
1:29:57actual AS logo. I have put in both our
1:30:00AIS brand guidelines and the AIS logo. I
1:30:03want to make sure that in the top left
1:30:05corner of every single infographic
1:30:06that's created, the AIS logo appears
1:30:08exactly as I've given you. But I think
1:30:10you guys get the point here. I'm going
1:30:11to answer a few more questions and then
1:30:13I'll just show you when we have a
1:30:14result. And now that we've done that,
1:30:16what you can see is it is going to
1:30:18create the skill. It's going to create
1:30:19the logo overlay. It's going to create a
1:30:21supporting reference markdown file for
1:30:23all of the API details that it's going
1:30:24to need. So that's great. It's going to
1:30:26register the skill in claw.md and it's
1:30:29going to log its decisions. All right,
1:30:31so it fully built the skill. It created
1:30:32all those files for us. We just have to
1:30:34give it a key API key so it can actually
1:30:37run this. Okay, so I threw in my API key
1:30:39and then I said test it out with an
1:30:41infographic about cloud skills. That's
1:30:42it. No other context. It invoked the
1:30:45skill right here and we will see what
1:30:46happens. Okay, this is really
1:30:48interesting. So what it's doing is it is
1:30:50generating the image and then it's just
1:30:52going to overlay the logo. So, it's
1:30:53going to be a lot more consistent than
1:30:54giving the AI image generator Nano
1:30:57Banana my logo. So, I didn't even tell
1:30:59it to do that. Let's see how it looks.
1:31:01Okay. Well, I don't love this. We're
1:31:03just going to go back and ask it to
1:31:04change some things. The logo on the top
1:31:06left doesn't look great. I gave you a
1:31:08logo with a transparent background. So,
1:31:10it should just be overlaid on top and we
1:31:12should be able to see the background
1:31:13behind it. The actual infographic itself
1:31:15is all right, but I actually want these
1:31:17to always be one by one aspect ratio.
1:31:21Okay, so I made some suggestions and
1:31:22it's going to try again and it's going
1:31:24to update its skill. So we'll see if
1:31:25that's better. All right, so second time
1:31:27we run the skill. Let's see if it's any
1:31:28better. All right, there we go. We've
1:31:29got the logo up top. We've got cloud
1:31:31code skills, custom AI workflow, command
1:31:33prompt, trigger, front matter, config
1:31:36triggers, AI agent delegation, document
1:31:38output. So just keep in mind all we said
1:31:40was build an infographic about cloud
1:31:42skills. And this was run number two.
1:31:44Every single time that we do this cycle,
1:31:46remember we talked about the feedback.
1:31:47We would basically watch it again, give
1:31:49more feedback, and then keep going. And
1:31:51after we run this probably five or six
1:31:53more times, this would be really, really
1:31:54good. And then every time I ask for an
1:31:56infographic, it's going to be
1:31:57consistent. And just to show you guys
1:31:58what was actually built, if we open up
1:32:00the infographic builder skill, we have
1:32:02the actual skill itself. So we have the
1:32:05front matter right here, the name, the
1:32:06description, we've got what the skill
1:32:08does, we've got context, so here's where
1:32:10it links to the actual brand guidelines
1:32:12and logos. We've got the step-by-step
1:32:14workflow right here. And we can see
1:32:15right here for full API reference and
1:32:17parameters just see the markdown file so
1:32:19that you don't have to actually go
1:32:21search the web and search through a
1:32:22bunch of tokens. You can just read this
1:32:24markdown file. All right. So we've
1:32:25talked about a lot of stuff about skills
1:32:26today and we just built one live. So
1:32:29what I want to talk about now is really
1:32:30how do you bridge the gap from like a
1:32:3290% good skill to making it pretty much
1:32:34100%. So testing, iterating, and
1:32:37debugging. There's different symptoms
1:32:38and there's different fixes. So let's
1:32:40just kind of go down this list one by
1:32:41one. The first symptom might be it does
1:32:43the wrong steps or in the wrong order.
1:32:45Well, you would just tell it to edit the
1:32:46skill.md instructions. You could get
1:32:49missing tone, style, or context. In that
1:32:51case, you're going to add reference
1:32:52files. And of course, those have to be
1:32:54pointed to correctly in the skill.md.
1:32:56You could get the same mistake happening
1:32:58over and over, then you're going to add
1:32:59a rule. If it struggles with a tool or
1:33:01an MCP or it keeps searching for the
1:33:03same things, then create some sort of
1:33:04reference dock for it. If it works good,
1:33:06but it could get better, then that just
1:33:08means you have to brute force it. you
1:33:10have to just run it over and over and
1:33:11over and keep nitpicking at what it does
1:33:13wrong or maybe not wrong but what it
1:33:15could improve on. If the skill isn't
1:33:16triggering, then check the YAML and make
1:33:18sure it is specific enough. If the skill
1:33:20triggers too often, then maybe try
1:33:21disabling model invocation and that is
1:33:24something that you can see in the claw
1:33:25docs, which basically gives you control
1:33:26over if the skill can only be invoked by
1:33:28natural language or only be invoked by
1:33:30the slash command directly or both. So,
1:33:32like I said, if you want to look at some
1:33:33more advanced stuff, then definitely
1:33:34head over here to the actual doc. But at
1:33:37this point, we've covered almost
1:33:38everything about these skills. One thing
1:33:39that I would call your attention to is
1:33:41the actual front matter reference
1:33:42because we saw the name and the
1:33:44description, which is what's required
1:33:45every time. But there's lots of other
1:33:47things that you can add in there. Here
1:33:48is the disable model invocation like we
1:33:50just saw. But you can also give it
1:33:52allowed tools. You can also give it an
1:33:54argument hint. You can give it a
1:33:56specific model to use. You can give it
1:33:58specific context. You can give it hooks.
1:33:59You can give it a specific agent. And so
1:34:01all of this lets you get really, really
1:34:03granular on the exact skill and how you
1:34:05want it to be used. But don't get
1:34:06overwhelmed. You really only get to that
1:34:08point once you've ran the skill a ton of
1:34:09times. Now, another thing that I need to
1:34:11hit on real quick is where do skills
1:34:13actually live because what we've seen so
1:34:15far is just building them right in
1:34:16ourcloud/skills folder. But when you're
1:34:18doing this, they only exist in that
1:34:21specific project. So whether that's my
1:34:23her two or my, you know, the one we just
1:34:24spun up, if I went to a different
1:34:26folder, that skill would no longer be
1:34:28able to be accessed by our cloud code.
1:34:30But you can also create skills that are
1:34:32actually global. And you do that by
1:34:34doing that in a different directory in
1:34:36your kind of overall home directory. And
1:34:38that's basically indicated by the little
1:34:39tilda right here. And so that means
1:34:41every project you use in cloud code, no
1:34:42matter where you are, that skill would
1:34:44exist. So for example, I have a
1:34:46front-end design skill that is installed
1:34:49globally. So that whenever I'm anywhere,
1:34:51if I need to do front-end design, it
1:34:52just is able to use it. And just in case
1:34:54you want to look at it in a different
1:34:55way, right now what we're doing is we
1:34:57have our project, right? So, herk 2 and
1:35:00then we have dotcloud and then within
1:35:02dotcloud we have skills and then your
1:35:04skill and then your MD, your references,
1:35:06whatever and then maybe another skill.
1:35:08But if it was global, you might not
1:35:09actually see it in your project. It
1:35:11would just be within your overall home
1:35:13directory. So, the reason why you might
1:35:15want to do this is if there's something
1:35:16very specific about you, your business,
1:35:18your workflows that you want applied to
1:35:20every single project, no matter what,
1:35:21maybe your company context, your company
1:35:23projects, your tone of voice, whatever,
1:35:25then you can install that globally
1:35:26instead. All right, great. Right. So,
1:35:28now that you guys understand skills a
1:35:29little bit better and you're excited to
1:35:31start building some, the last piece of
1:35:33the four C's is cadence. And what does
1:35:35cadence mean? Cadence means a couple
1:35:37things in my mind. It means that because
1:35:39you can now turn off your laptop and
1:35:41things will still run, you have to
1:35:43figure out the different triggers. And
1:35:45you have to figure out what are things
1:35:47that I should actually set up on some
1:35:48sort of cadence. Whether that means you
1:35:50want to have a skill running in the
1:35:51background or you want to have an actual
1:35:53deterministic automation running in the
1:35:54background or it means you want to have
1:35:56some sort of dashboard with live data
1:35:58across all of your different connections
1:36:00and you want to have one clean place to
1:36:01look at all of that. So let me address a
1:36:04lot of those things that I just
1:36:05mentioned and how you can start to work
1:36:06in a cadence around that stuff. So one
1:36:09of the very first things that you can do
1:36:11is you have this AIOS project, right?
1:36:14You should probably be putting this onto
1:36:16a GitHub repo. You can keep it
1:36:18completely private. That's 100% fine.
1:36:20But you should probably put that on
1:36:21GitHub. Now, if you're asking yourself,
1:36:22what is GitHub? Let me just explain that
1:36:24to you real quick. So, basically,
1:36:26everything that you build in Cloud Code
1:36:28is a bunch of folders and files, right?
1:36:30So, in this example, we've got the AIOS
1:36:32with our, you know, our readme, our
1:36:34claude, we've got our references, we've
1:36:36got ourcloud with our skills. This is
1:36:38basically what your project is at the
1:36:41end of the day. And this is all being
1:36:43stored locally. Meaning if you have one
1:36:45laptop that you're working on this with
1:36:46and then you open up your second laptop,
1:36:48they will not sync. You will only have
1:36:50that on one laptop. So GitHub is
1:36:53basically a place for you to upload your
1:36:57folders and files and basically your
1:36:59codebase. So if you put this into a
1:37:01GitHub repo, short for repository, then
1:37:04you could basically just pull that repo
1:37:06in. So if you guys remember at the
1:37:07beginning of this video when you cloned
1:37:09my repo, that's because I uploaded it
1:37:11there so anyone could go clone it. So
1:37:13you can upload one privately, meaning no
1:37:15one else can touch it, but you can
1:37:17because you have login access, right? So
1:37:19now if you open up your other laptop,
1:37:21you can pull in that GitHub repo and
1:37:23just keep working. So you have your AIOS
1:37:25no matter wherever you are. But the
1:37:27second thing you can do on top of that
1:37:29is you can actually use that GitHub repo
1:37:32with Cloud Code on the cloud or an
1:37:35OpenClaw agent or a Hermes agent or
1:37:37whatever of these other agent frameworks
1:37:39that you want to use, a codeex even. you
1:37:41could plug that into any different AI
1:37:43harness. So, if you wanted to be able to
1:37:45use your AIOS whenever you're, you know,
1:37:48out of town or if you are on the road
1:37:50and you don't have your laptop with you,
1:37:52you could connect it to like ClickUp,
1:37:54you could connect it to Telegram because
1:37:56you could put, you know, cloud code on a
1:37:58VPS or you could have it running on a
1:38:00Mac Mini or something. There's different
1:38:02ways that you can have access to your
1:38:03AIOS 247. You know, you could use, like
1:38:06I said, a Hermes agent or an OpenClaw
1:38:07agent as well. But that is one thing
1:38:09that you can do so you always have your
1:38:10AIOS with you no matter where you are.
