YT YouTube2Text

YouTube2Text → Transcripts → AI Agents are too good at marketing now

AI Agents are too good at marketing now — Transcript

by Greg Isenberg · 9,474 words · 1,319 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00If you're anything like me, you want to
0:02figure out how you can use AI agents to
0:04go and get you customers that make you
0:07money, that build high ROI, that could
0:09just take your company and make it a
0:12successful thing. But how do you
0:13actually do it? What are the workflows?
0:15How do you think about the skill files?
0:17How do you use Claude Code and Grockbot
0:19to actually [music] do all these things?
0:21Well, in today's episode, this is a
0:23fullon master class for how you can do
0:26that. We have Cody Schneider and he
0:28takes you through step by step in the
0:30most clear way I think I've ever heard
0:32for how you can be a marketing engineer.
0:35Basically a fancy way to use AI agents
0:38to do marketing on your behalf so you
0:41can make money 24/7. We learn how to do
0:44static ads. We learn how to do AI UGC.
0:47We learn just all sorts of things around
0:50paid media, AI content, how to do it.
0:54So, by the end of this episode, you're
0:56going into 2027
0:58understanding how to be the top 0.1%
1:02marketing engineer on the planet.
1:03Because let's be real, a lot of people
1:06don't know how to use these tools. A lot
1:08of people don't know how to pipeline
1:09data into this. A lot of people don't
1:11know how to put the full picture
1:13together to actually create gorgeous
1:15content, gorgeous ads that are going to
1:18get you paid. But I got good news for
1:21you. You're listening to this. This is
1:22100% free. Like, comment, and subscribe
1:25for more, and I'll see you at the end of
1:26the episode.
1:30[music]
1:35>> What's up, Cody? By the end of this
1:37episode, what are people going to learn?
1:39>> Totally. I'm going to teach you how to
1:41make Grockbot the best marketer in the
1:43world, top 1% or your cloud code.
1:45Doesn't matter what harness that you're
1:47using. And then the other thing that
1:49you're going to walk away from is you're
1:50going to know how to make static ads
1:52that are on brand that are talking to
1:53the pain points and desired outcomes of
1:56your target customer. And then we're
1:58going to make UGC avatar ads that look
2:01real, that talk real, that have motion
2:03graphics over the top of it. That entire
2:05workflow is going to be just yeah skill
2:08files that you have access to. They're
2:09literally in the description below. Give
2:11it to your agent and it can go and do
2:13this work. So that's what they're going
2:14to walk with G. Okay. So, you're going
2:17to teach people about how to use AI to
2:19make static ads, video ads, all these
2:22cra, you know, I've seen just like go
2:24viral all these like AI influencer type
2:26ads. I also want to know how I can layer
2:29on human beings into these videos. You
2:32think you can also let let us know about
2:34how to do that?
2:35>> Absolutely. Yeah, we'll talk through
2:37like how do you basically make the
2:38coordination layer that enables you know
2:41a person with a Grockbot or a Grockbot
2:43itself to like be this creator manager
2:46using a tool like tribe or side shift um
2:49to actually like do you know the UGC
2:52production or layer in the human that's
2:55the gate of like hey I need this human
2:57involvement before any of this goes live
3:00where that human taste is added to the
3:02process to the workflow um rather than
3:04it just being this totally autonomous
3:06thing. So, how do you co-work with this
3:08but solve the labor problem for
3:10yourself? So, you don't have to have a
3:11graphic designer, you don't have to have
3:12a video editor, you don't have to have a
3:14sales development rep. Uh you go down
3:17the list, all these pieces can basically
3:18be handled and solved. So, this is the
3:21advice that if you had a 17-year-old son
3:24and your 17-year-old son had built a
3:26startup and they had no customers and
3:29you know, you sat you sat him down and
3:31you said, "Son, I'm going to teach you
3:34in under an hour everything you need to
3:36know about how you can use AI agents to
3:40create ads um and create content that is
3:44going to get you customer. Is that is
3:46that what we're talking about right
3:46now?"
3:47>> That is exactly what we're talking
3:48about. it. Ads are one of the most high
3:50lever things that you can do to get
3:53initial traction and also to validate an
3:55idea. If you have a business idea that
3:57you're trying to get uh like to see if
3:59the market wants uh that this is the way
4:02to to basically make the media that's
4:05necessary to sell that product or
4:07advertise that product, right? Um this
4:10is the exact thing that I would teach
4:11somebody. Once you do it this way, you
4:13will never go back. you will never do it
4:15the other way because it feels archaic
4:17to do it any other way than this once
4:20you are in the weeds and basically you
4:22know building out these solutions. So
4:25>> thank you father let's get into it.
4:27[laughter]
4:28>> All right let's jump in. So first off I
4:30just want to talk about this a little
4:32bit deeper and I don't know if you want
4:34to bring up that diagram that we were
4:36kind of working on right before this uh
4:38Greg of like what is happening in this
4:40category. So the the big unlock and I
4:44think the the you know the mind shift
4:46that I want you to get into here of like
4:48what what is changing is you know
4:50historically if I wanted to go and make
4:52for example a static ad right I would
4:55have to go and get a graphic designer
4:58and I would have to go and like do the
5:00copywriting ideate off of you know
5:02here's these different angles desired
5:04outcomes pain points that I want to to
5:07focus on. Um now instead that is just
5:11code and tokens that are being burned
5:13right and so all marketing is just going
5:16to this like all marketing is now just
5:18code it's just software and tokens and
5:22so when you look at you know an static
5:24image for ads that you generate when you
5:26look at uh you know a seed dance video
5:29that you create where before I used to
5:31hire have to hire an a like an avatar
5:34you're sorry hire a UGC creator to make
5:36this and then have a video editor to
5:39edit it and then have a motion graphics
5:41designer to do the motion graphics,
5:43right? That whole thing is just stacked.
5:45That work is now handled by a Grockbot
5:48or a claw code or a dots or what
5:50whatever it is that you're using.
