ai automation builds 100+ ads in 24hrs - research, creative, and data analytics

0:00Everybody's talking about Vibe

0:01Marketing, but it's basically all

0:02vaporware. It's very rare that you

0:04actually see people that understand what

0:06they're doing. This is another one of

0:08those episodes that's talking to an

0:10individual that's actually building

0:11these out at scale, actually

0:13implementing these within his workflows

0:15and his process, who's an expert in this

0:16thing. I think the biggest thing for

0:18people to understand is that you have to

0:20understand marketing to automate a

0:21process, right? So, you have to be an

0:23expert at that thing to be able to go

0:25and actually build out these

0:26automations. But when you do that, you

0:27can automate 80% of the work that you

0:29previously were doing. Like all the tool

0:31sets are available. And really in the

0:33last three months has this become

0:34possible. Today I'm hosting my friend

0:35Jonathan. He is going insanely viral on

0:38Twitter right now with the marketing

0:40automations that he's building using

0:42NADN and a bunch of piec together APIs.

0:44You're going to learn how to do audience

0:45research at scale to understand how your

0:48customer talks and what pain points they

0:50have. specifically, he's scraping

0:51Reddit, pulling in all of that data, and

0:53then analyzing it for all the insights

0:55related to your audience and the product

0:57that you're trying to promote. He's

0:58going to share with us how to make

0:59creative at scale. We're going to talk

1:00about the principles of why you should

1:02be creating a bunch of different

1:03variations of creative now that for you

1:05page content is kind of ruling the

1:08world. Broad targeting and more creative

1:10is more effective at creating the

1:12outcomes that you're looking for than

1:13really defined audiences and less volume

1:16of creative. Finally, he's going to

1:17share some of the automations that he's

1:19starting to experiment with for the

1:21actual data analysis of what's working,

1:23why it's working, etc. And also share

1:24how to get started with this, where do

1:26you even begin, how do you learn this,

1:28and specifically how to use stuff like

1:30perplexity AI with Claude 4 to go and

1:32write complex NAND JSON outputs that you

1:35can then just copy and paste into an NAN

1:37to create the visualization. If you're

1:38listening to this on the podcast, I

1:40would highly suggest going and watching

1:42it on YouTube for the second half of

1:44this video. The majority of it is a lot

1:45of screen sharing, so you'll want to see

1:47that. And then finally, if you're trying

1:48to hire engineers for your startup, go

1:50to talentfiber.com to hire the best

1:53offshore talent. They place 300 plus

1:55engineers for startups. You can get

1:56engineers with US experience that have

1:59worked for the last 7 plus years, have

2:01perfect English, work US time zones, and

2:03are embedded with your team for half the

2:05cost of a US equivalent. We're getting

2:07people for 5,500 as an example, or I've

2:09seen that. And the best part is they

2:11function as a outsourced HR department.

2:13So they do all the technical interviews.

2:15They actually find you people that know

2:17how to do what they say they do. So by

2:19the time that you're even talking to

2:20candidates, it's really just a vibe

2:21check. Do you want to work with them on

2:23a daily basis, etc. Go to townfr.com to

2:25learn more. And with that said, let's

2:26get started with today's show.

2:30[Music]

2:37Jonathan, what's good, man? How you

2:38doing? What's going on? What's going on?

2:40We're here. Hell yeah. I've been looking

2:42forward to this, man. all weekend. Um I

2:44was traveling this weekend. I was just

2:45telling you about it. But um I think you

2:48know there's a lot of people talking

2:50about a you know AI automations or a

2:52like vibe marketing, right? I feel like

2:53is this like vaporware term that's

2:56thrown around right now. Nobody is

2:57actually doing this. Like it's very rare

2:59that I actually talk to somebody that's

3:01building out um you know systems and

3:03processes. And really what it comes down

3:04to is a lot of people don't know how to

3:06market first off. So they can't define a

3:08process. And if they can't define a

3:10process, how can they go and automate

3:11it? Right? Like I see all these

3:12engineers a lot of the times trying to

3:13build AI agents that do marketing and

3:15it's they're not even doing marketing.