1:38:13Now, what else can you do? You can
1:38:15actually start to schedule these tasks.
1:38:17You can start to schedule all of your
1:38:18capabilities. So, all of your skills,
1:38:20you can have them run every hour. You
1:38:22can have them run every Monday at 6 a.m.
1:38:25You can even have them triggered on some
1:38:26sort of, you know, web hook. So, those
1:38:28are the things that you can be looking
1:38:29at next. Now you guys remember earlier I
1:38:31mentioned that you could also use the
1:38:33desktop app of claude because in claude
1:38:35you have code, you have co-work and you
1:38:37also have claude chat and you can use
1:38:40claude code right here in the desktop
1:38:42app and you can run your AIOS just fine.
1:38:44I just prefer VS Code. But what's cool
1:38:47in here is that you have routines and
1:38:49when you have routines they can either
1:38:50be local or they can be remote. The
1:38:53difference here is that a local routine
1:38:54has to run with the desktop app open.
1:38:57But a cloud routine can run with your
1:39:00desktop app closed and with your
1:39:01computer off. As you can see, local
1:39:03routines only run while your computer's
1:39:04awake. But on my plan, which is 200
1:39:07bucks a month max plan, I can only have
1:39:0915 remote cloud routines running per
1:39:12day. And you can even open up the
1:39:14calendar and see, okay, today's
1:39:15Thursday. Here are all of the routines
1:39:17that I have running on a Thursday. On
1:39:19Friday, here are all the routines that I
1:39:20have running on a Friday. And this is
1:39:22where you really start to set up some
1:39:24nice cadence of things that actually
1:39:25happen while you sleep, which is just
1:39:27absolutely awesome. And what's cool
1:39:28about these is they actually run skills.
1:39:31So for example, this is my school wins
1:39:33engagement scheduled task or routine.
1:39:36What happens is the first instruction
1:39:38here is run the wins engagement skill.
1:39:41So a routine basically just injects this
1:39:43into a real claude chat. It doesn't run
1:39:46some sort of, you know, Python script on
1:39:48the cloud. it runs a real clawed session
1:39:50the same way that if I copied this exact
1:39:52prompt right here and I dropped this
1:39:53into my AIOS it would do the same thing.
1:39:56So let me just real quick show you today
1:39:58at 10 a.m. when this ran it did it like
1:40:00this. It dropped in this exact prompt.
1:40:02It ran the skill called wins engagement
1:40:05and then you can see down here cloud
1:40:07code basically ran the skill found all
1:40:09the results and then it executed
1:40:11everything for me in my school
1:40:12community. So that is kind of the highle
1:40:14overall really cool aspect of routines.
1:40:17I'm going to go ahead and play a clip
1:40:18right here just to dive into the cloud
1:40:20routines a little bit more because I
1:40:22think that those are very important to
1:40:23understand. And there's a few gotchas
1:40:24where if you don't watch this video, you
1:40:26might end up getting stuck. So, I'm
1:40:27going to play this video real quick and
1:40:29then you guys will be right back here
1:40:30with me. Cloud Code has finally brought
1:40:32us routines, which basically means you
1:40:33can inject a prompt into Cloud Code, but
1:40:35it can be running on the web, so your
1:40:37laptop does not have to stay open. And
1:40:39I'm so excited about it. I've already
1:40:40been playing around with it. I've been
1:40:41migrating my automations over there, but
1:40:43there are a lot of little gotchas. So,
1:40:44I'm here to explain exactly how you can
1:40:46actually set up these automations so
1:40:47that they work. So, today, April 14th,
1:40:49Claude tweeted, "Now, in research
1:40:51preview, routines in Cloud Code, you
1:40:53configure a routine once, which is
1:40:54basically like a prompt, and it can run
1:40:56on a schedule from an API call or in
1:40:58response to an event, and it runs on
1:41:00Anthropics web infrastructure. So,
1:41:02that's awesome. So, you can call a
1:41:04routine from an API, you can have GitHub
1:41:05events trigger it, or they can be
1:41:07scheduled, which are like the scheduled
1:41:08automations that we already have, but
1:41:10now they run on the web. So, you really
1:41:12can create these from anywhere. You can
1:41:13do it right here as a scheduled trigger
1:41:15to run scheduled remote agents which is
1:41:17in the terminal. You could also go to
1:41:18clawai/code. So you could do it on the
1:41:20web. And right here you see I have three
1:41:22web- based routines right here. Or what
1:41:24I'm going to be showing you guys today
1:41:25is just doing it in the desktop app.
1:41:27Because right here, if I go to my
1:41:28scheduled tasks, you can see that I've
1:41:30got some like these four that are local.
1:41:32And then I've got these four that are
1:41:33running inside of a GitHub repository.
1:41:35So these are the remote ones. If I go up
1:41:37here and click on new task, this is
1:41:39where we could set up a new local task
1:41:40or a new remote task. It's very similar.
1:41:43You set up the name, you set up what
1:41:45Claude should do, and this is the actual
1:41:46prompt. So, I'll talk more about that in
1:41:48a sec. But then you would configure your
1:41:50model, your repository, and your cloud
1:41:52environment. You set the cadence,
1:41:54hourly, daily, weekdays. I think the
1:41:56minimum is once an hour, like you
1:41:58couldn't go like every 10 minutes or
1:41:59something, but still not bad at all.
1:42:01This is where you could configure all of
1:42:03your connectors. So, if you need to
1:42:04connect Slack or Gmail or, you know,
1:42:06whatever it is, you can connect them
1:42:07right here. But you can also just do
1:42:09your regular API endpoints with your API
1:42:11keys. And then, of course, you've got
1:42:13your permissions. So, you can choose how
1:42:15Claude should be acting. Now, the one
1:42:17thing about these are these are meant to
1:42:19be a oneshot prompt. You're not around.
1:42:21So, you probably want to make sure that
1:42:22it doesn't ever have to stop and ask you
1:42:24questions. Otherwise, what's the point
1:42:25of the automation? So, like I said,
1:42:27there's tons of things to dive into
1:42:28here, and I'm not going to try to bore
1:42:30you guys, but some of this is really
1:42:31important because when I first got this
1:42:32set up, my automations weren't just
1:42:35migrating over and working. So, I'm
1:42:36going to tell you guys the issues that I
1:42:37ran into, and hopefully answer
1:42:39everything that you need to know so that
1:42:41you won't have to go into the comments
1:42:42and ask these common questions. I can
1:42:44just answer them right here for you. So,
1:42:45let me just first of all, real quick,
1:42:46show you guys what I tested out. The
1:42:48first thing I wanted to test out is if I
1:42:50came in here and I created a new routine
1:42:53for just shooting a message to my
1:42:54ClickUp. Obviously, that's not any
1:42:56value, but I just wanted to see how it
1:42:58worked because what I wanted to do is
1:42:59see if I could do this without adding my
1:43:01connector of ClickUp. And I was able to
1:43:03actually get this to fire off, but it
1:43:05didn't work right away. So, let me show
1:43:06you guys what I ran into. So, the way
1:43:08that this works is you need a GitHub
1:43:09repository to sync it to in order for
1:43:12this to actually run. So, it's going to
1:43:13clone my Herku project right here in the
1:43:15web. It's going to be able to read my
1:43:16cloud.mmd. It's going to be able to read
1:43:18my scripts and my skills. And then after
1:43:20it finishes the job, it basically just
1:43:22destroys that little cloud GitHub clone.
1:43:25But as you guys know, you don't push
1:43:26your secrets into GitHub because if you
1:43:28see here my my Herk 2 project, this is
1:43:31myv file with all of my API keys. And
1:43:34this is listed in the git ignore, which
1:43:35basically says, hey, when you push to
1:43:37GitHub, you don't include these files.
1:43:39So what that means is in here, if this
1:43:41is only looking at your GitHub repo,
1:43:43there's nov. So how do you get your API
1:43:45keys into this routine that runs on the
1:43:49web? Well, what you do is inside of this
1:43:51scheduled task, you have a cloud
1:43:54environment. So if I click on this one,
1:43:55you can see this one is called Nate
1:43:56Hertz cloud. So if I open up the
1:43:58settings, what do you see? You have the
1:44:00name of this cloud environment. You have
1:44:01the network access and you have
1:44:03environment variables. So right here is
1:44:05where I put in my YouTube API key, my
1:44:07ClickUp API key, any of the other API
1:44:09keys that I need to give this cloud
1:44:11environment access to. And then the
1:44:12other thing you have to do is you have
1:44:14to look at the access levels because
1:44:15right here you can see that this one is
1:44:17on full but by default this will be on
1:44:20trusted I believe and that means you can
1:44:22only download packages from verified
1:44:24sources from Anthropic. And when we talk
1:44:26about this later I'll have a link which
1:44:27you can go see all of them. You could
1:44:29even do custom if you wanted to allow
1:44:30specific domains that aren't on that
1:44:32list. But in order for ClickUp to work
1:44:33in this case I had to go on full because
1:44:35when I went on trusted it said hey we
1:44:37can't actually do that. But when I
1:44:38change this to full, it let me send a
1:44:40message to my ClickUp. And that is how I
1:44:41got this message right here that says,
1:44:43"Just testing that the remote tasks work
1:44:44and the credentials work." So basically,
1:44:46when these run, whatever you have here
1:44:48as your instructions is what gets
1:44:50prompted. And that's exactly the same
1:44:52way that the scheduled tasks locally
1:44:53work. So right here, you can see I say
1:44:55send a message in the internal ClickUp
1:44:56channel. And right here, the actual
1:44:58thing that it says was send a message in
1:45:00the internal ClickUp channel. So, think
1:45:02of a scheduled task or a routine as you
1:45:04basically typing in a prompt and then
1:45:07someone coming in to your laptop and
1:45:09typing it in for you. So, it's the exact
1:45:11same type of interaction as you talking
1:45:13to Cloud Code. But that's why once
1:45:15again, you want to make sure it's
1:45:16specific enough so that it can basically
1:45:18oneshot it. Okay. So, let's take it a
1:45:20little deeper. Now, what I tried to do
1:45:21is I did another one which I wanted it
1:45:23to be able to use the YouTube data API
1:45:26in order to grab some YouTube comments
1:45:28for me and give me a little analysis in,
1:45:30you know, ClickUp or whatever. So, this
1:45:32is the prompt I said, right? Analyze 50
1:45:34of my most recent comments from YouTube
1:45:35and give me a quick bullet rundown. My
1:45:37YouTube API key is available as an
1:45:39environment variable. Use it directly
1:45:41from the environment. Don't look for av
1:45:44because what happens is in your repo,
1:45:46right? So in this her 2 project um when
1:45:49I normally run this it grabs all my API
1:45:51keys from the env.mmd
1:45:54and realizes that's where a lot of those
1:45:56live. So by default it's maybe going to
1:45:58try to look in thev and it's not going
1:46:00to be smart enough to figure out and so
1:46:02for clickup it was fine. It figured it
1:46:03out but for some reason with this
1:46:05YouTube one it didn't. So I had to
1:46:06explicitly tell it hey look in the
1:46:08environment variable rather than in
1:46:10thev. So you can see this first time I
1:46:12ran it 1241 I didn't say that and it
1:46:15couldn't do it. It said like, "Hey, I
1:46:17can't find that. I'm getting an error."
