YT YouTube2Text

YouTube2Text → Transcripts → Building AI Agents that actually work (Full Course)

Building AI Agents that actually work (Full Course) — Transcript

by Greg Isenberg · 10,815 words · 1,754 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I think AI is confusing. There, I said
0:02it. I think there's a lot of terms,
0:04skills, MCPs, agent harnesses that are
0:08difficult concepts to understand. So, I
0:10had my friend Remy come on the podcast
0:13and explain it in the most simple terms
0:16possible. In this free course on how to
0:19master AI agents, he breaks down exactly
0:22what each piece is, how they connect
0:25together, and the simplest ways
0:27beginners could start using them today.
0:30Enjoy the episode.
0:34>> [music]
0:39>> I begged them to come on. Remy Gaskill's
0:41on the pod. You've structured your
0:43company where you basically have these
0:45folders and dot MD files that run your
0:49company. And what I want to do today is
0:53I want you to teach people in a
0:54beginner-friendly fashion. This is only
0:56for beginners.
0:58How they could do the same thing. How
1:00they can set up their own executive
1:01assistant, head of marketing, chief
1:03financial officer. Basically, I want you
1:05to base to tell us the concepts behind
1:10all this. By the end of this episode,
1:12Remy, do you think you can do that?
1:14>> 100% Greg. We're going to go through all
1:17the concepts that make up an AI agent.
1:19And by the end of this video, you will
1:20know exactly how you can build up
1:23agents to run complete departments of
1:25your life and your company within any
1:28agent platform you choose, whether it's
1:29Claude
1:30Codec, Open Claw, Manas, all of them.
1:33>> All right, let's do it.
1:35>> Sweet. So, one of the reasons why I
1:37really wanted to make this episode is
1:39because I feel like the AI landscape is
1:42moving into like stage two, from chat to
1:46agents. And most people are getting left
1:49behind right now, just using the chat
1:51models. And
1:53the founders and employees that are
1:55utilizing agents
1:57are no word of a lie 10 to 20 times more
1:59productive in their day. And when you
2:01stack that up over days, weeks, years,
2:04you're going to just be miles ahead of
2:06the competition. So, I really want to
2:07make this episode today to help bring
2:09everyone up to where the AI landscape is
2:11at the moment and to start using agents
2:13to manage every department of your
2:15business.
2:17So, the key thing to understand here is
2:19chat models versus agents because the
2:22word agent is thrown around lots online.
2:24I'm sure you've seen it, Greg, like AI
2:26agents this, agents this, use this agent
2:29for this, and it's kind of lost a lot of
2:30meaning. So, I wanted to give start by
2:32giving a really clear definition of what
2:34an agent actually is.
2:36So, the way I think of it is a chat
2:38model is question to answer.
2:41But then an agent is goal to result. So,
2:44moving from just like uh you asking, AI
2:47replies, then you do the work to you
2:50giving the agent a task, it planning out
2:53the task and then executing and then
2:54delivering you a result. Does that make
2:57sense?
2:58>> Crystal clear. I mean, the way I think
2:59about it is chat is kind of like ping
3:01pong, back and forth, back and forth.
3:03>> Yeah.
3:04>> And agent is
3:07uh
3:08you know, you're giving it it's a goal.
3:09I mean, the best way to Yeah, you're
3:10giving it a goal and you're hoping that
3:12over time it gets better and closer to
3:14that goal.
3:15>> Exactly. Yeah, that's exactly it. And I
3:17just think that's a nice way to lay it
3:18out in your head is chat is question to
3:20answer, agent is goal to result. So,
3:23when you chat to an agent,
3:26you might give it a task like build me a
3:28website for XYZ, and then it goes away,
3:32it does its work, and it outputs this
3:34wonderful website to you.
3:36But it's really important to understand
3:37what's actually happening in this step
3:39here.
3:40So, inside this agent step, we have
3:42what's called the agent loop.
3:45So, you give it your prompt or task,
3:48and it goes through these three steps
3:50here, which is observe, think, and act.
3:54So, let's just say for example, um we're
3:56actually going to do this demo after
3:57this, but if we gave the agent a simple
3:59task like
4:01"Build me a minimalist portfolio site
4:03for Greg Eisenberg."
4:05It's going to start by like you've
4:07loaded in that prompt. It's going to
4:09check if there's any files in the
4:10workspace that it can work with, like
4:12maybe you've got some information on
4:13Greg Eisenberg.
4:14Um and then it's going to think about
4:16what to do next.
4:17It's going to act, and then it just
4:19keeps going through this loop. So, for
4:21that actual example of building the
4:23portfolio site for Greg Eisenberg, let's
4:25just say it was a blank agent. We hadn't
4:27given it any context. The first thing
4:30is it's received this prompt to build
4:32the website.
4:33And the first thing it's going to be
4:34thinking about is, okay, well, I need to
4:36build this website about Greg.
4:38Who the hell is Greg Eisenberg? So, it's
4:40going to then decide to do some research
4:43into Greg Eisenberg.
4:44It's going to research everything about
4:46Greg, and then feed it back into this
4:48observe step. So, then it's going to
4:50think to itself,
4:51"Okay, so I've got this prompt to build
4:53a website. I've now got my research
4:55here, so I know exactly who Greg
4:57Eisenberg is."
4:59And then it's going to start thinking,
5:00"What is the next step?"
5:02And the next step is probably to write
5:04up a plan to build the website. So, it
5:06might write up that plan, feed that back
5:08in. Now it's got the research, the
5:10prompt, the plan, and it will think,
5:12"All right, what next? I should probably
5:13write the code." It'll write the code,
5:16feed it back in, and it just keeps going
5:18through this loop as many times as it
5:20needs until it can conclude uh that the
5:23task is complete. And how it concludes
5:25that the task is complete is based on
5:27the parameters that you set in your
5:28prompt. So, you know, if you're giving
5:30it a research task, you might say
5:31compile 10 sources, and then create a
5:35report as a PowerPoint. And then once
5:37it's compiled 10 sources and build the
5:39report as a PowerPoint, it can conclude
5:42that the task is complete, and then give
5:44you the output as the user. The The
5:46agent itself
5:48is made up of these four components.
5:51So, it's the LLM, which is the brain
5:53behind it. So, think like, you know,
5:55Claude Opus 4.6 or GPT 5.4 or Gemini 3.
6:00It's the model.
6:02Uh it's got the loop, which means it
6:04just keeps going until the task is done.
6:05It doesn't stop after one response. So,
6:07you're going from ping-pong to like it
6:09continuing to go rather than you having
6:11to sit there baby-sitting it.
6:13Uh it connects in all your tools.
6:15And then it connects in all the context.
6:17And a platform that facilitates this
6:21process and basically facilitates this
6:23loop to happen is known as an agent
6:25harness.
6:27And all of the popular AI agent
6:30platforms on the market that you'd be
6:31familiar with are just agent harnesses.
6:34They're just applications
6:36where this loop is facilitated.
6:40And I want to actually run this little
6:42prompt I prepared earlier.
6:44I want to open up CodeX, Claude Code,
6:46and Anti-Gravity. And I'm going to show
6:48you this loop uh actually happening in
6:51action.
6:52So, I have a nicely prepared before the
6:54episode these three demo folders, which
6:56we're going to run in. So, I'm going to
6:57open up demo one
6:59to work in in Claude Code.
7:01And the way these folders work is if
7:03you've used If you're familiar with like
7:05any of the chat models like Claude and
7:07ChatGPT, there's a projects feature,
7:09which is where um if I open it up
7:11actually,
7:13try not to get dizzy with me switching
7:14tabs so much. But, you know, if we
7:16create a a project here,
7:19it contains all your chats in one place.