5:51There's, you know, all all these
5:53different uh uh vehicles for this coming
5:55out. But I I think the big thing I want
5:57you to understand is that that labor
5:59that was necessary previously is no
6:02longer gone. Like you there's a
6:03renaissance that's happening in long
6:05form video editing. I imagine your
6:07editor, Greg, is starting to already do
6:08this where like you can literally use a
6:11software called Mirage Tessact. Um, and
6:14it will go and oneshot edit, you know, a
6:17podcast that's an hour and 30 minutes
6:18long. It'll cut out the silence. It'll
6:20cut out the takes. It'll like put it in
6:23frames. It will do the edits of like
6:25showing the screen when it's
6:26appropriate. Does the active speaker
6:28identification. All of this work that
6:31like used to be a human on a computer.
6:33We can now get agents to do this. I did
6:35not think it was going to get here this
6:36fast. Like to be totally blunt. Like I
6:39did not think long form video editing or
6:40like a YouTube video editing was going
6:43to get here uh you know this year. But
6:45here we are. And I think that today like
6:48I the the big thing I want you to
6:50understand is that all of the work that
6:52used to be a uh you know a human
6:55touching keys, moving the mouse that is
6:58now that can now be agents that are
7:00basically doing doing uh this work. But
7:02you have to give those agents like you
7:04have when I say agents like you have to
7:06give your grockbot, your cloud code,
7:07whatever it is that you're using all of
7:09the necessary like infrastructure that
7:12it needs to be able to do that work,
7:14right? You have to how I've been
7:15describing it is you have to make like
7:16an environment for it that is like
7:19healthy for it to be like to work,
7:21right? It's like this petri dish that's
7:23like, you know, cozy and hea and
7:25whatever, right? That like it it it can
7:28basically like thrive in. And so to do
7:30that what you need is you need to
7:32provide it data. So access to your dis
7:34business data. You need to provide it uh
7:37access to tools. So tools when I say
7:40that that's like API calls to uh you
7:43know a chat GPT image 2.5 endpoint or
7:45API calls to seed dance or API calls to
7:49uh you know you want to scrape your
7:51competitor's ad library. We're going to
7:52do all this today but this is like you
7:54know that's what I mean when I say
7:55tools. And then the other piece is you
7:57need to give it access to your ad
8:00platforms. For example, your Facebook
8:02ads, your Google ads, your websites, you
8:05know, CMS so that it can actually go and
8:08publish that content uh that it produces
8:11or or get the final product, you know,
8:13into the wild actually, you know, into
8:15into the public. So, that's kind of the
8:17high level. Any questions about that I
8:18can try to answer before we jump into
8:20this and actually do these builds? I
8:22mean the only thing I want to just say
8:23is your the whole everything is for you
8:28to do that job like your job now person
8:31listening is to basically build the
8:35system that Cody is describing the petri
8:38dish that Cody is talking about the
8:40machine around all these models agents
8:45uh the machine around uh context the
8:48machine around you know evals That whole
8:52thinking is the job of the person who's
8:55building a business in 2027 if you
8:58actually care about it getting customers
9:01on day one.
9:02>> Exactly. And I I think that this is
9:06again like the shift that's starting to
9:08happen is like get out of the way of the
9:10agents. Like you are actually in the way
9:13like your your job is to be a conductor.
9:15There's this running meme right on
9:17social media of like the best like you
9:19know the most effective employees will
9:21be people who played Starcraft
9:23[laughter]
9:24which is this like game where you would
9:25like coordinate you know swarms of
9:27actions and like you know it's basically
9:30it's that same philosophy that we're
9:32applying here but um okay so how do you
9:35actually do this? I'm going to talk
9:36about the open source version that you
9:37can like monkey patch together uh if you
9:40want to basically build everything that
9:41I uh you you know you need for these
9:43agents. The other thing uh that I'm
9:46going to uh talk about is then okay how
9:48do you get started today if you're like
9:50an hour from now you want to do this
9:52exact thing how do you go and you know
9:54get this basically live and into the
9:55wild. So first off on the uh open source
9:58infrastructure uh if uh let's solve the
10:01data problem first. So on the data
10:04problem side, you can use a software
10:06called Airbet. Uh and so you need to
10:09build two things. You need to build uh
10:11basically a data pipeline and a data
10:14warehouse. Why do I need to build these
10:16things? If you're just querying against
10:18the APIs of the uh your ad channels or
10:22you know any of the other solutions that
10:23you're doing, you're going to run into
10:25the same problems where it's like a
10:27massive amount of data that are all
10:29going into context and the agent is very
10:32likely to hallucinate on this or what
10:35it's going to go do is it's going to go
10:37write its own like database locally and
10:40try to stuff all of that data into
10:42there. But then you run into like these
10:44duplication errors and all these pieces.
10:46So how do you solve this? You use
10:47infrastructure that's built specifically
10:50for this that exists for this exact
10:53purpose. So the open source versions of
10:54this are a company called Air Bites. Um
10:57they are the data pipeline. So when you
10:59think about them like what they're doing
11:01is they're basically going and they are
11:05um uh like taking all of these different
11:08sources of data, right? Your Facebook
11:10ads, your Google Analytics, your HubSpot
11:13CRM, right? anything that you store data
11:16in and they are uh basically the
11:20pipeline that is going into uh this data
11:25warehouse. Right? So when I uh point to
11:27here, this is where Arabte fits in.
11:30Right? So this is Arabte
11:32and in contrast
11:35where this all goes uh is a data
11:38warehouse. And so that data warehouse
11:40what you can use is a company called
11:42Clickhouse. Shout out Clickhouse. They
11:44are an awesome partner for us. Uh so
11:47[laughter] we love them for everything
11:49that they do. Uh and that data warehouse
11:52is then where all of this data from all
11:55of these sources gets stored. So why am
11:57I putting it all into a single place?