1:46:18And I even tried to tell it here, and it
1:46:20still didn't work. But then on this most
1:46:21recent run, when I updated the prompt a
1:46:23little bit, it was able to fetch it
1:46:25right away using the API key. And now I
1:46:28have a remote, you know, routine that
1:46:30would work. Obviously, I need to update
1:46:32this. I'm going to migrate over my other
1:46:33automations, but this was just for
1:46:35testing purposes. And then another one
1:46:37that I do is I have some automations
1:46:39here which basically opens up a browser
1:46:41using playright CLI and it does some
1:46:44stuff in my school community because
1:46:45there's no publicly accessible API.
1:46:47We've kind of figured out a way to
1:46:48automate it without using browser. I'm
1:46:50not really going to dive into that right
1:46:52now. But what I wanted to tell you guys
1:46:53about that is I tried to basically move
1:46:55over this school wins engagement post or
1:46:58sorry automation into a remote session.
1:47:01So, I copied the exact same prompt that
1:47:03was in my regular scheduled task and
1:47:05then I just added this little snippet at
1:47:06the end. But what happened is this
1:47:08wasn't working because it basically
1:47:10said, hey, you know, like when you do
1:47:11this, it spins up a browser, but there's
1:47:13no cookies because all of this is
1:47:14running remotely and all I have to look
1:47:16at is the GitHub repo. I can't look at
1:47:18the local, you know, cookies that we've
1:47:19used in the last couple sessions of this
1:47:21automation. And so, it doesn't seem like
1:47:23this would work because once again, it
1:47:25has no access to that stuff. So if I
1:47:27wanted to do an automation like this, I
1:47:29would have to use um an endpoint that
1:47:31takes authentication in the form of like
1:47:33actual cookies or a header or you know
1:47:36like an API key because every single one
1:47:38of these runs is going to be stateless
1:47:40and after the run the GitHub clone just
1:47:42gets deleted. Now the exception of that
1:47:44is if the automation is changing
1:47:45something in your codebase or doing a
1:47:47review. If it does do that it will
1:47:48create a new branch for you or it will
1:47:50give you some sort of output and not
1:47:52just delete everything that it just did.
1:47:53But for an automation like this it would
1:47:55just delete it. But hopefully after you
1:47:57guys have seen those examples, you now
1:47:59have the ability to come in and you know
1:48:01make some changes if you need in order
1:48:03to make sure that your automations are
1:48:05running. And what I mean by that is you
1:48:07understand this should be a very
1:48:08specific prompt. This is how you change
1:48:10the model. You have to have a GitHub
1:48:12repo. You can change the settings for
1:48:13your cloud environments right here. You
1:48:16set the schedule. You add any connectors
1:48:17you might need, which would honestly be
1:48:19a little easier if you added just like a
1:48:20Slack connector. And then you can set
1:48:22your permissions here. Now, the other
1:48:24thing to be aware of is you do have
1:48:26limits. So, if I come over here to my
1:48:28settings, you can see if I go to my
1:48:29usage, we have our regular session
1:48:31limits, our model limits, but for
1:48:32additional features, we have daily
1:48:34included routine runs, and I haven't run
1:48:36any yet on the actual schedule. I've
1:48:38just been testing them. Um, but we are
1:48:40at zero for 15. So, I could only have 15
1:48:43automations running with routines per
1:48:46day because I'm on the max $200 a month
1:48:48plan. Your limits would be less if
1:48:49you're on Pro. I think maybe three or
1:48:51maybe five. I'll I have that information
1:48:53later on, but just something to keep in
1:48:55mind. All right, so let's just dive into
1:48:57a little bit more of the details here
1:48:58that may answer some questions you guys
1:49:00have. I think it's pretty clear at this
1:49:01point what it is. Um, I'm going to give
1:49:04you guys this entire doc as well as
1:49:05anything else I've talked about in my
1:49:07free school community. The link for that
1:49:08is down in the description. So, some of
1:49:10the stuff I may not cover. If you want
1:49:11to read more about it, then just go
1:49:13ahead and grab that free resource. So,
1:49:15we know what it is. I think we know how
1:49:16it works, right? Like you define a
1:49:18routine, which is a prompt. You connect
1:49:20a GitHub repo. You could also trigger it
1:49:22by APIs or by a GitHub action and then
1:49:25you can connect your connectors and
1:49:27basically it acts as you talking to your
1:49:29own cloud code. Because of the fact that
1:49:31this is working off of a cloned repo,
1:49:33it's going to read the cloud.MD file
1:49:35automatically every time. So if you have
1:49:37a massive project like a Herk 2 project
1:49:40for example with tons of context and
1:49:41tons of stuff maybe you don't want to
1:49:43put that repo into the cloud to be a
1:49:46routine run because there's a lot of
1:49:48context in that cloudMD and in that
1:49:50whole GitHub repo that might not matter
1:49:53for this automation. So maybe you're
1:49:54better off setting up a specific GitHub
1:49:56repo per scheduled routine. But of
1:49:59course, cloud.m best practices putting
1:50:02in the information that's important
1:50:03because this stuff is going to drain
1:50:05your cloud code session limits the exact
1:50:08same way as it would if you were open up
1:50:11in cloud code just talking to it. So
1:50:12once again, three trigger types,
1:50:14schedule API, which I think is really
1:50:16cool. You could have a different
1:50:17automation make a post request to some
1:50:19sort of routine. And then of course
1:50:21GitHub so you can have it automatically
1:50:23fire off kind of on a web hook based on
1:50:25new PRs, new pushes, new issues, new
1:50:27releases, things like that. So how does
1:50:30this compare to what already exists? We
1:50:32have routines which is the new feature.
1:50:33We have desktop scheduled tasks and then
1:50:35we have something like just a /loop
1:50:37command. So routines run on Anthropics
1:50:40cloud and these other two run on your
1:50:42machine. Do you need the machine on? No,
1:50:44for routines that's huge. But for
1:50:46desktop scheduled tasks and for loop,
1:50:48you need your machine on. Do you need a
1:50:49session open? No, that's the same across
1:50:51all three. Do they survive across
1:50:53restarts? The first two do, but loop
1:50:55does not. That has to live within a
1:50:57specific session. Local file access, no
1:50:59for the routines because it works off of
1:51:01the GitHub repo. And for the next two,
1:51:03yes, you have local file access.
1:51:05Permission prompts with routines, it's
1:51:07fully autonomous. And for these two,
1:51:09they are configurable. And then the
1:51:10minimum interval routines is 1 hour. And
1:51:12these two are both could go every minute
1:51:14if you want. Okay, so let's talk about
1:51:16the environments. Obviously, your env
1:51:18get ignored unless you push it into the
1:51:20GitHub repo. You know, ultimately, if
1:51:22you push it into a private repo, you're
1:51:24probably okay, but you want to be really
1:51:26really careful because then, you know,
1:51:28there's history there and if other
1:51:29people, you know, end up collaborating
1:51:31on it, you just don't want to do that.
1:51:33So, you want to put your API keys in the
1:51:34environment variable like I showed you
1:51:36guys earlier. You want to look at the
1:51:37network access, whether that is full or
1:51:39trusted or none or custom, and
1:51:41potentially some setup scripts. So,
1:51:43that's not something I showed you guys
1:51:44yet. If you're creating a new remote
1:51:45task, you can do a setup script, which
1:51:47is basically just a script that will run
1:51:49when this new session fires up before
1:51:51cloud code launches. So if you need to
1:51:52install any packages or anything like
1:51:54that. Okay, so what's the difference
1:51:56between trusted and full? So trusted
1:51:58only reaches the known vetted services
1:52:00from Enthropic, which I thought I linked
1:52:03right here, but I just linked it there.
1:52:04This basically shows you all of the
1:52:06different domains that are allowed. So
1:52:08right here you can see we've got
1:52:09enthropic services, we've got version
1:52:11control, we've also got some cloud
1:52:12platforms like Google, stuff like this
1:52:14right here. These are the ones that are
1:52:16kind of already verified. So what is the
1:52:18risk of going on full? Well, if Claude
1:52:21reads malicious content during a run,
1:52:23then it theoretically could be tricked
1:52:24into sending data to an external server
1:52:27and with trusted that outbound request
1:52:29would get blocked. Now practical risk
1:52:31for private repos where you control the
1:52:32inputs is very low, but I definitely
1:52:34just wanted to at least acknowledge
1:52:36that. So connectors. This is different
1:52:38than just adding your API key. This is
1:52:39more of like the connectors you would
1:52:41add to your actual claude chat or like
1:52:43claude co-workth into like Slack or
1:52:46ClickUp or stuff like that. Here are
1:52:47some security details. I'm not going to
1:52:49go super deep into this. You could also
1:52:50do some more research and download this
1:52:52doc. But of course, there are some
1:52:54things to be thinking about like your
1:52:55API triggers or what's going on with
1:52:57your GitHub repos and the branches
1:52:59because once again, everything is going
1:53:01to be running as you. So if you're not
1:53:03testing out these routines before you
1:53:04just kind of send them off every hour or
1:53:06something, you just have to be thinking
1:53:07about what could happen without
1:53:09permissions and you know stuff like
1:53:10that. Limits and quotas. So it looks
1:53:12like on pro you can have five runs a
1:53:14day. On max you can have 15 runs a day
1:53:16and on team and enterprise you can have
1:53:1825 routines a day. If you hit the cap
1:53:21the orgs with extra usage enabled can
1:53:23exceed it on metered overage. And then
1:53:25we have the minimum scheduled interval
1:53:26which is one hour. And there are also
1:53:29resource limits. So every one of these
1:53:31routines in the cloud runs on four
1:53:33vCPUs, 16 gigs of RAM, and 30 gigs of
1:53:36disk space. So once again, just be
1:53:37thinking about, are you putting an
1:53:39absolutely massive GitHub repo up into
1:53:41the cloud right now to run? That could
1:53:43just be wasting resources for no reason.