7:21It allows you to upload all your sources
7:23here, which is your context.
7:25And then you can even add custom
7:27instructions,
7:29which tells it how to behave within this
7:30project, and that's also known as a
7:32system prompt, which we're going to dive
7:34into how to do this with agents as well
7:36later. But, it's a similar concept that
7:38you'd be familiar with if you've used
7:39projects before. But, instead of the
7:41project being here on the cloud, we're
7:43actually working within projects that uh
7:46local on our computer.
7:48So, I've just selected this demo one for
7:49now.
7:51Then, we're going to run
7:52build a minimalist portfolio site for
7:55Greg Eisenberg.
7:56And then, this little bit here just
7:57tells it to actually spin it up like to
7:59publish it on the web
8:01uh in a preview mode so we can see what
8:03it's done.
8:04So, I'm going to run that.
8:05>> So, so this is this is um Claude Code?
8:10>> Yes. Yeah, right now we're in Claude
8:12Code, and this is just accessing it
8:13through the desktop app for Claude.
8:16Um so, I'm just going to run that.
8:19And then, I'm also going to give the
8:21same prompt to Codex here. So, this is
8:23the Codex app. And you can see same
8:25concept. It says, "Let's build." We can
8:28choose a folder on our computer to work
8:29in, like demo two.
8:32And then, we're going to give that a
8:33prompt as well. And we're going to tell
8:35it to host it on
8:37a different one.
8:39And then, also in Antigravity. So, you
8:42can see same concept. We're going in,
8:45selecting a folder.
8:47And then, we will
8:50give it the prompt as well.
8:52>> How should people think about security
8:54and these different products?
8:57>> I like to think of security as in just
8:59like
9:00scoping what they have access to. So, by
9:04default, Antigravity, Claude Code, and
9:07Codex, they're very, very secure because
9:08they're built by these massive companies
9:10that have a lot on the line um
9:12to protect.
9:13And
9:15I just you know, if you're if you're
9:17building out these agents to manage
9:18different elements of your business,
9:19like the other week I built one um that
9:22does manages meta ads. And obviously,
9:24that's quite a risky thing to give an
9:26agent control over managing ad budgets.
9:28So, it's just comes down to like what
9:30you feel comfortable giving the agent.
9:32And also, you can control what
9:33privileges or you can control what um
9:37tool permissions it has access to so
9:39that if it was compromised for whatever
9:41reason, the worst case like isn't that
9:42bad. And that means, you know, just
9:43giving it like read-only access to
9:46certain important platforms and stuff
9:48like that.
9:49Does that make sense?
9:50>> Yeah, totally. I mean, comparing it to
9:53like open claw, which is like
9:55>> Yeah, which I want to touch on at the
9:57end as well cuz that's the same thing,
9:58just another harness, but it's just like
10:00the wild west.
10:02>> Cool.
10:03>> Uh, and one thing like
10:06a nice little analogy to think about
10:07these harnesses is
10:10what we're going to learn today is we're
10:11going to learn to drive. So, we're going
10:13to learn about how to, you know, steer
10:14the car, like how the pedals, the brakes
10:17work, the accelerator works, the
10:18handbrake. But then once you know how to
10:21drive, you can kind of jump in any car
10:23whether it's like a old Toyota, a Range
10:25Rover, and you inherently sort of know
10:27what to do. And that just comes down to
10:29understanding all these key concepts
10:31that we're going to go through today.
10:32And you can think of the agent harnesses
10:34like different cars. And some of them
10:36will have better features like seat
10:37warmers and cruise control, but it's all
10:39once you know how to drive, you can
10:40pretty much jump in any of them and use
10:42them.
10:43So, we've just got our thing over here
10:45building to
10:46building the website for Greg. And it's
10:48going through this agent loop right now.
10:50So, you can see here it's actually
10:52decided that it's going to launch, um,
10:54an agent to go and research Greg
10:56Eisenberg.
10:57And I've connected it up to Perplexity,
10:59so it's now using Perplexity to research
11:01Greg. So, it's going through its first
11:03step of the loop. And I imagine that,
11:06uh,
11:07Codex
11:08has also done something similar here.
11:10You can see it's still working, but it's
11:12gone.
11:14And, um,
11:16started to build this out through the
11:18loop. I think Claude code does the best
11:20job of actually displaying that loop,
11:22um, and allowing you to see what it's
11:24thought about compared to Antigravity
11:26and Codex. But it's all just going
11:28through the same sort of loop process
11:30that I described earlier.
11:32And I think So, when
11:32>> you say that when you say you hooked it
11:34up to Perplexity
11:36didn't it's not like you asked it to
11:38hook it up, right? It just sort of did
11:41it.
11:42>> Yeah, because I've
11:44I've given Claude code Perplexity as a
11:47tool via MCP.
11:48>> Okay.
11:49>> Which we're going to get into
11:51um
11:51very very shortly all about MCPs, which
11:54is just connecting tools up.
11:56So we can see that in anti-gravity
11:59it's gone you can see this thinking
12:00process. It's gone um
12:04I'm now examining the current directory
12:05to figure out if there's an existing
12:07project or if I build one from scratch.
12:09It's then going
12:11I'm now going to start to build this
12:12thing.
12:14And then it's built the the website and
12:15it's given us a little localhost preview
12:17here. So it's created this nice little
12:19portfolio site for for you, Greg.
12:22>> What's interesting is like it's super
12:24minimalist and I mean
12:28it it did its job, right? Like
12:31this is
12:32I would totally launch something like
12:34this.
12:35>> [laughter]
12:36>> It actually looks really nice. Did it
12:37Did it scrape your email address
12:39correct?
12:39>> That's not my email address and I don't
12:41live in South Canada anymore. So but
12:43yeah, there's a few copy things, but
12:45other than that
12:46uh
12:46>> Yeah.
12:48It's done a pretty good job.
12:49>> It did.
12:50>> And if we go So that was anti-gravity.
12:53If we go into
12:55Codex as well, you can see here it's
12:56finished doing its website, which is
12:58somewhat similar.
12:59I think I prefer Gemini's.
13:01>> Yeah, agreed.
13:03>> If we check out Claude as well
13:05um
13:06it's still going. But you can see this
13:09loop, right? It's gone
13:10Okay, so first off, who is Greg
13:12Eisenberg? It's gone and researched
13:14Greg, then fed it back into that observe
13:16step and it's gone, all right, what
13:18next?
13:19Now I need to create the HTML file.
13:22So it's written the the code.
13:24And then now it's gone, okay, so he
13:26wanted it spun up on this local server.
13:28So, now I'm going to spin it up on the
13:30server.
13:31And then the last iteration of the loop
13:33is to check that it's actually done and
13:35can conclude the task is complete. It's
13:37opening it up and screenshotting the
13:39website and then reviewing the
13:41screenshots to check that the website is
13:43complete. And you can see here it's done
13:46another pretty good job. This one's very
13:48similar to the Gemini one, hey?
13:49>> It's true.
13:51>> Um but yeah, that's just like demoing
13:54how that loop is actually working in
13:56real time.
13:57>> Yeah. I mean, what comes to mind just by
13:59watching this is like
14:01how many people on the planet would
14:03benefit from a very clean website and
14:06like
14:06>> Yeah.
14:07>> how how do you set up these agents so
14:10that like
14:11you know, maybe it's like a cold email
14:14loop, right? Like you're sending cold
14:15emails, "Hey, I built you this website,
14:17so-and-so business. Do you want it? It's
14:19going to cost $250."