3:17It's not even doing a marketing

3:18activity. So I'm I think to start I'd

3:20love to learn like how how did you get

3:22into this piece of this like on the AI

3:24you know doing the automation stuff for

3:26for marketing? I know your your

3:27background specifically in um uh paid

3:29ads management. Yes. So I'm curious like

3:31when you got when you started using

3:33these uh automations within your

3:34workflows like was it in the last six

3:36months? Was it more recently? I feel for

3:38us uh we saw in like the last like three

3:40months something got better that just

3:42got smarter and so it's like way more

3:44useful to use it. But I'm curious what

3:45you're seeing on your side. Yeah, I

3:47definitely I I definitely feel like in

3:48the past 3 months for sure I feel like

3:50things have really started to go I guess

3:52like even more exponential. I say I I've

3:54probably been playing around with this

3:56since the the start of the year. That's

3:58kind of like when I started looking into

3:59like workflows and ended specifically. I

4:02actually started off by hiring a

4:05different um automation team to build

4:07out a workflow for me on the paid ads or

4:11um yeah creative and advertising side of

4:13things where I wanted to kind of like

4:14automate a little bit of the reporting

4:16and also a little bit of the I guess

4:18like daytoday insights that you can pull

4:20from like ads manager stuff like that.

4:22So I actually ended up hiring them to

4:24build out a pretty complex workflow

4:26which was very helpful. Um, and that

4:30kind of like inspired me to like dive a

4:33little deeper and kind of like figure

4:35out what options are possible and kind

4:37like what we can actually do with these

4:38tools because I mean I I just saw like

4:41the I can show you later the the flow

4:42that they built and like it was just

4:44like this huge thing and I was like oh

4:46my god this is this is actually insane.

4:48Um what's what's actually possible with

4:50these tools now? Um cuz like we were

4:52pulling creative reports, we were

4:53pulling kind of like scaling

4:54recommendations, writing new ad copy. it

4:57was like this whole giant thing. So

4:59yeah, I saw that and that kind of like

5:01triggered something inside of me that

5:03was like, okay, I think this is like

5:05really something to look into. So yeah,

5:07I used that workflow for a little bit.

5:09Um, and then I just decided like, okay,

5:11I can either keep hiring someone else to

5:13kind of like build out these things for

5:15me or I can try to like double down on

5:17this myself and really understand

5:18because yeah, I feel like that

5:21understanding kind of learning what

5:23these automations can do for you now

5:24makes a lot of sense. So you you're kind

5:26of like ahead of the curve and I like

5:28always staying up to date on what's new

5:29so you don't have to like always pay

5:31someone else to do this for you. So

5:33yeah, long story short, just started

5:35like diving through like classic YouTube

5:38tutorials, walkthrough, beginner guides,

5:40joined a couple of school communities

5:42and that's kind of like how I got

5:43started with Yeah. specifically and it

5:45ended like workflows uh in particular.

5:48Yeah. any any resources on the learning

5:50side specifically like school

5:51communities or stuff like that or even

5:53YouTube channels that you feel like is

5:54are making good content like specific

5:56ones? 100%. I can uh send some over to

5:58you so you can link them later. But for

6:00sure yeah Nate Herk is is probably the

6:03main one that I've watched a lot. He has

6:04a lot of amazing like beginner like full

6:07step-by-step walkthroughs or like un

6:09like actually from start to finish

6:11learning how to not just like build

6:13workflows in N&M but actually just like

6:15understanding all the different

6:16components cuz like there are a lot of

6:18moving pieces and things to understand

6:19right so you you can't really just like

6:21hop in especially for Nadm which is a

6:24little bit more complex than like a make

6:26or or a gum loop you do kind of like

6:28need to understand a little bit of the I

6:30guess like technical uh things

6:33underneath it. So yeah, he he is uh

6:36probably the the YouTube he has a school

6:39community as well, which I recommend uh

6:41the most, I would say. Awesome. Cool.