1:53:45So what persists versus what gets
1:53:46destroyed? The clawed branches gets
1:53:48pushed to your GitHub repo and the
1:53:51session also stays. So as you saw, if I
1:53:53came into here and I looked at all of
1:53:55these tasks, I could see all of the past
1:53:57runs and I could go look at them to see
1:53:58if something's going wrong. but the
1:54:00actual cloud environment that gets
1:54:01cloned will be destroyed. Basically, the
1:54:04rule of thumb here is if something's
1:54:05local or if cloud code can't reach it in
1:54:08your GitHub repo or via an API, then it
1:54:11won't work. We already talked a little
1:54:12bit about writing good prompts, but you
1:54:14definitely want them to be more
1:54:15specific. For example, with my um
1:54:18scheduled automation here, this is much
1:54:20more specific, right? I have a skill
1:54:21that I wanted to run. I give it the
1:54:22order of operations, but something more
1:54:25like this YouTube comments one, this is
1:54:27not what you'd want to put in there
1:54:29unless you were defining a skill to just
1:54:30let it run because once again, this is
1:54:32supposed to be a oneshot prompt. So, you
1:54:34wanted to make sure it gets it right on
1:54:35the first try. Okay, so why is this so
1:54:37exciting and why does this beat normal
1:54:39automation? Because we are actually
1:54:41keeping the agentic framework. If you if
1:54:44you know when I talk about the WAT
1:54:45framework where we have workflows and
1:54:47agents and tools when we actually push
1:54:49those automations to the cloud and it's
1:54:51just a you know sort of a Python script
1:54:53we're losing the agentic piece we're
1:54:55only sending off really the tools and
1:54:57the workflow but in this case we're
1:54:58keeping the WA and the T all running
1:55:01together because the agent is looking at
1:55:04the you know cloud MD it's looking at
1:55:05its scripts and it's figuring out what
1:55:07to do and if it runs into errors midrun
1:55:09it will selforrect and if you configure
1:55:11it the right way it will be able to sort
1:55:13like leave a memory trail and it can
1:55:15leave like you know updates even though
1:55:17each run is stateless you can still have
1:55:19them kind of continuously get better and
1:55:21real quick let's speedrun through these
1:55:23common questions do I need to know cron
1:55:25syntax nope you just can schedule in
1:55:27natural language super easy can it
1:55:29access my local files nope it only gets
1:55:30what's in your GitHub repo or your APIs
1:55:33what model does it use can choose any of
1:55:35the models as you guys saw can you watch
1:55:37it work in real time yes you can hit run
1:55:39now and then you can obviously watch it
1:55:41go right there same way you would in
1:55:42Claude You can even talk to it after
1:55:44it's done or interrupt it and then
1:55:46continue going. Can it use my MCP
1:55:48service? Yes, that is what the
1:55:49connectors are. Can teammates use my
1:55:52routines? Nope. These belong to your
1:55:54individual account. You might be able to
1:55:55share those if you're on a team plan,
1:55:57but I haven't actually yet tested that
1:55:58myself. What's the cost? It's just your
1:56:00normal subscription usage. So, keep that
1:56:02in mind. What happens when a run fails?
1:56:04Every one of them will be stored in your
1:56:06history. So, you can go see why they
1:56:07failed. You could maybe even have it at
1:56:09the end of every single routine, say,
1:56:11"Hey, if this does fail, just shoot me a
1:56:13Slack message to let me know." Things
1:56:14like that. And can I test a run before
1:56:16going live? Yes, in fact, you should
1:56:18test it multiple times before it goes
1:56:19live. You just go into the routine, you
1:56:21hit run now, and then it will pop up as
1:56:23running. And then you just watch it, you
1:56:24know, watch it go through its order of
1:56:26operations, and you can inject, and you
1:56:28can help it correct itself. So that you
1:56:30have confidence that once it shoots off
1:56:32the prompt next time, you won't have to
1:56:34get in the way at all.
1:56:38Cloud Code can now remind you to do
1:56:39things, check on things proactively for
1:56:41you, and work for days straight without
1:56:43you ever touching it or needing to give
1:56:44any input. So, here you can see I just
1:56:46said, "Remind me at 10:23 a.m. to check
1:56:48on my project." It goes ahead and uses a
1:56:50cron create tool to set this reminder.
1:56:54There we go. 10:23 just hit. I didn't
1:56:56touch it and it just said, "Hey, Nate,
1:56:57this is the reminder to check on your
1:56:59project." So, just shot off this one
1:57:00that says, "Every 10 minutes, check my
1:57:02ClickUp to see if there's any new
1:57:03developments on our project." It's using
1:57:05the loop skill, as you can see, which is
1:57:07a new built-in skill. And it creates a
1:57:09cron for every single 10 minutes. And
1:57:11now this would run for the next 3 days,
1:57:12every 10 minutes until I told it not to.
1:57:14And this doesn't have to be every 10
1:57:16minutes. It could be every hour. It
1:57:17could be every 5 minutes. It could be
1:57:18whatever interval that you want. And
1:57:20this is all thanks to the newly released
1:57:22feature or skill loop, which is a
1:57:24powerful new way to schedule recurring
1:57:25tasks for up to 3 days at a time. And
1:57:27this is so funny because less than 12
1:57:29hours before this was announced, the
1:57:31scheduled tasks include code was also
1:57:33announced. So, right off the bat, those
1:57:35two features might seem like they're the
1:57:36exact same thing, but they're actually
1:57:37super different in how they work, and
1:57:39they have different use cases. So, in
1:57:41today's video, I'm going to break all of
1:57:42that down and tell you everything that
1:57:43you need to know about it. And by the
1:57:44way, if you haven't watched my new
1:57:45scheduled tasks video, then check that
1:57:47out right up here and then hop back over
1:57:48to this one. All right, so as you guys
1:57:50just saw in the quick demo, we now have
1:57:51the ability to use loops, which means
1:57:53that we could say something like /loop
1:57:55every 5 minutes, check on the deploy, or
1:57:57we could just say that in natural
1:57:58language, which is awesome because it
1:57:59invokes the loop skill and then it
1:58:01creates that cron job right here in
1:58:03cloud code. And you'll notice that this
1:58:04is in my VS code. So this is available
1:58:06in your terminal, in cloud code desktop
1:58:08app, in VS Code extensions, wherever.
1:58:11This is just a core part of cloud code
1:58:13now. So if you're not seeing this, just
1:58:14make sure you update your extension or
1:58:16you update cloud code. And this lets you
1:58:18set up loop intervals or reminders. So
1:58:20reminders, like you saw that first demo,
1:58:21I just said, "Hey, at this time, just
1:58:22tell me this." And in that session, it
1:58:24will bump up a message without you
1:58:25triggering it. Or you could have them be
1:58:27intervals. So you could say every 2
1:58:28hours. You could say every 30 minutes.
1:58:30Whatever you want that actual interval
1:58:32to be. And what's cool about it is it
1:58:33does it all in the same session. So if I
1:58:35leave this session up, every 10 minutes,
1:58:37it would check everything right here,
1:58:39which means that it's able to
1:58:40continuously read through what happened
1:58:41in the past one, and it continuously
1:58:43sees what we're doing. Now, obviously
1:58:44there are some pros and cons there, but
1:58:46just wanted to point that out. The major
1:58:48con there being your context, making
1:58:49sure that if something does go off every
1:58:5110 minutes, you're not going to get a
1:58:52huge report and then every 10 minutes
1:58:54you just more tokens, more tokens, and
1:58:56then context rot. It's basically
1:58:58scheduling a prompt that you would be
1:58:59sending in here and then firing off,
1:59:02which means you can loop skills. So, if
1:59:04you want every 20 minutes, for example,
1:59:06run a skill called review PR, you could
1:59:08tell it to every 20 minutes run the
1:59:10skill. It would run it, it would wait 20
1:59:12minutes, and then it would do it again.
1:59:14And of course, you could use actual
1:59:15slash commands to invoke both the loop
1:59:17and the skill. Or you could just say,
1:59:19"Every 20 minutes, run my review PR
1:59:21skill." And of course, the onetime
1:59:22reminder feature. So at 3 p.m. or in 45
1:59:25minutes, remind me to do this or check
1:59:26in on that. And Claude will basically
1:59:28pin that time, it'll create that cron,
1:59:30and then once it's done, it'll just
1:59:31delete itself. So whether that's, hey,
1:59:33at 4:30, remind me I have to go do this,
1:59:35or every hour, remind me to just stand
1:59:37up and like look away from my screen for
1:59:395 minutes, it can do that. All right.
1:59:41All right. So, there's a couple things
1:59:42that I wanted you guys to understand
1:59:43about how this actually works. So, let's
1:59:45just play around a little bit. Hey, at
1:59:4710:40 a.m., can you please remind me to
1:59:49take out the garbage? Cool. So, what
1:59:51that's going to do is it's going to use
1:59:52the cron create tool, and it's going to
1:59:55create that basically schedule to remind
1:59:56me take out the garbage. And what you
1:59:59can see here is the actual prompt. So,
2:00:01at this interval, which is just how cron
2:00:02works, it's basically going to shoot a
2:00:04prompt into this window that says remind
2:00:07Nate to take out the garbage. You can
2:00:08see the recurring equals false. Now, of
2:00:11course, the key is if the session is
2:00:13closed, then that cron is going to
2:00:14automatically be killed. So, now
2:00:16something interesting. I'm going to open
2:00:18up a new session and I'm going to say,
2:00:20"Hey, every hour, can you just remind me
2:00:22I need to stretch my neck?" And I'm
2:00:24going to shoot this one off. And we'll
2:00:25see how this one is a little bit
2:00:26different because this once again
2:00:28creates a cron. We have a prompt. And
2:00:30you can see in this one, we don't have
2:00:31the recurring equals false. We just know
2:00:34that this cron is going to go every
2:00:35hour. But these loop jobs or task jobs
2:00:38are per session. So these two tabs are
2:00:40two different sessions. So if I came
2:00:41into this session and said, "Can you
2:00:43please tell me all of the scheduled loop
2:00:45tasks that we have today?" It's going to
2:00:47use a tool called cron list, and it only
2:00:49can see the 10:40 a.m. take out the
2:00:51garbage. It cannot see the task that
2:00:53exists in this session because they're
2:00:55independent and they're separate. Now,
2:00:57one interesting thing to notice is that
2:00:58this session didn't actually invoke the
2:01:00loop tool. The loop tool basically tells
2:01:02it how to set up cron jobs and how to
2:01:04use the cron create. So if you don't see
2:01:06loop, don't worry. It's still actually
2:01:08doing this in a loop. It's just kind of
2:01:10about the actual wording. So, if I was
2:01:11to open a new one, let's see if I
2:01:13actually call the loop tool right here.
2:01:14So, I do loop and then I just say, you
2:01:16know, um, check my ClickUp. This one is
2:01:21going to default, I believe, to 10
2:01:22minutes if you don't specify a time. And
2:01:24it might invoke the loop skill because
2:01:26we actually called it to, but looks like
2:01:27it didn't because it knows exactly what
2:01:29it needs to do already. So, the point
2:01:31being, all that matters is that the cron
2:01:33is being created. It doesn't always
2:01:34matter if it invokes the loop skill or
2:01:36not. And then if you wanted to cancel
2:01:38one of these jobs, all you'd have to do
2:01:39is either close out of the terminal or
2:01:41just say, "Actually, I don't need this
2:01:42anymore. Go ahead and cancel it." And
2:01:44that one invokes a different cron skill
2:01:46called cron delete, it shoots over the
2:01:48job ID. And now that it's canceled and
2:01:50one final thing to keep in mind is in VS
2:01:52Code, if you close out of a tab and then
2:01:54you just open up that conversation
2:01:55again, that still will kill those crons.