14:20>> Yeah. Yeah, that's actually a great
14:22idea. Um
14:23pre-making websites for companies. And
14:26it's like an off-the-shelf thing. It's
14:27like, "Hey, I I made you this website.
14:29If you want it like if you want to own
14:31it, it's $250." You can just do a mass
14:33cold email thing.
14:36Uh cool. So, I think that's like pretty
14:38much illustrated that agent loop
14:39example. So, I'm just going to go um
14:42back to our trusty board over here.
14:44But you can understand that it's just
14:46like all of these apps are just
14:48different flavors of the same thing.
14:50And then what we're going to be working
14:52up to today is my workspace looks
14:54something like this is I have, you know,
14:56a big like a a folder for each company
14:59or or client that I'm working in.
15:01And then I'll have folders underneath
15:03with all my heads of departments.
15:05And then
15:07um within those heads of departments,
15:08I'll have skills and MCPs, which we'll
15:10get into.
15:12And contacts.
15:13And then I've got like an overarching
15:15one at the top to just to sort of manage
15:17them all. But we're going to be focusing
15:18today on building out this executive
15:19assistant to take care of just your
15:21manual day-to-day tasks and free up at
15:24least 1 to 2 hours extra per day.
15:29Um cool. So, to build this out like we
15:33did uh with our demos, it's running off
15:36your local files. So, we're going to
15:37create a folder here
15:39called executive assistant.
15:43And also through building out this
15:44assistant, it's going to allow us to
15:47clearly explain each of the concepts um
15:49of building an agent in real time.
15:52And the way I like to think about
15:53building agents is onboarding them like
15:55a real employee.
15:57So, if you took on a real executive
15:59assistant, you couldn't expect to just
16:01for them to come into the office and you
16:03to give them a task without explaining
16:05your business first, your clients, what
16:07you do, the tools
16:09um because they just would not be a very
16:11good executive assistant.
16:12So, that's the first step that we need
16:14to go through uh when we're building out
16:16this agent. So, uh I'm actually going to
16:19work
16:21uh in Co-work at the beginning. So,
16:23Co-work is just another agent harness to
16:26do the pretty much the same thing as all
16:27the others, just that loop connecting in
16:29your tools and the context.
16:31So, you can see here um this was my
16:33little uh previous session where I was
16:35building some diagrams.
16:37But, we can go um and you can follow
16:39along in Cloud Code or Codex or
16:41Antigravity or whatever agent harness
16:43that you want to work in.
16:44But, I just think that Co-work has
16:46really nice simple UI for people to just
16:48understand really well what's actually
16:50going on.
16:51So, we're going to open up this
16:53executive assistant folder.
16:55And you can see here that if we ask it,
17:00"Write me a cold email."
17:03and send that off.
17:05>> So, peop- So, people are going to ask,
17:07"How How did you transcribe?" You did
17:09like a voice-to-text.
17:10>> Yeah, so that is um I use one called
17:13Monologue, but there's a lot out there
17:15on the market. Whisper flows another
17:17popular one, and it just allows you to
17:19hold a little button on your computer
17:20and
17:22just yap away, and it will just
17:23transcribe it neatly into text. And I
17:26find that
17:26>> good. My monologue looks really good.
17:29>> Yeah, I think it's built by the team at
17:30Every. Every. io.
17:32Um it's a cool product, but uh so
17:36what it's asking So it's it's straight
17:37away we we it's got no context here. So
17:40it's working out of
17:42uh that folder here on our computer,
17:44but there's nothing in the folder.
17:46>> Mhm.
17:46>> And it has no memory of our previous
17:48sessions.
17:49Um and it's asking like what like what
17:52do you even sell?
17:53Um and then we're going to kind of give
17:55it like who do you target? What tone do
17:57you want? These are all things that our
17:59executive assistant should know.
18:01Um so I'm just going to stop the
18:02response there.
18:04And one thing that's really important to
18:06know, which might be a bit of a shock
18:07moving from chat to agents, is that
18:09these agents' memory work a little bit
18:11different. So if you're used to using
18:14chat models like ChatGPT and Claude, if
18:16you open up a fresh session in the one
18:18of these chats,
18:19you don't give it any context, you don't
18:21upload any files, and you just say
18:23"Who am I and what do I do?" It's going
18:25to know a scary amount about you. And
18:28that's because with these chat models,
18:30they have memory built in automatically.
18:32So every time you sort of say things
18:33that are important, the chat model saves
18:35it to its memory in the cloud that you
18:37can't see and you can't control.
18:39And with agents, you have to set up
18:42memory and control exactly what you give
18:45it. And I think that's actually uh a
18:48benefit, not a limitation, because what
18:50happens is if you're using ChatGPT and
18:53it's got the auto memory, you're having
18:54conversations about three different
18:56companies, maybe you're asking for
18:59relationship advice, and then all of a
19:01sudden um when you ask it to write a
19:03landing page copy,
19:04it's pulling in context from all these
19:06other places that you don't really want
19:07in there.
19:09So with these agents, um
19:12you need to actually set up that context
19:14and memory. So, as you can see, when we
19:15asked it to write a cold email, it just
19:18had no idea about anything.
19:21So, we need to give it a context file.
19:25And the way you do this, right? So, you
19:26can see this example here. It doesn't
19:28know anything about us.
19:29And that's because we haven't populated
19:32what's called an agents.md file.
19:35And an agents.md file is just like a
19:37system prompt. Just like if you've
19:39created any custom GPTs before, you have
19:41that field for custom instructions. Or
19:43in the project, like I just showed
19:44before, you have that field for custom
19:46instructions. And it just gives it this
19:48context that's kind of always there,
19:50always on. And you put in there things
19:52like its role, context about you,
19:56um, your preferences for working.
19:58And then what happens is every new
20:00session, before it answers your your
20:03query or task, it loads in all this
20:05context to its brain as part of that
20:07observe step in the loop.
20:10So, I have pre-prepared
20:13pardon me.
20:16So, I've pre-prepared
20:18a, uh, agents.md file here. So, if we
20:22drag this in over here,
20:24this, uh, when you're working within
20:26Claude Code, it's called a Claude.md.
20:28When you're working within Gemini, it's
20:30called a Gemini.md. But when you're
20:32working in within Codex or Open Claw,
20:35it's an agents.md. But it's all the same
20:37concept.
20:38So, we can drag this into our folder
20:40here.
20:41And if we open up this file for a little
20:43preview,
20:46we can see here, I've got in here, um,
20:49all about me,
20:50what my business does, my working
20:52preferences, like the tools that I use
20:54and what for, like Notion project
20:55management, Stripe, um,
20:59we've got, you know, all the
21:01information. My ideal customer, it's
21:02loaded with context here.
21:05And I pre-prepared this, but if you want
21:07to make one of those, you can just use
21:08Claude chat or co-work whatever, and you
21:11can ask it to help you build out this
21:15agents.md file, and to just ask you
21:17interview-style questions to extract all
21:19the context from you, and then build the
21:20file.
21:21So, if I jump back in,
21:24now if I go to a new task,
21:27same folder,
21:28and we say,
21:30write me a cold email,
21:33it's going to have all that context.
21:35>> Yeah.
21:36That's what we hope.
21:38>> That's what we hope.
21:42There we go. And it knows, you can see
21:44these files over here. It knows
21:45automatically to load in this file if
21:48you title it correctly.
21:50>> Yeah.