6:44No, I'll link that for sure. Um I just

6:45found his YouTube channel. So um I'd

6:48love to hear about your stack kind of

6:50like the technical components of this of

6:52how how are you deploying this? Um like

6:53what we're doing currently is using

6:55railway.com and then just running an

6:57instance basically um for all the

6:59automations that we're building. Um, we

7:00found that to be the kind of the easiest

7:03way to do it. Basically, log in with

7:04GitHub, they give you like $5 of free

7:06credits to like get started. Um, you

7:08install N8, it's like three clicks

7:10basically to have it with workers. Um,

7:12but I'm curious what you're using uh

7:14kind of at the scale that you're doing.

7:15I know you're doing it at like a way

7:16larger scale than than the average

7:19person. So, yeah. Yeah. I mean, I'm

7:20honestly not even that like I guess

7:22fancy with it. So, for the most part,

7:24I'm just like running them inside of

7:25Naden just like trigger like just Yeah.

7:28Uh, firing. You're using their paid

7:30hosting just like their Yeah, I'm using

7:32their paid hosting and then I'm just

7:33like firing in everything like within

7:35trigger nodes inside of N. I am actually

7:37in the process now of kind of like

7:39building out custom um UI frontends

7:42using like bolt and kind like vibe

7:43coding a little bit more just kind like

7:45have a cool um interface for the

7:47workflows. I think especially like if

7:48you're going to start, you know, having

7:50other people use these workflows, it

7:52makes a lot of sense to kind of like

7:53have a a bit of a cleaner front end for

7:55them to interact with so they don't have

7:57to like go into edit and do that. So

7:59yeah, that's something I'm I'm uh

8:01working on right now as well. But for

8:03the most part or like for the actual uh

8:04flows that I use daily, I would say I

8:06just use and end for now. Nice. Nice. Um

8:10the So just to give people context of

8:13like what you can do with this, like

8:14what are your top three, you know,

8:16workflows that you're using specifically

8:18for marketing? Um I I you talked to me

8:20previously about uh your like bulk ad

8:23generator y uh system. Um, so may maybe

8:26uh talk through that and may some of the

8:28other ones that you're using most both

8:29for yourself and for clients or or what

8:30have people also been most interested

8:32in? I mean your Twitter is growing like

8:33crazy. It's wild. I feel like I every

8:35day or every other day that I see you

8:37pop up and look at it, it's like another

8:39thousand followers that follow you. So I

8:41feel like this category people don't

8:42understand like how much people want to

8:44like to do uh like AI automation for

8:48marketing like marketing uh uh tactics.

8:51Uh, so anyway, I I I'd love to learn

8:53like what you're seeing people be

8:54receptive to and also what you're using

8:55most. Yep. I guess probably a little bit

8:58just due to to my I guess like core

9:00audience being like uh advertisers and

9:03kind of like Yeah. D brand. So for me,

9:05I've seen a lot of traction with either

9:08like um creative research and kind of

9:10like avatar audience research and kind

9:12of like marketing insights. That's uh

9:14that's a big one. And the other one is

9:15obviously like creative production. So

9:17like using uh uh OpenAI's image gen API

9:20to create static ads automatically. Um

9:23I'm working on like another one right

9:25now which is kind of like a video scroll

9:26stopper. We discussed that on our last

9:28call as well to kind of like create

9:29video like hooks for uh video ads. That

9:32one is almost finished. But yeah, I

9:34would say the the top two or top three

9:36are like marketing insights uh

9:38generating new copy and new um hooks for

9:42ads. And then lastly, like creative

9:44production on the uh like image and then

9:47soon video side of things as well. Video

9:49is obviously a lot harder just due to

9:51kind of like the models that are that

9:52are out there right now. Images and

9:54static ads are a lot easier to kind of

9:56just prompt and and and get working

9:58right away. But I think soon, very soon

10:01videos are going to be like working

10:03really well too. And then it's that's

10:04going to be like a a total game changer.