2:01:58So you guys just saw how pretty much all
2:02:00of this worked. We have cron create to
2:02:02schedule. We have cron list to list them
2:02:04and then we have cron delete to cancel
2:02:06them and all of those can be invoked
2:02:08with natural language which is awesome.
2:02:09So now let's get into some of the
2:02:10limitations and then I'm going to
2:02:11compare them to the actual scheduled
2:02:13tasks feature. So the first big one is
2:02:15that we have a 3-day loop expiry which
2:02:17is just basically for safety. It auto
2:02:19cleans things up if you forgot you had
2:02:20all of these loops running. So once you
2:02:22create a loop it basically has a 3-day
2:02:24timer on it. It can run for day one. It
2:02:26can run on day two and then on day three
2:02:27it can run up until that last fire and
2:02:30then it will autodee. And if you want
2:02:32anything longer than 3 days, then you
2:02:34would either just recreate that loop or
2:02:35that probably indicates that you should
2:02:37just turn this into a legitimate
2:02:38scheduled task. Now, the other thing
2:02:40that you can do is if you want to
2:02:42completely disable scheduling, so maybe
2:02:43in your natural prompting it's
2:02:45accidentally creating all these crrons,
2:02:46you could go into your environment
2:02:48variables and just disable that and it
2:02:50would probably be able to help you
2:02:51figure that out. So the other things
2:02:53here are that if you close the terminal,
2:02:55your tasks are gone. It doesn't have
2:02:56catchup. So the scheduled tasks, if you,
2:02:59you know, opened up your desktop app and
2:03:01you missed a bunch, it would catch up
2:03:02automatically. This doesn't do that. And
2:03:04there's no persistence, meaning after
2:03:06your 3 days and you wanted to do that
2:03:08same loop again. It would be a fresh
2:03:09session. But obviously, there's tons of
2:03:11things you can do here with context
2:03:12management and reading different files
2:03:14in order to kind of Frankenstein your
2:03:16own fix there. So now that all the
2:03:18features have been explained and you've
2:03:19seen a demo, I think that probably you
2:03:21understand the difference between the
2:03:22loop and the schedule tasks a bit better
2:03:24now. But let's just go over some of the
2:03:26key highlights. The loop has your 3-day
2:03:28expiry. It's all done within one session
2:03:30and there's no catch-up. It's basically
2:03:32a help me now or help me on this project
2:03:34for today type of function. The schedule
2:03:36tasks are dis stored. They're longived.
2:03:39They have catchup and these are like
2:03:40daily, weekly, monthly functions that
2:03:42can run indefinitely. Of course, with
2:03:44both of these though, the terminal or
2:03:46you know the app has to be open and this
2:03:48one is only currently available in the
2:03:50desktop app. But I can imagine with how
2:03:52fast Enthropic is shipping things, maybe
2:03:54by the time you watch this video,
2:03:55scheduled tasks are already out for the
2:03:58terminal and extensions as well, the way
2:04:00that loop is available in cloud code
2:04:02everywhere. So basically, it's one
2:04:04simple question. Do you need help right
2:04:06now on a project or do you need help
2:04:08with something every day or every week?
2:04:09And that's how you decide if you use the
2:04:11new loop feature or if you use scheduled
2:04:13tasks. So I thought I'd end off real
2:04:15quick by giving a few maybe practical
2:04:17ways that you could actually use loop
2:04:19rather than something scheduled. So
2:04:20maybe all day you're waiting on a very
2:04:22urgent email. Just set up Cloud Code to
2:04:25check in on that email every 5 minutes
2:04:27and if it's there, it can automatically
2:04:28let you know. Maybe you're working on a
2:04:30deploy and you want to just pull that
2:04:31and check every, you know, hour or so if
2:04:33everything's working okay. Maybe you've
2:04:35got a deadline due at the end of the
2:04:36week and you need a 3-day sprint to be
2:04:38constantly checking in on the team and
2:04:39checking in on progress. Maybe you're
2:04:41testing and iterating. Maybe you're
2:04:42watching logs. Maybe you're tracking
2:04:43changes. There's so many different use
2:04:45cases here. There's so many different
2:04:46ways to use the loop to prompt an agent
2:04:48to have different files, to use
2:04:49different skills, and it's really,
2:04:51really cool the way that you could
2:04:52potentially set these things up. All
2:04:53right, so now you guys understand the
2:04:55routines, whether that is a local
2:04:56routine or a cloud routine, you
2:04:58understand how you can set them up.
2:05:00You're basically turning a skill into a
2:05:01schedule. You notice the pattern, you
2:05:03identify the trigger, you map it to a
2:05:05routine, and then you start to have more
2:05:07of your cadence. And I'm going to talk
2:05:08about this in a few more slides, so just
2:05:10give me a sec. But there's other things
2:05:12that you can start to do here as well.
2:05:13you can start to visualize your system a
2:05:15little bit. So, if I open up Obsidian
2:05:18real quick, do you guys remember how I
2:05:19talked about putting all of my YouTube
2:05:21transcripts into my AIOS? So, it can
2:05:24answer questions about them and it knows
2:05:26what tools I've used and what videos
2:05:27I've made and all that kind of stuff.
2:05:29This is me just visualizing all of the
2:05:31markdown files, all of my transcripts.
2:05:33And I've used sort of like this Carpathy
2:05:35LLM wiki idea to be able to visualize
2:05:38this and to be able to see connections
2:05:39between every video, every tool, how
2:05:42they're referenced, how they relate to
2:05:44each other. And it's really, really cool
2:05:45for me to be able to visualize it. Now,
2:05:48Obsidian giving me this visual layer
2:05:50doesn't change anything fundamentally
2:05:52about how my AIOS actually uses the data
2:05:55and is able to look at it and talk to it
2:05:57and edit it. But sometimes it is nice to
2:05:59be able to have that visual layer on top
2:06:01of it. So, if you guys are interested in
2:06:03figuring out how you can turn all of
2:06:05your business context, all of your, you
2:06:07know, meeting transcripts, all of your
2:06:08YouTube video transcripts into a visual
2:06:10thing like this using the LLM wiki, then
2:06:12I'm going to head into a video right
2:06:15here where I explain exactly how to set
2:06:16this up. If you're not interested, just
2:06:18skip past this part of the video. But if
2:06:19you are, go ahead and give it a watch.
2:06:21What you're looking at right here is 36
2:06:23of my most recent YouTube videos
2:06:24organized into an actual knowledge
2:06:26system that makes sense. And in today's
2:06:28video, I'm going to show you how you can
2:06:29set this up in 5 minutes. It's super
2:06:31super easy. You can see here how we have
2:06:33these different nodes and different
2:06:34patterns emerging. And as we zoom in, we
2:06:36can see what each of these little dots
2:06:38represents. So, for example, this is one
2:06:40of my videos, $10,000 aentic workflows.
2:06:43We can see it's got some tags, it's got
2:06:45the video link, it's got the raw file,
2:06:47and it gives an explanation of what this
2:06:49video is about and what the takeaways
2:06:50are. And the coolest part is I can
2:06:52follow the back links to get where I
2:06:53want. There's a backlink for the WAT
2:06:55framework. There's a backlink for Claude
2:06:58Code. There's a backlink for all these
2:06:59different tools I mentioned like
2:07:00perplexity, visual studio code, nano
2:07:02banana, nadn. It also has techniques
2:07:04like the wt framework or bypass
2:07:07permissions mode or human review
2:07:08checkpoint. So as this continues to fill
2:07:10up, we can start to see patterns and
2:07:12relationships between every tool or
2:07:14every skill or every MCP server that I
2:07:16might have talked about in a YouTube
2:07:17video and I can just query it in a
2:07:19really efficient way now that we have
2:07:21this actual system set up. And the crazy
2:07:23part is I said, "Hey, Cloud Code, go
2:07:25grab the transcripts from my recent
2:07:26videos and organize everything." I
2:07:28literally didn't have to do any manual
2:07:31relationship building here. It just
2:07:32figured it all out on its own. And then
2:07:34right here, I have a much smaller one,
2:07:36but this is more of my personal brain.
2:07:37So, this is stuff going on in my
2:07:38personal life. This is stuff going on
2:07:40with, you know, Up AI or my YouTube
2:07:42channel or my different businesses and
2:07:43my employees and our quarter 2
2:07:45initiatives and things like that. This
2:07:47is more of my own second brain. So, I've
2:07:48got one second brain here, and then I've
2:07:50got one basically YouTube knowledge
2:07:52system, and I could combine these, or I
2:07:54could keep them separate, and I can just
2:07:56keep building more knowledge systems and
2:07:57plug them all into other AI agents that
2:07:59I need to have this context. It's just
2:08:01super cool. So, Andre Carpathy just
2:08:03released this little post about LLM
2:08:05knowledge bases and explaining what he's
2:08:07been doing with them. And in just a
2:08:08matter of few days, it got a ton of
2:08:09traction on X. So, let's do a quick
2:08:11breakdown and then I'm going to show you
2:08:12guys how you can get this set up in
2:08:13basically 5 minutes. It's way more
2:08:15simple than you may think. Something
2:08:17I've been finding very useful recently
2:08:18is using LLM to build personal knowledge
2:08:21bases for various topics of research
2:08:23interest. So there's different stages.
2:08:24The first part is data ingest. He puts
2:08:27in basically source documents. So he
2:08:28basically takes a PDF and puts it into
2:08:30cloud code and then cloud code does the
2:08:32rest. He uses Obsidian as the IDE. So
2:08:34this is nothing really too gamechanging.
2:08:36Obsidian just lets you visually see your
2:08:38markdown files. But for example, this
2:08:40Obsidian project right here with all
2:08:41this YouTube transcript stuff that
2:08:43actually lives right here. This is the
2:08:44exact same thing. and here are the raw
2:08:46YouTube transcripts and here's that wiki
2:08:47that I showed you guys with the
2:08:48different um folders for what Cloud Code
2:08:52did with my YouTube transcripts. And
2:08:54then there's a Q&A phase where you
2:08:55basically can ask questions about
2:08:57YouTube or about the research and it can
2:08:59look through the entire wiki in a much
2:09:01more efficient way and it can give you
2:09:03answers that are super intelligent. He
2:09:05said here, "I thought that I had to
2:09:06reach for fancy rag, but the LLM has
2:09:08been pretty good about automaintaining
2:09:09index files and brief summaries of all
2:09:11documents, and it reads all the
2:09:13important related data fairly easily at
2:09:15this small scale. So, right now he's
2:09:17doing about 100 articles and about half
2:09:19a million words. So, there's a few other
2:09:20things that we'll cover later, but the
2:09:22TLDDR is you give raw data to cloud
2:09:24code. It compares it, it organizes it,
2:09:26and then it puts it into the right spots
2:09:28with relationships, and then you can
2:09:29query it about anything. And it can help
2:09:31you identify where there's gaps in that
2:09:33node or in that you know relationship
2:09:35and it can go do research and fill in
2:09:37the gaps. All right. So why is this a
2:09:38big deal? Because normal AI chats are
2:09:40ephemeral meaning the knowledge
2:09:42disappears after the conversation. But
2:09:44this method using Karpathy's LLM wiki
2:09:46makes knowledge compound like interest
2:09:48in a bank. People on X are calling it a
2:09:50game changanger because it finally makes
2:09:51AI feel like a tireless colleague who
2:09:53actually remembers everything and it
2:09:54stays organized. It's also super simple.