21:52It's basically just like a a reminder
21:54file.
21:55>> Yeah, pretty much. It's just like
21:58loading it in so it has all this set
21:59context before you even start working.
22:02And one of the other big shifts to make,
22:03which comes with moving from chat to
22:05agents, is prompt engineering used to be
22:08the big thing. It was like, here's the
22:10ultimate prompt for going viral on
22:12social media, or use this prompt for
22:14this. And now it's all about context
22:17engineering. It's about how well can you
22:19load up your agent with all the
22:20information about your business,
22:22so that your prompt can be stupidly
22:24simple, like write me a cold email, and
22:26you're still going to get an amazing
22:27result. Um you can see already here,
22:30it's already asking like, um is it a
22:32brand or sponsor, potential partner, or
22:35consulting client. So, it's already got
22:36that context. Um book a call.
22:40Um you know, it's it's it's loaded in
22:42everything that we've given it from that
22:43agents.md file.
22:47And then now we've got a a pretty decent
22:50cold email there, ready to go.
22:55So, that's basically agents.md files for
22:58you. And you want to create one of those
22:59to onboard your agent with all the
23:01context it needs. And if you have lots
23:03of context,
23:04without getting into too many advanced
23:06concepts here, sometimes what I will do
23:09is I will create like a um
23:12a folder called context.
23:15Load that in. And in here it's got
23:17different files about me, brand voice,
23:19ideal customer profile, etc., etc.
23:23And then in order to keep this smaller,
23:25I will then just say in this Claude.md
23:27file,
23:28um before answering any questions or
23:30before doing any tasks, read my context
23:32folder to understand about myself and my
23:35business.
23:36Because by default, if you just have
23:38this context file in here but no
23:39Claude.md, it won't load all that into
23:42the session by default. But if you tell
23:44it in this file that it always loads in
23:46to then check this file, you can start
23:48to like string all your context
23:50together.
23:53And a lot of people have done that with
23:54Obsidian. So, they'll have like in the
23:56Claude.md file, they'll tell it to go
23:58check their Obsidian vault for their
24:00second brain to go and find context.
24:02>> Mhm.
24:03>> So, that is agents.md files explained.
24:05So, that's how you actually when you're
24:06onboarding your agent, like our
24:08executive assistant, you can train it up
24:11on who you are and your business.
24:13And then as you can see here, you know,
24:15I've got folders for all these different
24:16roles in my business. And in the head of
24:18marketing, that Claude.md file would
24:20look somewhat similar, but in the top it
24:22would say instead like you are my head
24:24of marketing. You speak like this. These
24:27are your tasks. These are your roles.
24:29And then the second thing here is about
24:31memory
24:33and the self-improving loop.
24:34So, we've solved the problem now. Um try
24:37not to get too dizzy with me switching
24:39tabs.
24:40But we've solved the problem of our
24:43executive assistant not knowing anything
24:44about us or our business.
24:46But now we have a new problem, which is
24:49it doesn't really remember
24:51the intricate details or your
24:52preferences across sessions unless
24:54you're manually going and updating that
24:56Claude.md file.
24:58So, you can see here if we go
25:00um
25:02my favorite color is lavender.
25:06It'll probably say something like got
25:08it, noted.
25:09>> Yeah, that makes sense, right? Cuz it's
25:12and it's it's adding Where Where's that
25:14adding it?
25:16>> Well, it's not adding it, that's the
25:17thing. So, we can tell it my favorite
25:19color's lavender.
25:20And it's gone. The user's just shared
25:22you That's that thinking step. It's like
25:24the user's just shared this
25:26>> No like no nothing needed. Good to know,
25:29I'll keep that in mind. But then if we
25:31go into a new session, same folder, and
25:33we go
25:35what is my
25:38fav color?
25:41Mind my spelling.
25:43It's going to say I've no idea what your
25:44favorite color is, even though we just
25:46told it.
25:47And that um is an issue, you know, cuz
25:50if you're working, you know, in uh
25:52you've got like a head of sales or
25:53something and it keeps it signs off your
25:55emails wrong.
25:56And you tell it you correct it, you say
25:58never sign off emails with cheers, say
26:00warm regards.
26:02And it will go Okay, got it, noted. But
26:04then the next day you start working and
26:06it does the same thing again. It's like,
26:07what like my agent's broken. But really
26:10it's not it's running off those context
26:12files in the back and unless you are
26:14manually updating it, it won't know to
26:17save that preference.
26:18So, what I like to do is I like to add
26:21in
26:23something like this to my agents.md
26:27file.
26:28So, this is just a little simple thing.
26:29You can pause the video and copy it.
26:35But I like to
26:37I'm just going to remove that context
26:38file for now. That was just to
26:39illustrate that example of adding more,
26:41but we're just working with this one
26:43file for now. So, I'm just going to open
26:44this up so I can edit it.
26:47And I will quite often add something on
26:49the bottom like that little snippet.
26:52And this basically just says
26:55Actually, you know what? I might just
26:56add it at the top.
26:58Just so it's there.
26:59Top of mind for my agent, cuz I think
27:02this is really important.
27:05So, you can see I've just added this in
27:06and it just says
27:07read all files in context.
27:10Read memory.md. This is what you've
27:12learned over time.
27:14And then when I correct you or you learn
27:15something new, update the relevant
27:17section in memory.md. And it's just got
27:20a couple little things here. And it just
27:22says keep memory.md current. When
27:24something changes, update it in place
27:26and replace outdated info.
27:29So, we can do command S to save that.
27:32And then I'm going to add another file
27:34here.
27:35We can actually just duplicate this.
27:39And this one
27:41I'm going to call
27:43memory.md.
27:46And then we can open up this one.
27:47And I'm just going to remove all of this
27:49context here.
27:51Um except I'm just going to keep those
27:53sections.
27:55>> So, memory.md
27:57is basically I mean, it's just
28:01what it sounds like, right? It's
28:02basically like you want to
28:04>> [sighs]
28:05>> you know, if the goal is to build, you
28:07know, AI employees that do things for
28:08us, they're going to need to
28:10need to remember our preferences, right?
28:12A good employee remembers preferences.
28:14>> Exactly.
28:15>> And learns over time and that that
28:17compounds. So, memory.md
28:20is just a place that you can just make
28:22sure that
28:23uh over time it you know, your whatever
28:27you're using co-work or whatever, it
28:29ends up it ends up getting compounded,
28:31getting smarter. So, ultimately
28:33you might be trying things like co-work
28:36and and you're you're not getting good
28:37results. And a big part of that is you
28:40don't have uh
28:41a clo.md and memory.md sort of
28:44>> Yeah.
28:44>> set up.
28:45>> Exactly. Exactly. And now the thing is
28:47some of these agent harnesses
28:50have started to add in this memory
28:52system that we're doing manually,
28:54telling it to update. Some of them have
28:56have got that built in automatically,
28:57like Open Claw, and I believe like Manas
29:00and some of the others have that built
29:01in automatically. But it's still
29:03important to understand because it's
29:04just doing the same thing under the
29:05hood, except they've just set this up
29:07for you.
29:09Um so, we've got this here now. We've
29:11got our memory.md, our claw.md.
29:14And then now, if we go back into
29:18Cawork,
29:19if we do a new session
29:22in that same folder,
29:23and we say,
29:25"My
29:26favorite color is lavender."
29:31>> Better remember.
29:34>> For the sake of the demo, I hope that it
29:36that it does what it's told.
29:37>> It's going to remember it.
29:40>> You see here?