10:06Obviously like with Google VO3, I mean

10:08that's that's insane. So I don't have

10:10access yet unfortunately over here in

10:12Norway but yeah just from what I've seen

10:14it looks it looks insane. So yeah once

10:16like you're able to integrate those

10:17workflows in like an ended end as well I

10:20think it's going to be it's going to be

10:21crazy. But yeah I would say yeah um

10:23audience research and um creative like

10:26production are the top two kind of like

10:28segments that I'm uh seeing a lot of

10:30people interested in for me at the

10:32moment. No, that makes total sense too,

10:34right? It's like historically, you know,

10:36when I was doing uh paid ads management

10:39for for companies at scale, like I mean

10:41we try to go and make like a 100 plus

10:43variations of create actually, it's

10:44funny, I'm doing this right now actually

10:45for a mobile application. Um we're just

10:48doing like AI avatar variations with

10:50like different hook, different painpoint

10:51combinations. Um and then using the hey

10:54gen API basically to like go bulk and go

10:56bulk generate all these videos. Um, but

10:59the, you know, anymore these ad channels

11:02are so effective at finding the right

11:04people if you have like the correct

11:06conversion event set up. So like what

11:07we're doing is basically having a

11:08conversion event on the free trial start

11:11uh um uh user action like in the mobile

11:14application, sending that event back and

11:16then really your job as a a marketer or

11:18an advertiser is just like finding the

11:20creative that actually gets that event

11:22to happen as cheaply as possible, right?

11:24So your game turns into like creative

11:26development and testing rather than uh

11:29you know audience definition or

11:31understanding which I think is a for me

11:34it's been a mind switch. You know I I

11:35I've been doing this 10 years so it's it

11:37I'm from the the um you know the vintage

11:39of of marketers with Facebook ads where

11:42it's like super complex you know

11:44targeting that was the only way that you

11:45would actually get better. Totally. Um

11:48but anyway you you mentioned something

11:49uh a minute ago and I just want to come

11:51back to it. uh this idea of like

11:53connecting frontends to N8 end flows.

11:56I've seen people doing this where they

11:57like use lovable basically to build like

11:59the front end uh kind of application. I

12:02don't know what you would call it or UI

12:03of the of the flow and then N is

12:06basically the the backend uh that's like

12:08actually doing the functions. Could you

12:10talk about that more and kind of like

12:12how you're setting that up, how you're

12:13using it? Yeah, for sure. I mean, again,

12:15it's not something I've done uh a lot

12:17yet, but um what I'm basically doing

12:20right now is is pretty much exactly what

12:21you said. So, you have a working

12:23automation inside of Nen. Um and then

12:26instead of like so like yeah, for the

12:28most part like I I uh fire all of these

12:31flows like on a on a like a form submit.

12:33So like for example for the for for

12:36Reddit scraper like you add in the

12:37keyword that you want to scrape and then

12:39you send that over to like Reddit's API

12:41scraper tool and then and it uh scrapes

12:44entire um yeah post based on based on

12:47the keyword but instead of like instead

12:49of firing the flow based on a form that

12:52you have inside of nen you can kind of

12:54like have a uh a UI inside of a lovable

12:57or or bolt which I'm using and then kind

12:59of like have a custom UI where you have

13:01like the form there instead and then you

13:02pass the the keyword that you submit

13:05through bolt over into n and then that

13:07triggers the flow instead. So it's

13:09basically just kind of like beautifying

13:11the the flow a little bit. But I mean

13:13there's tons of there's tons of uh I

13:16guess like customizations and ways you

13:17can kind of like

13:19uh visually present the output as well

13:22which I'm also interested in because

13:23like there's obviously you're kind of

13:25limited inside of edit in and kind of

13:27like how you I guess present the the the

13:30data or the output that you get. For the

13:32most part I'm just like using a Google

13:34sheet for kind of like the output in

13:36terms of like uh yeah the like copy or

13:38marketing insights or whatever. But if

13:40you pass that data back to like a Bolt,

13:43you can present it a lot, I guess, like

13:45cooler, if that makes sense. No, 100%.