2:09:57It will take you 5 minutes to set up.
2:09:58I'll show you guys. You don't need a
2:10:00fancy vector database embeddings or
2:10:02complex infrastructure. It's literally
2:10:03just a folder with markdown files.
2:10:05That's it. You literally just have a
2:10:07vault up top. So in this example, it's
2:10:08called my wiki. You've got a raw folder
2:10:10where you put all of the stuff. And then
2:10:12you've got a wiki folder, which is what
2:10:14the LLM takes from your raw and puts it
2:10:16into the wiki. So in here, you have all
2:10:18the wiki pages which it will create. But
2:10:20then you also have an index and you have
2:10:21a log. So for example, in my YouTube
2:10:23transcripts vault, here is the index.
2:10:26You can see that I have all these
2:10:27different tools which I could obviously
2:10:28click on and it would take me right to
2:10:29that page or after that I have all the
2:10:31different techniques agent teams sub
2:10:33agents permission modes the WAT
2:10:36framework and then we've got different
2:10:37concepts MCP servers rag vibe coding
2:10:40we've got all these different sources
2:10:41which are you know the YouTube videos
2:10:43and then when I have people or when I
2:10:44have comparisons they will be put in
2:10:46here in the index and then we also have
2:10:48a log which is the operation history so
2:10:50in this case in the YouTube project the
2:10:51log isn't huge cuz I only ran one huge
2:10:53batch of the initial 36 YouTube videos,
2:10:56but now every time I have one, I say,
2:10:58"Hey, can you go ahead and ingest the
2:11:00new YouTube video into the wiki and then
2:11:02we'll see every single time we update
2:11:04this." And then, of course, you need
2:11:05your claw.md to explain how the project
2:11:08works and how to search through things
2:11:09and how to, you know, update things.
2:11:11It's also a big deal from a cost
2:11:13perspective, token efficiency, and
2:11:14long-term value. One X user turned 383
2:11:17scattered files and over a 100 meeting
2:11:19transcripts into a compact wiki and
2:11:21dropped token usage by 95% when querying
2:11:24with Claude. And obviously token
2:11:26management and efficiency is a huge
2:11:27conversation right now and will always
2:11:29be. The other thing that's really cool
2:11:30about this is there's not really like a
2:11:32GitHub repo you go copy or there's not a
2:11:34complicated setup. You literally just
2:11:36say hey cloud code read this idea from
2:11:38Andre Karpathy and implement it. And
2:11:40people on X are now talking about like
2:11:41this is how 2026 AI agentic software and
2:11:44products will be made. You just give it
2:11:46a highle idea and it goes and builds it
2:11:48out. And Karpathy even said, "Hey, you
2:11:50know, I left this prompt vague so that
2:11:52you guys can customize it." And I'll
2:11:54show you the ways in my two different
2:11:55vaults right now that it changed things
2:11:57a little bit based on the context and
2:11:58understanding of what the project is
2:12:00actually for. Okay, so this was the
2:12:01original tweet I just showed you guys
2:12:02and then he followed up and said, "Hey,
2:12:04this one went viral. So here is the idea
2:12:06in a gist format." So if you open this
2:12:07up, this is basically just another
2:12:09explanation of the core idea of how this
2:12:11works and why the architecture indexing
2:12:14all this kind of stuff. And by the way,
2:12:15this is the part where he says, "Hey,
2:12:16this is left vague so that you can hack
2:12:18it and customize it to your own
2:12:20project." So we're going to come right
2:12:21back to this in a sec, but the first
2:12:22prerec that we're going to do, it's not
2:12:24necessary, but I like to have a nice
2:12:26little front end to see the
2:12:27relationships is we're going to go to
2:12:29Obsidian and download it. So, if you
2:12:31just go to obsidian.mmd, you can see
2:12:33this is the completely free tool and
2:12:35you're going to go ahead and download
2:12:36it. So, just for your operating system,
2:12:38download this and then open up the
2:12:41wizard and open up the app. So, when you
2:12:43open up the app, it'll look like this.
2:12:44And what we're going to do here is we're
2:12:45going to create a new vault. So, down
2:12:47here, you can see I have Herkbrain and I
2:12:48have YouTube transcripts. I'll just make
2:12:50it a little bigger. I'm going to go to
2:12:52manage vaults. I'm going to create a new
2:12:54one. And now, we just have to give this
2:12:55a name. So, I'm just going to call this
2:12:57one demo vault. and you're going to
2:12:59choose a location where you want to put
2:13:00this. So, I'm just throwing this on my
2:13:02desktop and I'm going to go ahead and
2:13:03create this vault. Then, what you're
2:13:05going to do is go to wherever you like
2:13:07to run Cloud Code. So, in this case, I'm
2:13:09doing it in VS Code. And I open up that
2:13:11folder. So, demo vault. We get an
2:13:12Obsidian and then we get a welcome.md.
2:13:16So, I'm going to open up Claude. So, I'm
2:13:17going to do it in my terminal. I'm going
2:13:18to run Claude. And lately, I've been
2:13:20liking using my terminal better for
2:13:22Claude. I like to do it inside of VS
2:13:23Code, but the reason is just because I
2:13:25like to see the status line and I have,
2:13:27you know, a little bit more
2:13:27functionality. So, anyways, now that we
2:13:30have Cloud Code open, here's what we're
2:13:31going to do. We're going to go back over
2:13:33to the LLM wiki thing that we got from
2:13:34Andre Carpathy. We're going to copy all
2:13:37of this and we're going to go back into
2:13:40Cloud Code and then just paste it in
2:13:42there. So, that is the prompt from
2:13:44Carpathy that's going to build out
2:13:46everything we need. And then before we
2:13:47send that off, we're dropping this in
2:13:49which you guys can screenshot and then
2:13:50just throw into yours. But I'm saying
2:13:52you are now my LLM wiki agent. Implement
2:13:55this exact idea file as my complete
2:13:57second brain. Guide me step by step.
2:13:59Create the cloudmd schema blah blah
2:14:00blah. So this is just telling it what it
2:14:02needs to do with this idea that we just
2:14:04got from Kapathy. So anyways, on the
2:14:06right we have this cloud code running
2:14:08and on the left we have our obsidian
2:14:09vault and you can see it just created
2:14:11those two folders. So it created the raw
2:14:13and it created the wiki as you can see.
2:14:15Now, by default, it threw in four
2:14:16folders. It threw in analysis, concepts,
2:14:18entities, and sources. Once we start to
2:14:20populate stuff, we can talk to it to see
2:14:21if that's actually the way we want to do
2:14:23it or not. Because it's interesting in
2:14:25my personal kind of second brain, the
2:14:27wiki is literally just markdown files.
2:14:29There's no structure to it. And in some
2:14:31cases, that's good. Carpathy actually
2:14:32said, "Sometimes I like to keep it
2:14:34really simple and really flat, which
2:14:36means like no subfolders and not a bunch
2:14:38of over organizing." But then you guys
2:14:40did see in my YouTube transcript one,
2:14:42there were different subfolders. And I
2:14:43think that in this case it actually
2:14:44makes more sense. But you can see that
2:14:46it went ahead and it created a claw.md.
2:14:48It created an index and a log and then a
2:14:50few different folders in our wiki. But
2:14:52now it's saying, "Hey, let's go ahead
2:14:52and try it out. Drop in your first
2:14:54source into the raw folder and tell me
2:14:56to ingest it." Okay, so I'm at this
2:14:58website called AI2027. If you guys
2:14:59haven't read this before, it's kind of
2:15:01an interesting read. So go check it out.
2:15:03And now let's say I want to get this
2:15:04into my vault. What I could do is just
2:15:06copy the whole page, right? And it might
2:15:08just come through a little weird. or we
2:15:10can just get an Obsidian extension which
2:15:12lets us basically take articles right
2:15:14from the web and just put it right into
2:15:15our vault. Super easy. So search for
2:15:17this extension called Obsidian Web
2:15:19Clipper. You would go ahead and add this
2:15:20to Chrome. So then when you're in the
2:15:22article that you want, you basically
2:15:23just click on your extensions, you open
2:15:24up Obsidian Web Clipper, and then you
2:15:26can just chuck it into your vault. And
2:15:28then right here, you're going to want to
2:15:28set this to RAW because this is the
2:15:30actual folder that it's going to put it
2:15:31in. So you can go ahead and click add to
2:15:33Obsidian. Open Obsidian. And then now
2:15:35you can see in my raw section we have
2:15:37this AI 2027 source with the title the
2:15:41source and it's not super super
2:15:43populated yet because the LLM in cloud
2:15:46code is going to do that. So here is our
2:15:48file. I'm going to open up cloud code
2:15:49and say awesome. I just threw in an
2:15:51article called AI 2027 into the raw. Can
2:15:54you please go ahead and ingest that? It
2:15:56might ask you some questions. It might
2:15:57also be helpful to before you start
2:15:59ingesting stuff say hey by the way this
2:16:01project is specifically for my second
2:16:02brain. So, personal things, business
2:16:04things, whatever. Or this is just a
2:16:06research project. This is where I'm
2:16:07going to chuck you all the articles and
2:16:09all the things that I want to learn
2:16:10about and all the things that I know.
2:16:11So, there's different ways that you can
2:16:12set up the project as you saw with mine.
2:16:14One for YouTube, one for just personal
2:16:17second brain. So, now what it's doing is
2:16:19it's going to read through this article
2:16:20and then it's going to figure out where
2:16:22should I chuck everything into the wiki.
2:16:23It's not just going to create one MD
2:16:25file for this. It might create five or
2:16:27it might create 10. And there's going to
2:16:29be relationships between each of the
2:16:30different sections that it creates. So,
2:16:31it's kind of doing its own method of
2:16:33chunking. Now, one thing I want to call
2:16:34out real quick is with this extension,
2:16:36if you go here and you open up the
2:16:38options for it, you can see that you can
2:16:40actually change where by default the
2:16:43folders are dropped, which is in the
2:16:45location section. By default, it'll be
2:16:47going to a place called clippings, but
2:16:48just go ahead and change that to raw.