11:59It's so that I can look at the data
12:02across the entire organization. I want
12:05to be able to see the ripple effect of I
12:08ran this ad and I can connect it to my
12:10Google Analytics. I can connect it to my
12:12HubSpot. I can connect it to my post
12:15hog. I can connect it all the way down
12:17to my payment processor and say, "Hey,
12:19when people see ads like this, I get
12:22more revenue. So, let's go put more
12:24money into that." The the thing that
12:27gets even crazier though, Greg, is like
12:29you can actually see who's a better
12:31customer lifetime value. So, we can tag
12:34an ad all the way down to like better
12:37customers come out the other side. like
12:38they are a better customer lifetime
12:40value for us when they see media like
12:43this. So not just like it made revenue,
12:46it actually made like more lifetime
12:48revenue when we are doing these
12:51activities, right? So all this goes into
12:53the data warehouse and then that data
12:55warehouse you're giving access to that
12:58data warehouse for your agent. So when
13:01you think about this, this is where
13:02Grockbot is, right? Or claude code or
13:05whatever it is that you use as your
13:07driver. And this is back and forth. It's
13:09basically querying that data. And why
13:12you we do this is it's constantly
13:14syncing in the background. So this data
13:17is constantly streaming in, right? It's
13:20constantly streaming in to the data
13:22warehouse. And you can query this data
13:24warehouse as much as you want. You can
13:26nuke this data warehouse with queries,
13:28right? And this is something we see
13:30happen all the time with businesses is
13:32they're like they're going directly to
13:34these sources and they're running into
13:36okay rate limits from the API or these
13:38challenges of like collecting it or they
13:40have to have like a claude function
13:42that's running on an hourly cadence or
13:44you know you know some some monkey patch
13:47solution that that that is they're they
13:49built locally to solve this. This is how
13:50you get around this problem. So you give
13:52it the data. So that's the first step of
13:54this. That's how you give it the data.
13:56The second step is you need to give it
13:58access to these platforms. So let's say
14:01this is Facebook ads, right? So Facebook
14:03ads, how you give it access is you need
14:07to create uh what's called a uh Facebook
14:10ads API token. Um so it's called the
14:14marketing API. and the marketing API how
14:18that's used uh is basically to do writes
14:23to the uh uh to the platform. So I'm
14:28never actually reading the data from
14:30Facebook ads. This is only
14:32onedirectional. I'm only using this
14:36Facebook ads API
14:38to upload ad creative
14:41to turn off ads
14:44to move ads around. All of the data
14:47reads are only ever happening from the
14:49data warehouse. Now, why do I do that?
14:51What is the point? Like what is the
14:53reasoning for this? The reasoning is
14:55because you're trying not to abuse the
14:57Facebook ads API or the Google Ads API
15:00which they have terms of service and
15:02agents will go and just like do a
15:05thousand requests at once and this that
15:08is against their terms of service. So
15:10you probably have seen these online
15:11where like people like okay uh Grockbot
15:14got my Facebook ads account banned right
15:16like don't use agents for this. That's
15:17not what's happening. your Grockbot
15:20violated the terms of service of the API
15:22because it was sending thousands of
15:24requests per second
15:26>> because it was trying to do some data
15:27pull. So,
15:29>> you should just you should assume your
15:30agents are like drunk patrons at a
15:33restaurant.
15:34>> Absolutely. Absolutely. Like it's your
15:36drunk uncle at the like local Irish pub.
15:39He's eight Guinnesses deep. He's split
15:41the G four in a row for the last like
15:43hour and he's basically like this is
15:47that's how he's interacting with the
15:48Facebook ads API, right? Um
15:52this same strategy, you want to do this
15:53for everyone though, right? You want to
15:55do it for Facebook ads, you want to do
15:56it for Google ads, the same thing, their
15:58API. We have documentation on this. I'll
16:01send this over to you, G, so that you
16:03have it in the notes. It's just a notion
16:04doc. It takes you about five minutes to
16:06get access to this. It's not super hard.
16:09Um, you literally can give the notion to
16:10your agent and it will just like go and
16:12walk you through the setup process for
16:14this. So, I've now solved two things,
16:16right? I've solved the data. How do I
16:19give that to the agent? Let me zoom in
16:21so we can see this. So, one, how do I
16:23get it the data? Two, how do I get it
16:26the access it needs? And now three, this
16:28is the thing that people are just
16:30starting to discover. How do I give it
16:32the tools that it needs to go and do the
16:35work that's necessary for it to be able
16:38to do media buying to function as a
16:41sales development rep to function as a
16:43video editor etc. And for these tools,
16:47what we like what you can do the again
16:49the open source version of this is like
16:51you go and get a Kai AI subscription and
16:54you go and get a foul subscription or
16:57put tokens in there and you go and you
16:59get an open router you know credential
17:01and you go and you get a uh uh you know
17:05an Apollo subscription, right? You can
17:07go down this list of all of these
17:09different tools that you have to go and
17:11basically get access to. And when you
17:14have all these tools, all these APIs and
17:16you provide them to the agent
17:20it now. So I'm just going to take this
17:22tool piece and we'll just you know put
17:23this over here of a list of a handful of
17:25these. So like KAI, like Val for
17:29transcription, um like Open Router for
17:32LLMs,
17:34um again Apollo or and we uh actually
17:37we'll just do like Git Leads because
17:39they're a great partner. um or like get
17:41leads for your uh email data provider
17:44and like person data. When you do these
17:48tools and you give it access to the
17:51agent, it now can do the middle work for
17:55you. So, for example, it can make an ad
17:57using Kai AI, a static ad. And I'm going
18:00to show you literally how to do this.
18:02Like it it'll you're going to just paste
18:04the skill in and it's going to ship ads.
18:06It's ridiculous. Um but when it has all
18:10these tools, it can now do this middle
18:12work. It has the environment that it
18:15needs. So I was talking about this
18:16earlier like this is the environment
18:18that's necessary for these coding agents
18:21to do this work to do marketing, right?
18:25And when you have this environment,
18:28suddenly your agent becomes the most
18:30like capable marketer in the world. I'm
18:33talking like top.1%.
18:35I've been doing this 10 years. It will
18:37come up with stuff where I'm like, "Holy
18:39shit." Like, I would have never thought
18:41about that. An example of this recently
18:44is we have an agent that runs Google
18:46ads, right? This is like a template that
18:48we have within Graph. Um, what we found
18:51is that we can in real time audit the
18:54search terms that we're coming up for
18:56for paid search and we can have an LLM
18:59be like is this search term like is the
19:03search intent buyer ready for the
19:05landing page it's sending it to? So be
19:07before we're even spending money,
19:10we're having like this, you know, labor
19:15audit whether this is a good thing for
19:17us to even be spending money on. Now, I
19:19could do this previously, right? I could
19:21have a human that was running that same
19:23process. But I getting a human to audit
19:275,000 search terms [laughter] in 20
19:30seconds is impossible. And that's the
19:32scale that we're talking about here. and
19:34you can get it to do this work and and
19:36and basically function uh as like a team
19:38of people that was necessary. So this is
19:40the environment that you need to make.