29:42Perfect. It's gone good. I'll remember
29:44that. Let me save it to memory. And now
29:46you've got this big memory file that
29:47builds up over time. And whether For
29:50example, this is your executive
29:51assistant, so it might be saving
29:52preferences like how to sign off emails,
29:55or don't connect with clients on Slack.
29:58I always want to keep client comms on
30:00email. But if you're building out like a
30:02head of marketing, it might be
30:04preferences about how you like your ads
30:06structured in Facebook Manager. If
30:08you're building If you've got a folder
30:09where you're working on a website or an
30:10app, it might be things like don't use
30:13dark mode, and then it will update so
30:15it'll never use dark mode again. And
30:17these just compound over time. So, as
30:19you start to build up these rules,
30:21the amount of errors go down. And And
30:23this just compounds and compounds over
30:25weeks and months.
30:27>> Remy, have you seen some of these
30:29memory.md files get so big that at a
30:32certain point it's just ineffective?
30:34>> Great question. I personally haven't
30:37happen have that happen to me yet. I
30:38haven't hit that threshold. Um but a
30:41best practice for those Claude MD files
30:44is to keep it around like no more than
30:46200 lines.
30:48And yeah, I could imagine if you started
30:50to build this up over years and years,
30:52you'd eventually hit a point where all
30:53the the rules are stepping on each
30:55other's toes.
30:56And you know, you could probably go
30:57through and do a bit of like a manual
30:59clear. But, I haven't hit that threshold
31:00yet.
31:01>> Cool. So, people don't need to worry
31:03about
31:04cluttering their memory MD.
31:08>> I wouldn't worry too much. I mean, if
31:09it's saving like this like the silliest
31:11little things, like the tiniest
31:13corrections, you could maybe update that
31:15Claude MD to say only save like
31:17substantial corrections, you know? And
31:19then you can have a bit more control
31:20about what it's saving.
31:21So, that's probably a yeah, what I would
31:23do there. But, once you've set this up,
31:26now when you say something like quit
31:27writing so formally, it's going to
31:30do the task, then update its agents
31:33or in this case Claude MD to keep tone
31:36casual, never formal. And then now in
31:38any new sessions, it's going to keep
31:40that preference over time, which is
31:41pretty cool.
31:43So, now we've got our executive
31:45assistant set up with
31:47memory and we've given him his role. We
31:50now need to connect our tools cuz by
31:52default, most of these agent harnesses,
31:53they just
31:55have web search baked in. But, if you
31:57want to actually start linking it up to
32:00your tools like Gmail, calendar, and
32:02everything else, which is where the real
32:03productivity gains are made, you need to
32:05do so via what's called MCP.
32:09And I actually got Greg, I got this MCP
32:11explanation from when you had on
32:13Is it Ross Mac?
32:14>> Ross Mac. Yeah.
32:15>> Yeah. So, this he did a great
32:17explanation and it just dropped into my
32:18head really nicely. And it's basically
32:20that
32:21before MCPs,
32:23your agent or your LLM, in order to
32:26speak to tools, it had to kind of learn
32:27their language cuz Claude speaks
32:29English, Notion speaks Spanish, Gmail
32:31French, your browser speaks Japanese,
32:34and Slack speaks Chinese. And it was
32:36capable of connecting to those tools,
32:38but it like required these extensive
32:40custom developments that took a long
32:42time.
32:43But then
32:44Anthropic actually created MCP, is that
32:46right?
32:48>> Yep, it's right. Yeah.
32:50>> Anthropic built MCP to basically sit as
32:52this translator
32:54in between your tools so that Claude can
32:56still just speak English and your tools
32:58can just speak their languages and this
33:01MCP speaks every language and then just
33:03translates your calls from your
33:06agent to the tool and then from the tool
33:09back to your agent. So just set a really
33:11easy standardized way to connect tools
33:13up.
33:14And that's what we're going to be using
33:15to connect all of the tools to our
33:17executive assistant.
33:19So if we go back into co-work here, you
33:22can see that Claude make it really,
33:24really easy to connect up your tools.
33:26You can just go to connectors, browse
33:28connectors and they've got like hundreds
33:29of all the like biggest apps that you
33:31probably use.
33:32And you can just, you know, add them,
33:33sign in, pretty self-explanatory.
33:36But I believe
33:37Codex would be the exact same. You know,
33:39you can go
33:41skills or
33:44If we go settings, they probably have
33:46like uh
33:47And then like Manas is the same. For
33:49example, if you go into Manas, we can
33:51see
33:52we can go and connect our tools.
33:55Very, very simple process and then same
33:57with Perplexity computer. You know, you
33:58got your connectors
34:00and you can connect all your tools in
34:01here. It's just all using that
34:03model context protocol, MCP.
34:07So I've already, before the episode,
34:09gone and connected all of the tools that
34:11I use most, like Gmail, Google Calendar,
34:13Granola, Notion. They're all set up
34:15already as MCPs.
34:17And what I'm going to do now is I'm
34:18actually going to
34:20open up this executive assistant folder
34:22in Claude code to sort of demonstrate
34:24how these harnesses are all the same and
34:26they work off your local files and the
34:29real future-proof AI stack is just
34:31having those markdown files on your
34:33computer.
34:34And the reason why I like to work in
34:36markdown files is because it's just the
34:38easiest sort of format
34:40for your LLM, for your agent to actually
34:43digest and understand compared to if you
34:45were to give it your files as like a
34:47docs or a PDF file.
34:49So, I like to use Claude Code within
34:52Visual Studio Code.
34:54Um so, you can see here, I'm just going
34:56to It looks very similar to
34:57Anti-Gravity.
34:58I'm just going to open up our executive
35:00assistant folder here.
35:02And the way that I see the future of
35:04this all going, Greg, is I think that
35:06everyone's going to have their what I
35:07call an AI OS, like an operating system.
35:10And this will just compound over time,
35:12like you saw with adding the rules and
35:14getting less errors.
35:15But, with adding your tools and then
35:17skills, which we'll get into, which is
35:18basically just training AI on your
35:20processes.
35:22And I think that everyone's going to
35:23have like an AI operating system they
35:25work in. And everyone will just have
35:27personal agents and agents to manage
35:29each department of their company.
35:30And people won't actually use these apps
35:32anymore. Like, I've connected up Gmail,
35:36uh Google Drive, calendar, Granola for
35:38my meeting notes, Stripe for payments,
35:40Notion for project management. And I
35:42don't even enter these tools anymore. I
35:45just sit in Claude Code as one central
35:46place. And an example here is I sent
35:49myself before the episode, I sent myself
35:51an email from a fake prospect.
35:54And I also entered in Granola a fake
35:57meeting with this prospect.
35:59So, now I can say things like um
36:03summarize my inbox from today.
36:07>> Um you know,
36:08so someone might ask like
36:10well, how important is that really, you
36:13know, like
36:15is that such a high value task? Like our
36:18What are high value tasks that you're
36:19actually getting done here?
36:21>> [sighs and gasps]
36:22>> So,
36:24>> Or maybe Or Or maybe you get a lot of
36:27emails, you know?
36:28>> Well, if you know, emails is a big thing
36:30if you do get a lot of emails, but
36:33just having like all those tools
36:34connected in one place and not having to
36:35switch and and copy-paste context. So,
36:38you'll see an example here, right? So,
36:39we've got summarize my inbox from today,
36:41which is like one of the most basic
36:43agent tasks ever. But, we can see this
36:45is one I sent earlier. We've got this
36:46one email here like our call today,
36:48excited after your call wants next
36:50steps.