2:16:50Okay. So, here it came back with all
2:16:52these questions, right? It said, "Here
2:16:53are my key takeaways from this article,
2:16:55blah blah blah." And now it'll ask you,
2:16:56"What do you want to emphasize from this
2:16:58article? What's your focus? How granular
2:16:59do you want to be? what's your plan? So,
2:17:01I'm just going to say I want this to be
2:17:03extremely thorough. This is my passion
2:17:06looking at where AI is going to go. Um,
2:17:08and this whole project, by the way, that
2:17:10you're setting up in this vault is
2:17:11basically just going to be my place to
2:17:13dump in research about AI. So, help me
2:17:15keep all that organized so that I can
2:17:16query it and that I can, you know, keep
2:17:18my thoughts related. So, that's just a
2:17:21quick example of what it might look like
2:17:22for you to give it some more context to
2:17:24continuously build your project. So, I'm
2:17:27going to switch over over here to the
2:17:28graph view because I think it it'll be
2:17:30interesting to see as it is starting to
2:17:31go through and create those different
2:17:33wiki files. It's going to go ahead and
2:17:35it's going to create all those
2:17:36relationships and we'll be able to watch
2:17:37it in real time. All right, so it's
2:17:38creating all of the wiki pages now and
2:17:40you can see that it said it's going to
2:17:42make about 25 because there's so much
2:17:44stuff going on in the original AI 2027
2:17:47article. Okay, so our first one just
2:17:48popped in here and there a second one
2:17:50just came through and now you can
2:17:51understand you're starting to see where
2:17:52do you have hubs or where do you just
2:17:54have little individual nodes? So this is
2:17:56a major hub someone named Eli someone
2:17:59named Thomas Daniel and you can see all
2:18:01the different relationships here with
2:18:02things like AI governance with things
2:18:04like open brain superhuman coder. Okay
2:18:07so that ingest took about 10 minutes. So
2:18:09sometimes you have to be a little
2:18:10patient with you know it reading through
2:18:12everything and organizing everything but
2:18:13it does a lot of heavy lifting of
2:18:15course. When I uploaded the 36 YouTube
2:18:17transcripts in batch it took about 14
2:18:19minutes. So it kind of just depends but
2:18:22it created 23 wiki pages. We have the
2:18:24source. We have six people, five
2:18:26organizations and one AI systems page,
2:18:29different concepts, so technical
2:18:30alignment and geopolitical and then an
2:18:33analysis and then it asks some questions
2:18:35about it so that it can help make the
2:18:36relationships and make the structure
2:18:38even better. Now let's just open this
2:18:40one up a little bit deeper and see what
2:18:42it actually did in here with this stuff.
2:18:44So we have this is the source with all
2:18:46the main relationships. So as we start
2:18:48to add other articles, we will see other
2:18:50big kind of like nodes and maybe in some
2:18:52cases we'll have relationships between
2:18:54like compute scaling with different
2:18:56articles that we upload as well. So
2:18:57let's just see if I click into the main
2:18:59source, we can see the tags that it got.
2:19:01We can see the authors and we can click
2:19:03around. So here's a link to OpenAI.
2:19:05Okay, what's OpenAI? Here's references
2:19:06in AI 2027. Here's some other
2:19:08connections with OpenAI like modelsp
2:19:10spec. Okay, we're in model spec. Let's
2:19:12take a look. We can see other things
2:19:13about modelsp spec. And we could also go
2:19:15to how the LLM psychology model works.
2:19:17So this is just super super cool all the
2:19:19relationships that we get. And once
2:19:21again, all of this stuff that we're
2:19:22looking at was derived from one article
2:19:24and automatically organized and related.
2:19:27So the question now is like what do we
2:19:28do from here? Do we query it inside of
2:19:31this environment? Do we query it from
2:19:32somewhere else? And that's completely up
2:19:34to the way that you want to use this. So
2:19:36for example, with my YouTube project,
2:19:38I'm probably just going to keep this
2:19:39here. And whenever I want to ask
2:19:40questions about YouTube or if I want to
2:19:42turn this into like a website, I can
2:19:43just do that from here. Or if I need to,
2:19:46I can point a different project at this
2:19:48folder since everything's here and it
2:19:49can crawl through the wiki, it can read
2:19:51the index and it knows how this stuff
2:19:52works because you can give it the clawmd
2:19:55so it understands the project as well.
2:19:56So for example in this one which is just
2:19:58my second brain where we have all of the
2:20:00different things about like I drop in my
2:20:01meeting recordings, I drop in, you know,
2:20:03ClickUp channels, summaries and things
2:20:05like that. This is something that I want
2:20:06to use in my executive assistant. So
2:20:08what I did in my executive assistant
2:20:10here called Herk 2. If I go to this
2:20:11cloud.MD, MD. You can see that we have a
2:20:14wiki path. So, whenever you need to read
2:20:16things about me and my business that you
2:20:18don't have already, you would basically
2:20:19go to my Herkb brain vault. You would go
2:20:21to that directory and then you would
2:20:23read through the wiki. You can read the
2:20:24hot cache, which I'll explain in just a
2:20:26sec. You can read the index. You can
2:20:28read the domain subindex. And then you
2:20:30can also just search through everything
2:20:31here. And I said don't read from the
2:20:32wiki unless you actually need it. Here
2:20:34are some things that you might do that
2:20:36you don't need to go read the wiki for.
2:20:37And all of this is my business
2:20:38knowledge. Now, if you guys remember, if
2:20:40you watched my video on setting up an
2:20:41executive assistant, I used to do this
2:20:43with context files inside of this
2:20:45project. And when I changed over to this
2:20:47method, I actually saw a reduction in
2:20:49tokens that I was actually calling in
2:20:51this project. So, the thing about the
2:20:53hot cache, right, I didn't actually have
2:20:55this in my YouTube one. So, if I go to
2:20:57YouTube, you can see there's no hot
2:20:58cache. But, if I go to the herk brain in
2:21:01the wiki, you can see there's a hot.md
2:21:03right here. And this is basically just a
2:21:05cache of like 500 words or 500
2:21:07characters that it saves, which is like
2:21:09what is the most recent thing that Nate
2:21:11just gave me or that we talked about. In
2:21:13the context of my executive assistant,
2:21:14this is really helpful. You know, it
2:21:16might save me from having to crawl
2:21:17different wiki pages. But in something
2:21:19like the YouTube transcript project, I
2:21:21don't really need a hot cache. So,
2:21:23another thing that I alluded to but
2:21:25didn't really cover was the idea of
2:21:26linting. So, Karpathi says that he runs
2:21:28some LLM health checks over the wiki to
2:21:30find inconsistent data, impute missing
2:21:33data with web searches, find interesting
2:21:35connections for new article candidates,
2:21:37things like that. So, it basically helps
2:21:39you run a lint, you know, every day,
2:21:41every week, whenever you want, which
2:21:42helps make sure that everything is
2:21:44scalable and structured in the right
2:21:45way. And it might even come back and
2:21:48say, "Hey, I don't fully understand
2:21:49this. Can you give me some more info or
2:21:50can you grab some more articles that
2:21:52might help me out here?" So now the
2:21:53final question about this that I wanted
2:21:55to cover is like does this kill semantic
2:21:57search rag? And the answer is no, but
2:21:59kind of yes. And it all depends on the
2:22:02goal of the project and the goal of the
2:22:04context, how much context you have. So
2:22:06here's a really quick chart that I had
2:22:08my cloud code make. I was in my Herk
2:22:10brain where I dumped in a bunch of
2:22:11information about Karpathy's LLM
2:22:13knowledge and I just said, "Hey, can you
2:22:16please explain Karpathy knowledge as
2:22:18simple as possible, keep it super
2:22:19concise and um compare it to typical
2:22:22semantic search rag." So, it found
2:22:25Karpathy's idea. Instead of a database,
2:22:27you just give the LM well organized
2:22:28markdown files and it compares it here
2:22:30to the actual semantic search rag. So,
2:22:33actually, I might as well just read it
2:22:34off from here. So it finds it by reading
2:22:36indexes and follows links rather than
2:22:38using similarity search. So we're
2:22:40getting a deeper understanding of
2:22:41relationships because they're links
2:22:43rather than just saying, "Hey, these
2:22:45chunks seem similar." As far as
2:22:46infrastructure, it is literally just
2:22:47markdown. So like I said, you don't even
2:22:49need the obsidian. You just need these
2:22:51markdown files. Whereas with semantic
2:22:53search, you need an embedding model. You
2:22:54need a vector database and a chunking
2:22:56pipeline. The cost over here is
2:22:58basically free. Your only cost is going
2:22:59to be tokens. Whereas over here, you
2:23:01might have ongoing compute and storage.
2:23:03And for maintenance, you just run a
2:23:04lint. You clean up things. You add more
2:23:06articles. You give it more context
2:23:08rather than having to re-mbed when
2:23:10things change. But right now, the
2:23:12weakness of course with the uh LLM
2:23:14knowledge wiki is that it doesn't scale
2:23:17huge across enterprises, right? Because
2:23:19it's just a bunch of files. Um and that
2:23:21is where the cost will probably get more
2:23:23and more expensive than going to
2:23:25something like standard semantic search
2:23:27or knowledger graph or light rag or
2:23:29whatever other tool is out there for
2:23:30that. So here you can see if you have
2:23:31hundreds of pages with good indexes,
2:23:33you're fine with wiki graph. But if you
2:23:35were getting up to the millions of
2:23:36documents, then you're going to want to
2:23:37actually do more of a traditional rag
2:23:39pipeline. At least for now with how the
2:23:41current models are and everything we
2:23:43know right now in April 2026. Okay,
2:23:46awesome. So that was Obsidian stuff. Now
2:23:49let's talk about what else you could do
2:23:50with clawed artifacts because you have
2:23:53all of these different connections,
2:23:54right? You're at this point now where
2:23:56you have a bunch of stuff hooked up. So
2:23:58you could have, you know, your school
2:24:00dashboard hooked up, your Stripe
2:24:01information, your YouTube stuff, your
2:24:03ClickUp tasks, but what if you want to
2:24:05be able to have a visual place to see
2:24:07that stuff? Because, you know, once
2:24:09again, we now have potentially a
2:24:11visualization of our data, or sorry, not
2:24:13our data, but like our knowledge. But
2:24:15what if we want a visualization of our
2:24:17data? Well, you could go ahead and
2:24:19custom build all of this. But the first
2:24:21step, I think, would just be to leverage
2:24:23cloud artifacts and see what you can
2:24:24already do. So once again, I'm in the
2:24:26desktop app. I'm going to go to co-work,
2:24:28which is the middle section, and I'm
2:24:29going to click on live artifacts. And
2:24:31you can see here that I've got three
2:24:32different dashboards here. I've got a
2:24:34Fireflies dashboard, I've got a weekly
2:24:36commitments dashboard, and I've got a
2:24:37QuickBooks dashboard. So, for example,
2:24:39I'm going to click on the QuickBooks
2:24:40dashboard. I'm not going to show you
2:24:42guys this data obviously, but what this
2:24:44does is as soon as I open up this
2:24:45dashboard, it makes a call to QuickBooks
2:24:48and it starts to pull in this data and
2:24:50populate it with real analytics. And so
2:24:53now I have this really simple place in
2:24:54my clawed environment where I can
2:24:56actually see our revenue, our expenses,
2:24:58our net profit, our cash on hand. I can
2:25:00see the trends. I can even get an AI
2:25:02breakdown of our runway and it's looking
2:25:04forward and it's doing a bit of a, you
2:25:06know, financial analysis. I can see all
2:25:08of this information right here. I could
2:25:10also do this for my weekly commitments.