19:42That's the open- source way to do this
19:44is go and get all those pieces. If you
19:46don't want to do that, you can use uh
19:49our company graph.com. Go to graph.com
19:53uh platform. It's literally free to get
19:55started. Everything that I just
19:57described
19:59uh is built into this. the data, all of
20:02the access pieces we handle for you. We
20:05also make it so it's multiplayer. So all
20:08your team members, for example, can be,
20:10you know, five people separately using
20:12their Grockbot or their Claude code can
20:14be managing the Facebook ads account,
20:16you know, through this without having to
20:17have all their own system set up. And
20:20then also all the tools, all the tools
20:22that are necessary to do all the work
20:24that I'm about to show you. Again, you
20:25can set this up yourself. That's just
20:27the way. If you're like 10 minutes from
20:28now, you want to be able to do this,
20:29this is how you go do this. All right.
20:31>> Two two things. Two quick things.
20:33>> One, just because people always ask, I'm
20:35not affiliated with Graph at all.
20:37>> He is not. Yeah. I'm here just teaching
20:40this work on how to do this. So
20:42>> two,
20:44there's one piece of this pie that this
20:48marketing environment diagram that I
20:50feel is missing, which is what I would
20:54call a swipe warehouse.
20:57So, in order to get the most out of your
21:01Grockbot, your cloud code, whatever it
21:03is you use, you need to have examples of
21:07inspiration that is going to help you
21:09get the top
21:12creative because a big part of this
21:14game, if you want to be the best
21:15marketer, if you actually want to get
21:17people to actually buy whatever it is
21:19you're buying, you need good creative.
21:22And the the the reality is if you have
21:24good creative, the algorithms in in
21:27terms of paid ads, the algorithms are so
21:29good now that they'll match your good
21:31creative to an ICP that'll buy your
21:34stuff. So the question becomes, how do I
21:39like I buy every I buy the data, I buy
21:41the access, I buy the tools, I buy this
21:43whole marketing environment, but how do
21:45you create a,
21:48you know, swipe warehouse? So when I say
21:50swipe, I'm basically saying a swipe
21:52house, a swipe file, but a swipe files
21:55made sense pre-ai. Now you need like
21:58thousands of swipe files. Okay. Yep.
22:01>> What am I looking at?
22:01>> We're just starting to build this. Okay.
22:03>> And so as an example, like you want that
22:05Google Ads agent that I was describing.
22:08>> Yeah.
22:08>> Literally, here's the code. [laughter]
22:10>> Yeah.
22:11>> Just paste that into your cloud code and
22:14out the other side will come out a
22:16Google Ads agent. Um, you want to like
22:19we're going to use this exact one here,
22:21the static ads in UGC. You want to make
22:24static ads, same thing. Just literally
22:26paste it in and it's going to be able to
22:28go and like do these builds for you. So,
22:31all right, let's actually build some
22:33Let's get this like publicly out
22:35there.
22:36>> Cool.
22:36>> So, I'm going to make some ads. Greg, is
22:38there what's a what's a product that uh
22:40should we do? Idea browser? We can do a
22:42bunch of them, but like you you rattle
22:44off some company names and we're just
22:45going to start like making creative for
22:47them right now.
22:48>> Okay. I would say idea browser and our
22:51design and strategy agency late
22:53checkout. Agency.
22:55>> Okay. Late checkout.
22:58>> We've never actually done paid ads for
23:00it.
23:01>> Curious to see what it does. So, okay.
23:04So, first off, what we're going to do is
23:07I'm going to take Idea browser, right?
23:10And I have installed this skill, the
23:13static ads. And again, I'm using Graph
23:15for this. You can do everything that I'm
23:17about to do with like an open source
23:19version of patching this all together.
23:21This is just the fastest way to do it.
23:23So, Graph is installed. And what I mean
23:25by that is we have the Graph CLI
23:28installed, which has like all of this
23:31knowledge like stacked in this. When I
23:34say the CLI, what I mean is it's like
23:37this tool set. It's just a package that
23:41has all of the knowledge on how to use
23:42the platform like the marketing
23:44engineering cloud is what we're calling
23:46it stored within that. So I've already
23:48got this static ad installed. I just do
23:51forward slashstatic ads and I'm going to
23:55drop in the brand URL and hit enter. And
23:59what's about to happen is it's going to
24:01go to the website and extract the brand
24:04style guides of the website. It's going
24:06to go and find the Facebook ads library
24:09of idea browser if it exists. It's going
24:12to go and then uh research the pain
24:15points and desired outcomes that this
24:17product uh solves by going and
24:20extracting conversations from Reddit and
24:23other social media. It's then going to
24:25propose five static ads creative. And
24:29then it's going to go and generate the
24:31ad creative
24:33based on all of this information that it
24:35just extracted. And then from there,
24:37we're going to be able to go and upload
24:39this directly into Facebook ads. Again,
24:43going back to that diagram, because we
24:45have all of this set up, all of this
24:47access, right? Perfect. So again, what I
24:50did is I did I forward slashstatic ads
24:53ID browser and it's now just working.
24:56It's running this skill. And when I say
24:58it's a skill, like I'm not talking about
25:00like it moves some, you know, text
25:02around. It's going to actually do real
25:04work. It's going to do research. It's
25:07going to actually pull in like what the
25:09human would do to make this creative.
25:11It's going to go and do that process.
25:13So, let's go do this for another
25:14company. Now, let's do this uh for late
25:17checkout, right? So, I'm going to create
25:20uh another session. So, for/static ads,
25:25I drop in the business. That's all I
25:27have to do. Out the other side is going
25:29to come these static uh pieces of ad
25:32creative. So, let's come back. We'll see
25:35what's happening. It's extracting the
25:37brand here. What other brands, G, are
25:40you guys work or that that we can do
25:42right now? Um any products that you're
25:44excited about or like I mean it could
25:46even be ecom brands as well. Like
25:48literally anything.