36:51So, I might just say here, um
36:54Okay, great. I review my meeting notes
36:57with Maltoshi from today and then draft
37:01up the email sending the proposal and
37:04creating the Stripe payment link and
37:05then go into Notion and set up the
37:08project.
37:09And where this starts to compound even
37:11more is when you start to build out
37:13skills for each of your processes,
37:15because every time I do a process, even
37:17like this, manually prompting it, um
37:20and I know I'm going to do it again at
37:21some point, I'll then just turn that
37:23into a skill.
37:24Uh and then you eventually end up if you
37:27automate like three to five tiny manual
37:29processes each week with skills, you
37:31eventually end up
37:32um automating like your entire life with
37:34these agents.
37:35>> Right. So, it's it's not so much in like
37:37summarize my my emails where it's super
37:41super valuable. It's like
37:43that's where the starting point is and
37:45then we want to like manipulate it and
37:47use it and go deeper and stuff like
37:49that. That's when
37:50>> Exactly.
37:52>> these tools really really are valuable.
37:54>> And you can see here it's now connecting
37:56all my tools. So, it's going into
37:58Grainola and found the the full meeting
38:00of what we went through today.
38:02It's now going into Stripe to create the
38:05product link. Um
38:06>> Yep.
38:07>> And then it's going into Notion to set
38:09up the project.
38:11And then it will it should create the
38:13draft ready for us to go to send out.
38:16>> That this is really like a new way of
38:18working.
38:19Right?
38:20>> Yeah, it is. It is. It it it it it's so
38:22new and I and I even this task, it's
38:25really simple, right? Just sending an
38:26email based on a call with a proposal
38:28link and stuff. But like even if you can
38:31just do something like seven times
38:34faster without having to go into all
38:36these tools, copy the meeting notes into
38:37the
38:38page to give it context on your meeting.
38:41It really starts to compound and you
38:42start to fit like a week in a day and
38:44then seven weeks in a week. Um and you
38:48stack that up over a year and you're
38:49going to be miles ahead of everyone
38:50else. Uh and when we get into skills,
38:52you're going to see how this continues
38:53to get even better.
38:55But um you can see here
38:56it's drafted the email.
38:59It's pulled in all these insights from
39:02our call in Granola, which is like where
39:04I do my meeting notes.
39:06And then it's created the Stripe payment
39:08link.
39:10Here, ready to go.
39:11>> That's cool.
39:12>> And then now I can just go
39:14um
39:15send this email.
39:17And it will use my Gmail integration to
39:18to go and send it.
39:20Uh and then
39:21>> really cool.
39:23>> It is, hey? I think this is the new way
39:25of working. And Cody Schneider, who you
39:26had on the pod the other week, I saw a
39:28tweet from him and he said
39:30that in the future
39:31everyone's going to have like an AI
39:33operating system like this and you're
39:34going to have like the 100X employee
39:36because everyone will come into their
39:38role with a pre-existing AI operating
39:41system and then build out skills for all
39:44their manual processes, similar to how I
39:45was describing, and just keep building
39:47skills each week for anything manual
39:49that comes up until eventually their
39:50entire
39:51life and work life is automated.
39:54Great. So, you can see here it's now
39:55created the draft here in Gmail ready
39:57for us to to go. And if we're happy with
39:59it in the platform, I could also just
40:01ask Claude to send it there and then.
40:04Uh but then what like also gets really
40:06cool is
40:08I'm going to demonstrate now how I
40:09actually build out skills for these
40:11processes. So,
40:12I know I've talked a lot about skills so
40:14far. I want to just give a little
40:15overview on what skills actually are.
40:18>> Yep.
40:18>> So, the easiest way to think about
40:20skills is SOPs for AI. So, standing
40:23operator
40:24standing operated Oh my god. Standard
40:27operating procedures for AI.
40:30>> The most
40:32>> [laughter]
40:32>> It means once you explain something
40:34once, you never have to explain it ever
40:35again.
40:37An example of this is without skills, if
40:39you are creating a proposal for a client
40:42and you're sitting in in in your cloud
40:43chat or whatever agent harness you're
40:44using
40:45and you ask it to create this proposal,
40:48you're probably going to go back and
40:49forth a bunch of times. Remove change
40:51the formatting here. Use this color blue
40:54for this part. Put the price at the
40:56bottom instead of at the top. And
40:58eventually, maybe after 15 minutes, half
41:00an hour, you land on a proposal that
41:02you're really happy with.
41:04And you send it. And the next week, you
41:07want another proposal written, but
41:08unless you're going and finding the same
41:10session and working in that same
41:11session, it's going to have completely
41:13forgotten all of these preferences. And
41:16even if you have that memory system set
41:18up, these kind of things you don't
41:20really want clogging up your memory.
41:22They're better off as skills, which is
41:24basically it packages up that process
41:27into a dot skill file.
41:29And in that dot skill file, it's
41:31basically just a markdown file that
41:33explains the exact process that you went
41:35through. So, you could create a proposal
41:37skill.
41:38And then every time you now you need a
41:40proposal written, it just
41:43takes that skill, knows exactly what to
41:45do, and then you can have that proposal
41:47the same way every single time.
41:50>> So, is a skill like a memory file? Like,
41:53what's the difference between
41:54essentially a memory.md
41:57and a skill? Like, is it just is it is
41:59it is it is it just like a memory.md
42:02file for particular job to be done?
42:06>> pretty much like exactly it. And all of
42:08these agent harnesses pretty much now
42:10have skills as a feature.
42:12So, you can see here like if we go into
42:16Codex, for example, they've got skills
42:18here.
42:19Um same with Claude as well.
42:22And these when you're working with these
42:24agent harnesses that operate like mostly
42:26locally off your computer,
42:28um you can see here
42:30it actually operates out of this hidden
42:33file called a dot Claude folder.
42:35Skills, and these are all of the skills
42:38that I've created. There's tons.
42:40And if we open [clears throat] one up,
42:41for example,
42:42um like this one here, let's find a good
42:44one.
42:46For example, I've got this one here for
42:47writing viral hooks.
42:50And in this skill, we have a dot skill
42:53file, which is basically like your
42:55memory.md, which explains the exact
42:57process for writing viral hooks.
43:01And then it's also got packaged in here
43:03some references like hook formulas.
43:06>> Okay, so wait. So,
43:08how did you create that skill?
43:11>> Okay, so there's two ways that I find
43:14useful to create skills is one you can
43:16have an idea of a skill you want to
43:18create off the bat. So, like viral hooks
43:20for example, I had this course on viral
43:23hooks, which I transcribed, put it into
43:25Claude, and Claude has this by default.
43:27It has a skill creator skill added into
43:30it. Same with all of the major agent
43:31harnesses, they'll have a skill creator
43:33skill. So, it's kind of like
43:34skill-ception. You use the skill creator
43:36skill, and you say, "Hey, take this
43:39course on viral hooks and create a viral
43:41hook skill." And it can create it like
43:43that. That's one way.
43:45Uh and then it will package it up nicely
43:47with that skill.md. It'll do the whole
43:48thing for you.
43:49>> Wait, you you you asked you asked it to
43:52take the course?
43:55>> Yeah, I I uploaded the course like the
43:57full transcript of the course.
43:59>> Literally.
44:00>> And I said uh
44:02yeah, based on this course on viral
44:03hooks, build me a viral hook skill.
44:07Uh and then I use that for my content
44:09team. So, just like we're building the
44:11executive assistant folder, I've got a
44:12folder called content team, and I've got
44:14that
44:15>> Yeah.
44:15>> that uses that skill for me.