13:48Yeah, I I I'm just thinking about it is

13:49like, you know, if you're doing bulk

13:51image generation for uh you know, image

13:53ads, pulling that back into a dashboard,

13:55you know, maybe there's some type of I

13:57am working I am working on that on uh

14:00for a video uh flow right now. It's not

14:02done yet, but I can see No, I don't have

14:04that saved. I can show you the one that

14:06I'm building a little bit later if you

14:08want for like a Tik Tok marketing

14:10insights where you basically just like I

14:11said you you submit the um you submit

14:14the the keywords uh and the amount of

14:16videos that you want to scrape and then

14:18you give a product uh description and

14:20then it's it scrapes the videos based on

14:21that like pulls pain points trigger

14:24events and rewrites the script from the

14:26scraped videos and then sends everything

14:27back. So I can show unreal. Absolutely.

14:30Yeah. I think the like second half of

14:31this will just go through all the flows

14:33that you built because I think for sure

14:35for a lot of people it'll kind of give

14:37uh you know really a jump point of like

14:39what's possible or like what people are

14:41thinking about with this and I I think

14:42that's the you know the biggest hurdle

14:44right now currently for marketers is

14:46like they know it exists but they don't

14:47really know what are the actual like

14:48tactical implementations of this um and

14:51so anyway yeah I think second half of

14:52this video we'll just have you screen

14:54share and talk through kind of some of

14:55the automations but sounds good um cool

14:58okay so companywise I the talk to me

15:02about how you're implementing this

15:03within your business. So you do paid ads

15:05management for e-commerce brands from my

15:07understanding. Um hit you know what were

15:09you doing before and then like what has

15:11this solved now like why you know why

15:13why are why why are we even talking

15:15about this today? Yeah you know what's

15:17the purpose of the implementation within

15:18the business for you know anybody that's

15:20listening. So I mean like you said with

15:23the with the image gen thing that you're

15:24working on like I mean obviously just

15:25like volume creative volume right now is

15:27huge with advertising in general just

15:29like the more you can feed the algo and

15:30that's just going to get get bigger or

15:32more and more important. So just the

15:34more I guess like yeah creative output

15:36and variation that you can generate the

15:38the more you can just feed that back

15:40into Facebook and obviously like you're

15:42not going to be able to generate all

15:43that manually yourself. So being able to

15:45just have these tools that can pretty

15:47much just like 10x 100x your own manual

15:51output is kind of like how I'm seeing

15:52it. So it's it's kind of like just

15:54having an extension of yourself and um

15:57yeah just maximizing output and also I

15:59mean obvious like obviously just saving

16:01time, right? And just like having having

16:03these workflows run 247 automatically

16:06just spit things out inside inside of

16:07these work uh inside of these sheets and

16:09inside of these databases. So you don't

16:10have to do all that manually. you just

16:12have a compilation of of sheets and

16:14databases and things that that are

16:16presented to you on like a daily or

16:18weekly basis, whatever, and then you can

16:19just kind of like skim through

16:21everything, take what you need, and then

16:22just run with that. So, yeah, for me it

16:25and again, I think we discussed this on

16:26our last call, too. I mean, also right

16:28now, I'm not even like super super

16:30worried about the actual output of these

16:33agents in a lot of cases because I mean,

16:35sure, they're somewhat limited in terms

16:37of like the the tech and the output like

16:39obviously in terms of especially like on

16:41the video side of thing for creative

16:42like it's not amazing yet, but it likely

16:47will be and it will likely be pretty

16:49soon, right? So I'm also kind of like in

16:51the mindset of like getting ahead of the

16:53curve right now and understanding the

16:54logic and the and the tools and kind of

16:56like how to build it these yeah these

16:58agents or these workflows. So when the

17:01output kind of like goes really

17:02exponential then you're already like

17:04really good position to kind like really

17:07um take advantage of the flows. Totally.

17:10Yeah. I think the thing that people get

17:11hung up on right now is like maybe it's

17:13not perfect, right? But like then your

17:15job all it should be is like a curator,

17:17right? rather than being the, you know,

17:19the grunt labor for the production. I

17:20mean, I think about a marketer, right?

17:22Like it, and again, everybody has done

17:24this where it's like, okay, I want to

17:25make 10 variations of ad creative, like

17:27that's a slog to go and generate like

17:29each of those variations. Um, the the

17:32real power of this is like, okay, cool.