25:50>> I would be interested to see like how uh
25:53like you know M Leon Dore, do you know
25:55that brand? I don't know. How do you
25:57spell it?
25:58>> A a i m e.
26:03>> Cool.
26:06>> All right, let's let's drop that into
26:09>> some something like that has like a
26:10strong brand and probably has some ad
26:12somewhere.
26:13>> Totally.
26:16So, we'll do them. Who else?
26:20Um,
26:21I would be interested to do
26:24like a SAS app probably.
26:27Um,
26:29>> what's a what's a big one that you guys
26:32have done recently? We can do cursor.
26:35>> Yeah, cursor is great.
26:37>> All right, let's do cursor. All right,
26:40we'll do cursor as well. Just firing
26:42these off and just to come back to this.
26:45So, idea browser, this is what it came
26:48up with.
26:50So, it's going to make five statics
26:53>> and headline, stop guessing which idea
26:55to build. Every idea is scored on
26:58demand. Pain and timing. Get uh today's
27:01free idea. Headline build what people
27:03already search for. Headline your chat
27:05GPT ideas are generic. Headline don't
27:08waste six months on a dud. Headlineund
27:11or 1948 ideas worth building. Okay. So
27:15these are basically the ad ideas or the
27:17concepts that it extracted from going
27:21and researching online like what are the
27:25pain points desired outcomes of somebody
27:27that would buy this product. Right? So
27:29now I said okay generate these ads and
27:32out the other side is going to come the
27:34actual ad creative uh you know based off
27:37of these concept points. And this is how
27:39you get good like outcomes to happen is
27:42you again you have to give the agent
27:45[laughter] what it needs the environment
27:47that it needs to do this work right so
27:50let's come back to these other ones
27:51we'll see where they're at oh they have
27:53zero ads that's fascinating um cursor
27:56cool so cursor we're pulling uh the
27:58Facebook ads libraries that's great and
28:02late checkout still uh building the
28:04brand style guide there so um let's do
28:07some other brands While we're waiting,
28:09uh, trying to think, uh, let's do David.
28:14Just like the protein bar.
28:15>> Mhm.
28:16>> What other brands are you stoked about
28:18right now? G
28:20>> stoked about
28:22>> or just come to [laughter] mind.
28:26>> Um,
28:27>> let's let's do an old classic maybe of
28:30like uh I don't know.
28:32>> I think Robin Hood's a pretty good
28:34brand.
28:35>> Cool. Let's do that. Robin Hood as well.
28:39All right, these are ready.
28:41>> You know what's a really cool brand and
28:43product? Teenage Engineering.
28:45>> Oh, that would be sweet, actually. Yeah,
28:48I'm going to say, okay, make these
28:52and let's go to idea browser. Here's the
28:55statics that are coming to are starting
28:56to come out.
28:57>> Wow.
28:58>> So, we'll take a look at those in a
28:59second.
29:01And same deal here. like make these and
29:06same thing make these. All right, so got
29:10a handful of these. Let's go back to
29:11idea browser. Let's look at these pieces
29:14of creative, right? So, stop guessing
29:17which idea to build.
29:21Again, this is just based off of the
29:22brand that it's pulling.
29:25>> You can modify this. You can also give
29:27it the skill files that define your
29:30brand style guides. This is just a
29:31markdown file. What are the fonts? What
29:33are the colors? Again, this is just
29:35extracting from the site, going and
29:37doing that research. But I've made these
29:39ads now, right?
29:41Once I've made these ads, I can then go
29:45and because I've already set up the
29:47Facebook ads API, I can say something
29:49like upload these
29:51into the Graph Facebook ads account into
29:56a new ad set, right? and it's going to
29:59go and upload those directly into it.
30:04I'm just going to stop this for the time
30:05being. And now what instead I'm going to
30:08do is show you how to take these same
30:10concepts and we're going to turn this
30:12into an AI avatar UGC that's doing a
30:17script read that also has motion
30:20graphics over the top of it. So I'm
30:22going to do forward slash UGC.
30:26write a script
30:28for this based on above. So this skill
30:32file it understands okay how do I need
30:35to go and make this seed dance ad
30:38while this is generating
30:41uh we'll just uh do these also or make
30:45them now come back to these
30:50uh those are still rendering so in the
30:53example they used the static ads were
30:55used uh using chat GPT image 2 the UGC
31:00ads are using CEN Is there an
31:02opportunity to like let's say I don't I
31:04don't want to use image 2 I get better
31:07results on nano banana or I get better
31:08results on whatever is there a way to
31:12swap models out
31:14>> yeah absolutely so you like for example
31:17if you get a kaii subscription um you
31:20could use nano banana etc um with the
31:25unified API has all those endpoints so
31:27you just literally say like you know
31:29change this the nano banana from Chad
31:32GBT image. Chad GBT image 2.5 is the
31:34best that we've seen. So like I would
31:36suggest it. Um but anyways, so we're
31:40going to make this UGC ad. Now there's
31:43already a pre-built avatar in here. It
31:45just wrote the script. I'm going to say
31:46okay, make the ad. And what's about to
31:50happen is a couple things. So people, we
31:54see a lot of people being like, I just
31:56spent $50 to make this UGC ad. like
31:59burning like tokens through this. So
32:02this workflow that we're about to run,
32:06how we've solved for this is we've done
32:07a handful of things. The first thing is
32:10that we're taking the script. So imagine
32:12you know this is a long script and that
32:16then goes
32:18and we do an ad read using not seed
32:23dance. So the audio is actually using a
32:26foul uh voice clone and then this makes
32:30an MP3 file. This is what you know comes
32:34out the other side is just the MP3. The
32:38next piece is we send
32:41the MP3
32:43and also an image,
32:47right? So imagine this is like the
32:49creator image, right? This is a
32:52reference image that we're sending at
32:55the same time with the MP3
32:57and also the script gets sent as well.