44:17And the second way to create skills is
44:20going through a process manually once
44:22with Claude.
44:23And then if you know you're going to
44:24have to do it again, like that proposal
44:26example,
44:27you can just say once you've done the
44:28task, "Hey, create a skill for what we
44:31just did." And it will package up that
44:32process you went through. And that's the
44:34second main way that you can create
44:35skills.
44:37>> So, in your viral hook example, if you
44:39go into that folder again,
44:41>> Yeah.
44:41>> so you have a references folder.
44:44>> Yeah.
44:45>> So, that is probably
44:49like was that Yeah, let's open it. I'm
44:51just curious.
44:53>> So, here it's got a full thing about
44:56like
44:56>> So, was this from the course?
44:57>> he is a bit screwed. Yeah, this was
44:59basically from like a a
45:00course I put into it.
45:02>> And did you ask it to create a
45:04references folder? Like how should
45:05people think about
45:07>> No, it just did it. It just did it. So,
45:09I think what would be great is if we
45:10could actually um, demonstrate building
45:13a skill live.
45:14>> Let's do it.
45:15>> Um, and so, for example, like this
45:18process here,
45:20um, I might I could create a skill
45:22called like a daily brief skill, you
45:24know, that goes through and summarizes
45:26like your calendar, your inbox, and your
45:29projects in Notion, and plans out your
45:31day for you in the morning. And then you
45:33can run that on a scheduled task,
45:35because
45:36uh, a lot of these agent harnesses now
45:37are starting to introduce scheduled
45:38tasks.
45:40So, you can just run it on 9:00 a.m.
45:41every morning, "Use my daily brief skill
45:44to prepare me for my day."
45:45But I think another cool one here, just
45:47to show you an example, right, of how
45:50my new like how intricate I make these
45:52skills, is let's just say for this
45:55fictional meeting I had with this
45:56person,
45:58I might say,
45:59um,
46:02"Can you draft up an email? I want to
46:04refer Maltoshi to my good friend
46:07Sebastian, who has an AI automation
46:09agency
46:11and can help them out better with their
46:12needs.
46:17And then we can just go
46:19um
46:20Sebastian's
46:22email is
46:25And we can just say that, right? And
46:27then now it's going to be able to
46:30take the notes from Granola, all the
46:32contacts, and then draft an email
46:34connecting these two,
46:36um a prospect with a friend. And you
46:38know, I have different like referral
46:39things set up like that with people in
46:41marketing agencies. And that's just like
46:43a little manual process there. Tiny. It
46:45maybe takes 15 minutes out of my day.
46:47But then I can just go
46:50I want you to use your skill creator
46:52skill
46:53and create a Sebastian refer skill so
46:57that whenever I ask you to refer someone
46:58to Sebastian
47:00you know exactly what to do and you know
47:03his email address.
47:10And then that'll build out that tiny
47:11skill for us, tiny process. But it means
47:14I like I know in the future I'm going to
47:16have to refer someone to Seb again. And
47:18even if this skill now saves me 15
47:20minutes another five or six times
47:23they start to compound when you create
47:25skills for every single little process
47:26in your business.
47:28>> Yeah, I guess it's like
47:31we should just be asking ourselves like,
47:33you know, in our day-to-day life, like
47:35what are all the jobs to be done?
47:37>> Yeah.
47:37>> are all skills that we need? Like what
47:39are the
47:40repetitive processes or SOPs as you
47:43>> Yeah.
47:44>> talked about. And then just
47:46setting up as many as possible.
47:48Right? To make our lives easier.
47:50>> Exactly. And just to give you a little
47:52demo here. So
47:54I sort of I to that folder structure at
47:57the start of the video. And this is it
47:59here. So we've got workspaces AI with
48:01Remy.
48:02And for example, I can open up my
48:03content team.
48:06And within this folder
48:08uh this is just like a more elaborate
48:10version of our executive assistant, but
48:12we've got our Claude.md in here, which
48:14explains
48:16um you are like the main orchestrator,
48:19you have these sub-agents. It's just a
48:22more elaborate version of that
48:23Claude.md.
48:25But I've got a skill within this for
48:28like a meta ads analysis. So that was a
48:30process, for example, if you're a
48:32marketing agency owner, this is probably
48:34like the kind of stuff that you can get
48:35inspiration from.
48:36Um like ads analyzing, you know, taking
48:38competitors ads libraries, breaking down
48:40all the creatives um and their landing
48:43pages. So I built out this ads analyst
48:46skill, where I literally just do ads
48:48analyst, and then I paste in like the
48:51ads library URL like that.
48:54And I'll click run, and I did an example
48:56yesterday with the Udi, which is a super
48:59large e-com brand, and it ran through
49:02and basically scraped all of It took
49:05screenshots of all the landing pages.
49:07It went and scraped all of the ads that
49:09they're running, all like 220.
49:11It then did a full deep dive here on all
49:14the ads.
49:16Visual analysis, copy analysis, why did
49:18this work, what could be improved. It
49:20basically did a breakdown of all the
49:21landing pages with screenshots.
49:24>> Hmm.
49:25>> And it did a master report here about
49:27everything that's going on. So it did it
49:28just did a full breakdown. And that was
49:30like a manual process that I would have
49:31gone through when I used to run like my
49:33marketing agency, and that probably
49:35would have taken me like
49:36three or four hours.
49:38And then I went through to build out
49:40this skill, I went through the process
49:42once with Claude. Like I started a fresh
49:43session, and I was like, all right, go
49:45to this ads library URL, scrape this, do
49:48this, do this, do this for for two
49:50hours. And then after I've done the
49:52entire process
49:54I just said use your skill creator skill
49:57to make a skill
49:58for ads analyzing and package up the
50:01entire process we just went through as a
50:03skill.
50:04And then now whenever I want to do that
50:06process again I can just
50:07invoke the skill and it and it knows
50:09what to do.
50:10Which is pretty cool.
50:11>> Crazy. Crazy.
50:13>> Absolutely crazy. So if So now the the
50:16refer Sebastian skill is live
50:18and now whenever I want to refer someone
50:21to Sebastian again I can just say um
50:23yeah refer to Sebastian and it will just
50:25start to use it will use that skill.
50:27That's example there tiniest process but
50:30you build like those up for all these
50:32little tasks you do day to day
50:34and then
50:35it just compounds and compounds and
50:36compounds. And I can already think of an
50:38idea here where you could then you can
50:41chain skills together.
50:43So you could example have like a a
50:45meeting prep skill
50:47that prepares you for a meeting by
50:49researching the guest
50:51um and compiling some talking points.
50:53You might have like a podcast research
50:55skill for example Greg for a guest
50:56that's coming on.
50:58And
50:59you might also create a a morning brief
51:01skill.
51:02And in the morning brief skill you can
51:04say if there's any meetings
51:07uh coming up or podcasts in my day use
51:09the podcast research skill
51:12um to research the the the guest. And
51:15you can like chain them together
51:17uh and build like some really really
51:18cool workflows.
51:21>> Yeah and you can have it so it sends you
51:23an email right?
51:25>> Yeah exactly. And then now these
51:27harnesses are starting to get more and
51:28more autonomous. Like they're starting
51:29to add like you know in the car they're
51:30starting to add like cruise control and
51:32stuff. Like now within most of these
51:33harnesses you can schedule tasks
51:37um like in co-work or cloud code now and
51:40you can like for example this one here
51:42you can go new task and I could say like
51:45uh
51:46run my morning briefing skill.