2:25:12So I'm looking right now, this is
2:25:13pulling data from ClickUp and it's
2:25:15organizing it to show me tasks,
2:25:17completion percentage, what's at risk,
2:25:19who I might need to follow up with. So
2:25:20once again, it's just another way for me
2:25:22to visualize all of this data. And you
2:25:24can set them up with artifacts super
2:25:26super quickly. You just come into Cloud
2:25:28Co-work. You would click on new
2:25:29artifact. And this basically just starts
2:25:31a new conversation with Claude and it
2:25:33says, "Okay, hey, let's build you a
2:25:35dashboard. What do you want to see? What
2:25:36MCP servers can we connect to? What
2:25:38other connectors do we have access to?
2:25:40Let's build you a visualization." Now,
2:25:42that is really nice, but it's not
2:25:43perfect. And ultimately where you might
2:25:45want to go is using Cloud Code to build
2:25:48you an actual dashboard that not only
2:25:50just has like, okay, your QuickBooks is
2:25:52this one and your ClickUp is this one,
2:25:54but maybe you want a dashboard to see
2:25:55everything because if you have all these
2:25:57connectors, wouldn't it be great to just
2:25:59have one dashboard for your business
2:26:01where you could look at all this stuff?
2:26:02Yeah, it would be great, but it requires
2:26:05a little bit more configuration because
2:26:07obviously you have to sync all the data,
2:26:09but then you have to build some sort of
2:26:10refresh cadence. Maybe you want every
2:26:12single night all of that data to
2:26:13refresh. Maybe you want it to refresh
2:26:15every 25 minutes. You would have to
2:26:17build out that logic and you would have
2:26:19to build it up as a routine or you would
2:26:21have to put that on something like
2:26:22trigger.dev or modal. So there's just a
2:26:24little bit more of, you know,
2:26:26complexity. There's more moving pieces,
2:26:28but it's definitely doable. And so the
2:26:30reason why I recommend starting with
2:26:31artifacts is because artifacts you can
2:26:34spin up in 5 minutes. And if you spin up
2:26:36an artifact and then you realize, "Wow,
2:26:38I didn't even check this all week." Then
2:26:40why would you have went through all the
2:26:42effort to build your own custom
2:26:44dashboard? You know, maybe you build an
2:26:46artifact and you're like, "Wow, I
2:26:47actually check this every, you know, I
2:26:49check this four times a day. I should
2:26:50probably build my own custom one." Then
2:26:52you know, so another one of Nate's
2:26:54mindset shifts is P. What does PC mean?
2:26:58Proof of concept. You want to build
2:27:00something that's super easy and it's
2:27:02lightweight enough that it proves yes or
2:27:05no. Don't waste time. Don't waste money
2:27:08investing into something that might not
2:27:10be proven yet. So prove the concept
2:27:12first. Lightweight, cheap, build a cloud
2:27:15artifact. If you like it, if the concept
2:27:16is proven, then you can dedicate some
2:27:18more resources towards building a custom
2:27:20solution. Because honestly, what I've
2:27:21realized is I don't really need that
2:27:23much data visualization, at least in in
2:27:25that sense, because of the fact that I
2:27:27can just message my assistant and I can
2:27:29get answers instantly. And if I ever
2:27:30want to generate trends and I want to
2:27:32see what's going on, I just say, "Hey,
2:27:34can you pull data from these five
2:27:35sources from the past month and generate
2:27:37me a dashboard or generate me a report
2:27:40because that way I'm able to get stuff
2:27:41when I need it on demand rather than
2:27:43having to spend time building and
2:27:44maintaining some sort of dashboard. All
2:27:46right, so now let's say we're at a
2:27:47really good spot. We've got a bunch of
2:27:48context. We've got a bunch of
2:27:49connections. We've got a ton of
2:27:50capabilities and cadence. What does it
2:27:52look like on the dayto-day and on the
2:27:54week to week? Well, first of all,
2:27:55obviously trying to do as much work as
2:27:57you can just in cloud code and not
2:27:59leaving and switching between different
2:28:00tabs and context switching. But on top
2:28:02of that, let's talk about the daily
2:28:03loop. So, what I would start to do is
2:28:05every morning go into cloud code and
2:28:07say, "Hey, help me plan my day." If it's
2:28:10helping you plan your day in a really
2:28:11good way, meaning it can look at your
2:28:13priorities and look through messages and
2:28:14look through your calendar, that's
2:28:15really good. Keep doing that. But if it
2:28:18doesn't, note down why not. What might
2:28:20it be missing? What context might it
2:28:22need in order to plan your day better?
2:28:25And then at the end of the day, look
2:28:27back at it. What skills did you use
2:28:29today? What did you have to correct your
2:28:30assistant on? What did you have to, you
2:28:32know, copy and paste in? And then
2:28:35tomorrow, you can sort of patch some of
2:28:36that stuff up and just make small
2:28:37iterations as you go. And then on the
2:28:39week to week, of course, you can use the
2:28:41audit. Maybe every single Friday, you
2:28:42run an audit and you see what you need
2:28:44to do. Think about how many skills got
2:28:45used. Think about which ones got used
2:28:47daily. If they're getting used so often,
2:28:49why not just automate them? And the
2:28:51other thing to think about is sometimes
2:28:52when you actually decide, I need to
2:28:54build an automation, it might be
2:28:56overkill to go with an agentic skill.
2:28:58What you might actually need is more of
2:29:00a boring is beautiful workflow
2:29:02automation because workflows,
2:29:04deterministic workflows, beat AI agents
2:29:08nine times out of 10. Most of the stuff
2:29:09that we were doing for businesses were
2:29:10just automations. We barely even used AI
2:29:12sometimes. And hardly did we ever give
2:29:14them like a fully autonomous AI agent.
2:29:16It just truly wasn't needed for most
2:29:18business processes, especially
2:29:21especially when you start to actually
2:29:22break them down by task. You know,
2:29:24automate one chunk at a time and you'll
2:29:26realize how little AI and how little
2:29:29autonomy is actually needed for each
2:29:31individual chunk of a process. And if
2:29:33you do get to that point where you want
2:29:34to be able to build something that is
2:29:36more of a regular Python script or more
2:29:38of a deterministic automation then then
2:29:40basically just ask cloud code to help
2:29:42you build that and then you can say hey
2:29:44I need to push this into trigger or
2:29:46modal so that it can actually run 24/7.
2:29:49It's really really easy. It literally
2:29:51just takes that natural language in
2:29:52order to actually do. So let's talk
2:29:54about some of the success criteria that
2:29:56I think about when I try to understand
2:29:58is my AIOS actually providing me value
2:30:01and is it working? And I wanted to call
2:30:02these KPIs, but I didn't because KPIs
2:30:05usually indicate that there's something
2:30:07objective. There's an objective metric
2:30:08that you're looking for. And we're not
2:30:10looking for really objective metrics
2:30:12here. We're just looking for kind of
2:30:13subjective success criteria. So the
2:30:16first thing is that your team could
2:30:18reach out to the AIOS. Meaning
2:30:20ultimately it'd be great if they could
2:30:21just message your AI OS instead of you
2:30:24because if the AIOS knows more info than
2:30:27you or has better memory than you and
2:30:29can pull the exact source and it never
2:30:31sleeps, it would actually be more
2:30:32efficient for people to message that
2:30:34instead of you. And I started to realize
2:30:35when someone would ask me a question, I
2:30:37would take their question and I would
2:30:38just paste it into my AIOS and then
2:30:40paste the answer back to them because
2:30:43once again, my AIOS had more data than I
2:30:46did. Number two is that you stop opening
2:30:47new tabs. You don't have thousands of
2:30:50browser tabs open. You don't have
2:30:51thousands of desktop apps open because
2:30:54you're doing as much of your work as
2:30:55possible right from cloud code itself.
2:30:59Number three is that knowledge leaves
2:31:00your head. So you stop having to worry
2:31:02so much about all these little things
2:31:03you have to remember. You stop having a
2:31:05million post-it notes everywhere. You
2:31:07just have all of your knowledge inside
2:31:09of your system. It can send you
2:31:10reminders. It can make sure you're not
2:31:11losing track of things. And it helps you
2:31:13out in that way. And honestly, it can
2:31:15get rid of a lot of your stress. And if
2:31:17even just two out of three of these
2:31:18things are true within the first month,
2:31:20then it definitely took and you're
2:31:21definitely starting to get that
2:31:22exponential growth from building an
2:31:25AIOS. And once again, just wanted to hit
2:31:27this one more time. This is personal.
2:31:30Like the whole point of me building this
2:31:32course was for your personal gain, for
2:31:35your personal productivity. It's great
2:31:37to be able to think about, okay, my
2:31:38whole team needs to adopt this or I want
2:31:40to start teaching other people how to do
2:31:41this. That's great. But you have to
2:31:43start with yourself. You can't scale a
2:31:45system if you haven't lived in it
2:31:46yourself. You can't help your team
2:31:48connect to all these different data
2:31:49sources if you haven't connected to them
2:31:51yourself. And once you've set up your
2:31:53own personal AIOS, your whole company
2:31:56can start to build around your AIOS. A
2:31:58company where every operator runs a
2:32:00personal AIOS is a company that is truly
2:32:03AI ready and all the data is AI ready.
2:32:06Because think about it, if I didn't
2:32:07realize that I needed to do everything
2:32:08through my Google Workspace, then I
2:32:10wouldn't have realized the benefit of
2:32:12having a shared drive and I wouldn't
2:32:13have been able to communicate to
2:32:14everyone else to use GWS CLI to get all
2:32:17of their stuff into the shared drive and
2:32:19organized. And now all of us are
2:32:21benefiting from that. All right, so I've
2:32:22been standing here yapping at you guys
2:32:24for way too long now. Um, I really hope
2:32:27you guys enjoyed. I hope that this gave
2:32:28you some structure and it got you really
2:32:30excited about what you can actually do
2:32:31and where you can take this thing. If
2:32:34you don't have a current voice to text
2:32:35tool, please go download Glyo. I
2:32:37actually see a future very soon where we
2:32:39have more of a voice OS where I could
2:32:41throw away my mouse and keyboard and I
2:32:43could just talk to Claude Code inside of
2:32:45my AOS. So, I've got my voice OS feeding
2:32:48into my AIOS and I'm literally just
2:32:51sitting here, you know, I'm just laying
2:32:52in bed talking all day and work is
2:32:54getting done. So, if you want to support
2:32:55and if you need a voice tool, check out
2:32:56Glido. Link is in the description. But
2:32:58that is going to do it for today. I
2:33:00really hope you guys enjoyed. And now
2:33:01that you're set up in an environment
2:33:02where you're using cloud code way more
2:33:05hopefully, you're going to need to
2:33:06really start to think about context
2:33:07management. So, I'm going to highly
2:33:09recommend you check out this video next
2:33:10where I talk about how to make sure
2:33:11you're not hitting your session limit
2:33:13super super quick. And I'll see you guys
2:33:15over there. As always, thanks for making
2:33:16it to the end of the video and see you
2:33:18in the next one. Thanks everyone.

About this transcript

This page contains the full transcript of Build & Sell Claude Code Operating Systems (2+ Hour Course) by Nate Herk | AI Automation, generated from the public captions YouTube serves with the video. The transcript has 36,834 words across 5,017 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.