33:00So all of this goes to seed dance at the
33:03same time. Why are we doing this? Why am
33:05I even talking about this? If I just
33:07send the script straight to seed dance,
33:10it's going to give me just like whatever
33:12it invents at that [clears throat]
33:14moment. So what I have to do is I
33:16basically have to be like here's this
33:18walled garden that you can work within
33:21the script exact script the exact voice
33:25example doing the ad read so you're not
33:28like inventing like if you've ever had
33:30seed dance where it's like and it's like
33:32it's not make it's not actually
33:34enunciating the words this is how you
33:35can solve for this and then also the
33:39reference image so that the avatar that
33:42you get out is what you're looking for
33:43it's not some random thing that seed
33:45dance thinks seed dance thinks about or
33:47you know makes up. So all of this gets
33:49sent to seed dance and then out this
33:52other side you get this ad that is ready
33:56for you to edit. This isn't the final
33:58form. This ad that you get is 480p,
34:03right? So we'll say that you know this
34:04is the first version the V1 of the ad,
34:07right? That it's this video. This is
34:09480p.
34:11Why are we doing that? It's because if I
34:13did a you you can use a higher model,
34:15but uh it just, you know, if you're
34:17trying to be costconcious, this is the
34:19way to get around this. So, you can do
34:20like 1080p, but it costs you, I think
34:22it's like, you know, $14 for that. In
34:25contrast, I can do a 4080p
34:27and then I upscale
34:31the video
34:32and I add blur using what's called
34:35ffmpeg. Again, you don't need to know
34:37any of this. All of this is happening
34:40within the skill that I just used to
34:43make this ad. Right. The other thing
34:45that it's doing is it's doing wordbyword
34:49transcription and it's matching the
34:53script to the transcript.
34:56Right? So it's matching the script to
34:59the to the transcript to confirm that
35:02it's an exact read that's happening. Now
35:04say it mispronounces kalani, right? and
35:07it says like kalandla, right? This is an
35:09exact example that like we had happen.
35:12Instead of that kalanla now being in
35:15there, what it does because I have the
35:17wordbyword transcript,
35:20it can isolate, hey, this word is wrong.
35:24Kalendla is wrong.
35:28Great. Here are the timestamps
35:31right
35:33of where that is said. And I'm now going
35:36to go and I'm going to replace this with
35:38the correct pronunciation of Kalani.
35:42Right? And we can swap basically those
35:45mistakes. And this is again this edit,
35:47this video editing that I just described
35:49here. This is not a human. This is the
35:52agent that is basically doing this work,
35:55right? Doing this within Cloud Code or
35:57your Grockbot. Everything that I'm
35:59describing as well. I think that the you
36:01know this is kind of the meta yeah of
36:03this whole thing. This can all run in a
36:05cloud.
36:07So like I can host a cloud code harness
36:10in the cloud and have it do this entire
36:12workflow. So like our engineers like
36:14this is what they do. You want to you
36:16want to you want a 100 pieces of
36:18creative using this exact process like
36:22this is exactly how it happens. It's
36:23hosted in the cloud. It's not actually
36:25on the machine locally that this is all
36:27going on. So out the other side, what I
36:30get from this when it does this whole,
36:32you know, stack of things on the video
36:34edit side, I get this complete ad
36:38that is doing the pronunciation right
36:42that's upscaled and that looks actually
36:45like the like a human like like a human
36:48videoed it on their phone. And the
36:50reason for that is we use a source
36:52image. I forgot to mention this. We use
36:54a source image that is taking a real
36:58creator and remixing their remixing
37:01basically that creator, right? I'm not
37:03saying a likeness. I'm saying like we're
37:05like change the background and make this
37:08person, you know, this age, this
37:10ethnicity, uh, you know, etc. is
37:13basically what this ends up being. And
37:16if you ever get that like plasticky like
37:18effect on seed dance, this is how you
37:20solve for this is you give it, you know,
37:22a real human version for it to go and do
37:24that remix. So I now have this video
37:27that is ready that is, you know, a UGC
37:30read. But there's things I want to do. I
37:32want to add captions and I want to add
37:33motion graphics to it. How do I go and I
37:36do that? So this next step is you use a
37:39product called Tesseract from Mirage.
37:41This is the exact thing that we're about
37:43to do.
37:45And [snorts]
37:46what it can do is it can take that word
37:48by word transcription [clears throat]
37:50that we have up here and it can add
37:52captions
37:54or it can go and add motion graphics to
37:57the video. So let's go and look at this
37:59now. Video is probably done. You know,
38:01takes about 5 minutesish to render. So
38:03let's see if it's completed. And it's
38:06completed. So let's take a look. Uh I'll
38:09screen share.
38:11Cool. So this video is now rendered.
38:14And
38:16this is it doing the lip read, doing the
38:19sync. It's done those entire edits.
38:23And I can now go and say, okay, add
38:26motion graphics
38:28to this video
38:30using Tessact.
38:34>> By the way, she looks insanely real.
38:37I So we have hundreds of these and we
38:41can make any person type like you name
38:44it, we can get it. We've got ones like
38:46you're trying to sell to the bluecollar,
38:48awesome. You're trying to sell to
38:49somebody who's in a corporate office,
38:50awesome. You're trying to you you name
38:53the demographic and like or the
38:55psychoraphic and we can go and hit it.
38:58So this cost $4 to make about $4. I
39:02think it was a little bit more. We'll
39:05ask it at the end. That was It says 475
39:07was the projected. Um, so
39:11this motion graphic piece will be a
39:13couple more cents. So basically, I'm
39:14getting this ad out for about $5. Um,
39:17we're about to work with another
39:19provider. We'll be able to get this down
39:20to about three is where we think we'll
39:22end up be in that vicinity. But I've got
39:25this ad now. So say I wanted to go and
39:26make a hundred of these
39:29um or we'll say 10 of these, right? Uh
39:33that's 50 bucks. I want to do a 100.
39:35That's 500. [laughter] If I was to go
39:38and get this made previously and add
39:40motion graphics to it, I mean, you're
39:42easily in the range of $500 plus in the
39:44in the the past, right, for just that
39:46single ad. And this is this way around
39:49this. Now, these ads, you don't have to
39:52only use these AI avatars. The other
39:54thing that we're starting to do is
39:56layering on tools uh for um like example
40:02is tribe where you basically make a UGC
40:04creator network. What's amazing now is
40:07instead of having to have a uh person
40:11that was managing all these
40:12relationships, you can use the API and
40:15your Grockbot and your cloud code to do
40:18to basically be the social media like
40:20social creator manager, right? the
40:23creator management person. So this is
40:25like literally how we are doing this. We
40:27use that. We also use a company called
40:28Sideshift as well. Um you know just two
40:31examples of like how do you run these
40:32campaigns?