51:51And then set that to go every every
51:53morning at 9:00 a.m. And then now it's
51:55like an automated workflow that like
51:58you've just got running every morning
51:59now, which is pretty cool.
52:00>> Yeah, I'm I'm doing this right now. Like
52:03for example,
52:04I'm I'm buying a new car right now and
52:07it's like a particularly unique
52:10like color that I want and feature set
52:13and there's just none none really
52:15available, so you know, every 3 hours I
52:18I have I'm scraping all the different
52:21car marketplaces
52:23>> That's great.
52:23>> and
52:24and then I'm getting a notification that
52:26you know, it when something comes up and
52:28it's
52:29it's crazy, right? Like it saves me
52:32I'm one of those people that like
52:35if I didn't have this, I would be
52:36spending an hour of my day just like
52:39checking
52:40religiously every single, you know,
52:42CarMax and cars.com and Autotrader and
52:46all these websites and refreshing like a
52:49insane person.
52:50Um so yeah, the schedule to not
52:53>> great great example there. But you know,
52:55this like this is a skill that would be
52:57relevant for my executive assistant.
53:00Same with that car one. That could be a
53:01good executive assistant skill. But then
53:03I've got those more elaborate skills
53:05built out for like, you know, my um
53:07content team that ads library scraping
53:09one. And then um I've got like, you
53:12know, re- a research weekly research
53:14skill for my newsletter team. And that
53:16runs on a schedule every Thursday
53:18morning to go and scrape like I've built
53:20the skill out so it goes and scrapes
53:21Twitter and Reddit to find what's new in
53:23AI.
53:25Um but yeah, skills are so so powerful.
53:28Um combine them with like your MCP so it
53:30can use your tools and then you can
53:32start to just train up your agent on all
53:33the processes in your business. And I
53:36did a um a build out on Open Claw for a
53:41agent to manage meta ads, and it went
53:43pretty viral.
53:44And
53:45the the way I built this was with all
53:47these key concepts. So, Open Claw
53:49functions the exact same way. So, I I
53:53hope it hasn't timed out, but I've
53:54remote accessed into my Open Claw
53:56dashboard here.
53:57And you can see
53:59it's just operating off an agents.md
54:01file in the backend. But, instead of the
54:03dot claw folder, it's in a dot open claw
54:05folder. And then it's got a couple of
54:06these other ones here. It's got a
54:07memory.md.
54:09It's got some of these other ones that
54:11it's added on, like a soul which tells
54:13it its personality, and an identity
54:15which tells it who it is.
54:17But, it's that same concept of markdown
54:20context files
54:22connecting your tools, and then creating
54:25skills. So, that
54:27uh meta ads manager one that went pretty
54:28viral,
54:29I just planned it out with Claude. I was
54:30like, "I want to have this Open Claw
54:31manage my meta ads.
54:33Help me write the agents.md file to tell
54:36it you are my meta ads media buyer. You
54:38do these processes." And then I created
54:41skills. So, I created a
54:44ad creative skill.
54:46Um so, it knew to go look in the Dropbox
54:49folder and create creatives.
54:51I created a copywriting skill, so it
54:54knew how to write good copy for the
54:55business.
54:56Um
54:58and I just built out there's probably
54:59maybe 15 different skills. And then I
55:01would combine scheduled tasks, which is
55:03cron jobs, with skills, and the context
55:07files, and then just give it all the
55:09tools it needed.
55:10Um and just following the same process
55:13as we just went through to build the
55:14executive assistant, I had an Open Claw
55:16meta ads media buyer, which was sick.
55:19>> I love it. And for the beginner, are you
55:21like would you recommend people, you
55:23know, use Open Claw, or should they be
55:25using Co-work or Manas and some of the
55:27ones you showed?
55:29>> So, great question.
55:31Uh
55:32I would say that Open Claw is probably
55:34like one of the hardest to learn and set
55:36up of these harnesses. I would say
55:39Claude Code is probably the easiest.
55:41I think Perplexity Compute, you did a
55:42video on it. Um it's pretty simple, too.
55:44Same with Mana.
55:45>> Easy.
55:46>> Um but I would definitely learn
55:50um and get comfortable using like Claude
55:52Code or um one of these other ones
55:54before I started to play around with
55:56Open Claude. And I would also
55:58uh have all the processes built out in
56:00Claude Code first. So, for example, that
56:02executive assistant, over the next
56:05uh 2 weeks, I might build out a bunch of
56:07skills, like the Sebastian refers skill,
56:10um like a daily brief, meeting prep,
56:13etc. etc. And then once I'm happy with
56:15how it's all functioning in Claude Code,
56:17then I could look to migrate that into
56:19Open Claude, where it has that more
56:21autonomous nature to it.
56:23So, that's kind of how I think about
56:24using Open Claude and those other
56:26harnesses.
56:27>> Yeah.
56:28Cool. All right, anything else you
56:30wanted to
56:31>> Um
56:31>> cover?
56:32>> I mean, like really, there's no right or
56:34wrong way to run these. Like that was
56:35the executive assistant.
56:37I've got one built out for all the other
56:39departments in my business, and then
56:40other businesses I work on, I have the
56:41same. Um
56:43and you can just kind of build out that
56:45structure with what works for you.
56:47Um you've got like one other thing to
56:50mention is global versus project level,
56:52which I'll just go over super quick. So,
56:54like those skills, for example,
56:56um you can add them at a global level,
56:58which means they apply to every single
57:01project you work in, whether it's the
57:02executive assistant, your head of
57:03marketing. And some skills you want
57:06globally cuz you might use them in every
57:08chat. Um like
57:11a
57:12um truncate skill that I created, which
57:15just makes whenever I want to make
57:17something shorter, it makes it shorter
57:20without compressing the sentences, but
57:21just removing sentences that don't need
57:23to be there.
57:24And that's something I want in every
57:25session. So, I've got that in global.
57:27But you can have project level skills,
57:29like that um
57:31Sebastian refers skill, I would not want
57:34that with my marketing head of marketing
57:36cuz it's just like plugs up the context
57:38and you don't need it there.
57:40Um, so I would have that as a project
57:41level for example. And you can have
57:44um, global skills versus project skills,
57:46global Claude dot MD versus project
57:49Claude dot MD. And same with MCPs, you
57:51have global MCPs and project MCPs.
57:53That's probably the other concept
57:55um, to go over.
57:57But look, other than that, that's pretty
57:58much the entire agents crash course. So,
58:01it's just that loop running in the back
58:03end to complete your task
58:05and connecting in your tools, your
58:07context, and the LLM all in one place.
58:11Uh, and I would just say to to work out
58:13what roles you want to start to build
58:15out an agent for,
58:16go into Claude or your favorite chat
58:18model and get it to help you build out
58:20those
58:21uh, context files through an interview
58:23style process. Just say ask me questions
58:25to build this out.
58:27I would connect all the tools that you
58:28need and then start building out the
58:29skills through daily use.
58:32And then pretty soon you're going to
58:33have like pretty powerful AI agents
58:35built for every single um, aspect and
58:37department of your business.
58:39>> Ramy, thank you so much. I'll include
58:41links uh, in the show notes, in the
58:43description, where you can go follow
58:45him, get to know him a little bit
58:47better. And uh, I appreciate you coming
58:50on dropping some sauce. Thank you, man.
58:52>> Thank you so much for having me on,
58:53Greg. It's been a blast.

About this transcript

This page contains the full transcript of Building AI Agents that actually work (Full Course) by Greg Isenberg, generated from the public captions YouTube serves with the video. The transcript has 10,815 words across 1,754 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.