40:33>> So wait, so how would that actually
40:34work? Cuz so what you're saying is Tribe
40:37is a network of UGC creators. So you're
40:40saying I can
40:41>> So we pay them like $40 a video as an
40:43example for a company that we were
40:44working with.
40:45>> I would take this script
40:47and I would be like, "Hey guys, here's
40:49this script, you know, a variation. feel
40:51free to modify it, but this is the D,
40:53you know, directionally basically where
40:55to go for this, right? Um, that via the
40:59API of Sideshifter Tribe gets sent to
41:01them with, you know, a prompt and then
41:02you pay them on like a pervide basis or
41:04a percentage. I I there's a video that
41:07went viral. I think his name is Hector.
41:10Yeah, I'm I don't know him, but it's
41:13super legit like what he's doing. This
41:15is the exact strategy that we implement.
41:18We do it on a per video basis and then
41:19have like a bonus based off of this. um
41:21the spend. But there's tons of ways to
41:23basically structure these. But um what
41:25this turns into is you can have the
41:27human creators. So you have instead of
41:29the avatar doing the read, right? You
41:32can have the human creator and then go
41:34to okay, I want the infograph or I want
41:37the motion graphics uh to be added to
41:39that human creator. So just a way to to
41:41make this happen, right? The exact
41:43workflow I just described, we can do
41:45that for all of these different ones. So
41:47here's David, right? Built on science,
41:49backed by doctors. Cool. For late
41:51checkout, your product is getting a
41:53second user. Your next product is your
41:56support ticket.
41:57>> Every company can have a lab now.
42:00>> Fire.
42:01>> Some examples. Uh cursor. Uh finally
42:06understand the code nobody wrote down.
42:08Uh looks like it's rendering the rest of
42:10them. I'm really curious what it's going
42:13to do here. So here's uh a may uh so
42:18this is ad one pay retail not resale
42:21bought once worn for years the gift
42:23they'd pick try it on stay for coffee
42:29and see what it did for so still working
42:31on the Robin Hood ones all right every
42:34one of those if I wanted to go and make
42:36an AI avatar ad I could do that I could
42:39also have the avatar hold the product
42:41can like modify the prompt for that
42:43again you just tell it I want you to
42:45hold the product here's the image
42:46variation etc and it can do that you
42:49just provide those as the reference
42:50images and once those ads are complete
42:54so we have this final ad that's doing
42:56the read it sounds human so it's doing
42:59the read it's adding the motion graphics
43:04it's doing the popup
43:06it's looking at this frame by frame it's
43:09adding the call to action and Again,
43:13didn't do this perfect on the first
43:14shot. You can have it just like be like
43:16vision model, go through this frame by
43:18frame. You know, we need to basically
43:19like have it be aligned, etc. Um, that's
43:22something that we actually probably need
43:23to build into this. But anyway, um,
43:26we've now made those ads.
43:29>> Like, does it make sense? I mean, you
43:31could do this on Claude, you can do this
43:34on Grockbot, right? Um, so does it make
43:37sense to have a
43:41uh like creative director essentially do
43:45the first pass and be like go and like
43:48watch the what comes out as the first
43:50pass 480p
43:52and and like score it just like how a
43:57code rabbit or a reptile does
44:02>> like code review eval says like this
44:03code is ready to push.
44:05to production.
44:06>> Um, like don't don't even show me the
44:11video unless like it fixes a lot of
44:12those little issues. And also go and
44:15make sure that it it's at least an eight
44:17on 10 or a nine on 10.
44:19>> Absolutely. You can 100% do that. Um, I
44:22wouldn't even like make this person like
44:25I think people think about this as like
44:27employees, but like I would just think
44:28about it as like jobs to be done. like
44:31what was the human doing and how am I
44:33turning that into code, right? Like you
44:36don't need God in a box to run a
44:37Facebook ads account. It's actually a
44:39terrible way to run a Facebook ads
44:40account. Um how do I know? We did this
44:44multiple companies and it's absolutely
44:46terrible. Um the better way is like
44:48what's the top 1% media buyer doing? How
44:51do I go and turn that into code with a
44:53thinking loop with a live data stream
44:55that it's making decisions on top of?
44:57That's actually what an agent, you know,
44:59a live autonomous agent ends up being.
45:01But to your point, yeah, absolutely.
45:03Like you could just add a step to the
45:04workflow that I just described. I'm
45:06gonna like ship this skill literally
45:07right after this. I'm going to make this
45:09update where it basically will look at
45:11the motion graphics frame by frame and
45:13it's like, oh, here's these adjustments
45:14that we need to make so that it matches
45:16doesn't, you know, you can have a gate
45:18within there as well. Like if it doesn't
45:19meet this quality standard, then you
45:21know, it has to go through another loop.
45:23This is actually how we do static
45:25creative generation for um any like
45:28regulated markets like for example we
45:30have clients that are in financial um
45:32and for them uh like you know they they
45:35basically have all these approval
45:36processes there's things they can and
45:38can't say so we have an agent that like
45:39is a basically an image model that's
45:42like looking at this like here's all
45:44these things that we can can't do can't
45:46promise etc and it goes in a like
45:49recursive loop modifying the creative
45:52based off of that. So creative, you
45:53know, one gets made, vision model looks,
45:57yes, no, okay, send it back, make these
46:00changes, yes, no, and just goes, you
46:02know, we can typically get in like two
46:03to three. So it's it's like you don't
46:05have a lot. But Chad GBT image uh 2.5 is
46:08like an incredible graphic designer spec
46:10like for specific modifications. Like it
46:13it we've just found that it's like very
46:15good at keeping like the original image
46:17rather than remixing it again. And all
46:20these things are going to get better,
46:21right? like yeah is does a random seed
46:25dance generation like you know one out
46:27of 50 suck and it just like is still
46:29garbled nonsense absolutely will this
46:31continue to improve yes doesn't matter
46:33right now because the cost is so much
46:35like this is the discrepancy right it's
46:37like again I can pay $500 or we'll say
46:40300 maybe you get this like all you know
46:41correctly done be $300 to have the UGC
46:44creator a motion graphics person you
46:47know actually do this or in contrast I
46:50can go and pay
46:5395% is good out the other side. And the
46:55funniest part with all this is like
46:57we'll see these AI avatars sometimes
46:59perform better. Same script read
47:02>> then the human,
47:03>> right?
47:03>> And it's like people are like, well,
47:04why? And it's I don't know why. It just
47:06hits. And and that's the game anymore,
47:08especially on Facebook ads with
47:10Andromeda, right? It's like what it
47:12comes down to is like you're just
47:13feeding the beast. You're feeding the
47:14machine. It is smarter than you. It is
47:17more capable than more capable than you.
47:18I know there's more than anybody else in
47:21this world, right? Um, and it will just
47:23show the right ad to the right person.
47:26All you have to do is give it like
47:27creative velocity. I This is a lot of
47:29the work that we're doing right now is
47:30like companies are like, "Sweet." Like
47:32we were shipping 10 ads a month and
47:34we're like, "Awesome." Like, "Let's go
47:36to 300
47:38tomorrow. [laughter] We'll just ship a
47:40cohort of 30, right?" And you know it
47:43depends on the ad budget size and all
47:45these pieces but like and your ad budget
47:47and the CPL like there's like a formula
47:49for like how many ads should I ship
47:51based off of this so that it actually
47:52gets through testing in these pieces. I
47:54mean, we'll see this especially for
47:55software companies where it's like we'll
47:56ship a cohort of 30 and like an ad will
47:59get like two conversions, like two demos
48:01booked at like, you know, $50 and maybe
48:04their average is like 150 and then it
48:06will just stop spending on that ad cuz
48:08Facebook found the two people that will
48:09convert and then it's like, "All right,
48:11cool. We're done here, right? Let's move
48:12on to the other set pieces." And so
48:14again, you just have to like feed this
48:16machine and everything that I've talked
48:18about is how you actually feed this
48:20machine. like how do you give your agent
48:23the necessary resources to be able to go
48:26and do this? So now we ship those ads,
48:28right? We upload those into Facebook
48:30because I have the data solution. I can
48:33be like what is working
48:36and all of this is just code, right? All
48:38this is just code and tokens when I'm
48:40sending like here let we can do this
48:42right now but like when I'm sending that
48:43prompt to seed dance to generate the ad,
48:46right? It's just a JSON blob. It's just
48:49code that's being sent. I can store that
48:52JSON blob in a database
48:55and I have literally like at you know
48:58this cellular level breakdown of like
49:00what is this ad
49:02I can tie the data on Facebook directly
49:05to that ad that JSON blob and now
49:07imagine that for a thousand ads right
49:09that's the scale that this is getting to
49:10this is the game that's happening this
49:12is like why you have to do this this is
49:14where every company is going to be in
49:16five years like you're going to have to
49:17do this to compete
49:19>> you will not be able to win unless you
49:21do this basically have the
49:23infrastructure and then like have the
49:25engineering talent to be able to do this
49:27work. So,
49:28>> the one step I would add on the creative
49:30director piece, like what I would be
49:33doing if I was creating this and I'm I'm
49:35going to go and like build this system
49:38is have your creative director agent go
49:42and ingest like David Oglev's books and
49:46a hundred of the best, you know, YouTube
49:49podcast over the last 24 months of the
49:52world's best creative ad people and
49:55their podcasts and go and like now now
49:58you understand all the fundamental
49:59techniques. That's one. And then two,
50:02the human being, you know, going back to
50:04the swipe warehouse, how do you create
50:07like a swipe warehouse of just like tons
50:09of really cool examples and interesting
50:11examples that you're just constantly
50:14adding to? And there's a bunch of tools
50:15that you can use to manage that. You can
50:17use like I think Sublime.app is one. Uh
50:20you know, Cosmos
50:21>> if you want to make it multiplayer. This
50:23is in graph. Again, I'm going to do this
50:24plug like sweet. you want to ship a tool
50:26to your whole team. Like we have a
50:28series A startup that like five people
50:29are working on the accounts across
50:31marketing and sales and like as they
50:33discover new workflows and things they
50:34just upload it into it and it now their
50:36whole team has access to it and all the
50:38tool usage and all the data usage all of
50:41that is there. Um yes that like repo of
50:44that storage of knowledge like that is
50:46one of the most valuable things that you
50:47can do. So
50:48>> yeah. Is there anything you haven't
50:50shown that you want to show or is that
50:53it? Maybe we do another one though, G of
50:55like I come back and we'll do like an
50:56outbound version of this like how do I
50:58build a total addressable market map and
51:01how do I go and find all of their emails
51:03using a waterfall enrichment and how do
51:04I get an agent like Grockbot to go and
51:06cold email all of these people, right?
51:08What are the what's the environment that
51:10it needs to do that? We can do that as
51:12well. We can do it for SEO. Like we can
51:14do a whole series on this if that's
51:16something that the audience is
51:17interested in. Let Greg know below if
51:19that's what you want. I will come and
51:20teach you literally how to be a
51:22marketing engineer anytime that you guys
51:25want me back on the show. So
51:27>> amazing. All right. Well, thank you
51:28Cody. This has been incredible.
51:30>> Yeah. Gee, one other shout out if you
51:32don't mind. I'm in San Francisco for the
51:34next two weeks. We're raising our seed
51:36round for graft. Uh if any of what I
51:38just described resonates with you and
51:40this idea that like there needs to be a
51:42cloud for marketing engineering, that is
51:45what we are here for and that is what we
51:46were building. Reach out uh on Twitter.
51:49You can find me at Cody Schneider and
51:51you can literally ride along as we do
51:53this uh you know basically fundraiser
51:55and just turning it into an event. So
51:57cheers. Thank you man.
51:58>> Love it. Thank you, Cody.

About this transcript

This page contains the full transcript of AI Agents are too good at marketing now by Greg Isenberg, generated from the public captions YouTube serves with the video. The transcript has 9,474 words across 1,319 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 choose a caption track available on the video.

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, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.