17:34If I can have every, you know, all the

17:36hard labor or the heavy lifting done and

17:38then my job is just to come in and like

17:40have, I guess, like, you know, taste or

17:43preference like that. That is the real

17:45opportunity here with all of us.

17:46Exactly. Um, and you know, maybe gets it

17:49only 80% of the way there and then you

17:51have to have a human in the loop to get

17:52it the last 20%. But a lot of the times

17:54I mean there there's so much time saving

17:56with that. I mean for example like we we

17:57have agencies uh like one of ours is a a

18:00SAS SEO agency right historically if I

18:03was doing uh you know any type of SEO

18:06services like content writing was the

18:08the most time consuming aspect like

18:10researching the article and then doing

18:11the content writing was the most

18:13timeconuming aspects of the whole

18:15process. Like now we have an automation

18:17that basically, you know, takes the

18:19target keyword, scrapes what's currently

18:21ranking on page one, and then writes the

18:23first draft of the article based on

18:24that. I mean, it's 90% of the way there.

18:27We have a human coming in over the top

18:28of it just to like edit it for dwell

18:30time so that it's like basically just

18:32more readable from a human standpoint.

18:34But that, you know, that process that

18:36used to take like a day to write an

18:39article, as an example, it now happens

18:40in, you know, 30 seconds. And I mean, we

18:44can get a hundred of those out the door,

18:46right? That in in the time that it used

18:48to take to only get maybe 10 of them or

18:51even less, maybe five of them. So, I

18:53think that's like the mindset to

18:54approach the Swiss. And and the other

18:56thing to piggyback on that with is we're

18:59not doing anything different, right?

19:01Like you're not changing how you're

19:02running your ads accounts or like how

19:04you're doing your your your creative

19:06development in any way. The only thing

19:08that's changing is the automation of

19:10those processes that already exist,

19:12right? So, it's like not a reinvent.

19:14We're not reinventing the wheel or any

19:15any uh big uh I don't know. It's not

19:19like this. How I think about it is we're

19:21not replacing anything that we were

19:22doing previously. Um we're just like

19:24augmenting it, right? And so anyways, I

19:26I think that's things that like people

19:28get hung up on where they're trying to

19:29like find these like new creative ways

19:31to like use this like because and that

19:33frontier is totally there. like there

19:35are people that are probably going to

19:36invent uh these new use cases but for

19:39the majority of companies it's just like

19:41what are you doing currently what can

19:43you automate of that and for a lot of

19:45companies that's like 80% of their human

19:47labor they can go and automate with

19:48these tools especially on the marketing

19:50side and I mean that's immediate cost

19:52savings right like we can I've seen I' I

19:54I have friends that run agencies that

19:56their their margin was like 30% and

19:59started implementing these AI

20:00automations and tools etc and I mean

20:02they took their margins to like 70%

20:04overnight just because they like

20:05automated that human labor. So I think

20:07again just that's the way to think about

20:08this if you're you know a founder

20:11listening to this ahead of growth

20:13listening to this etc. So um with that

20:15said man I let's just jump into it. I'd

20:17love to see some of the workflows some

20:19of the ways that you're you're

20:20implementing this. Um okay and uh yeah

20:23I'll just have you screen share and

20:24maybe we just like talk through the

20:25automations and and uh uh uh show have

20:28you show uh the audience. Let's do it.

20:31I'll pull up a couple of flows here and

20:33then we can get started. Let's

20:35see. Um, let's do let's start with the

20:38Reddit one because we already discussed

20:40that a little bit. Let me just pull that

20:41up right here. I'll see if I can find

20:43that. Yeah, here we

20:46go. I think we should be good now. Yep,

20:49I can see that. It's perfect. Cool.

20:51Yeah. So, this is a Reddit marketing

20:54insights um flow that I built inside of

20:57Nen. So yeah, basically I can just

20:59quickly go through the flow high level

21:01and I can just kind of like show you the

21:02output here as well later. I'll just

21:04pull this up first. So I just have it

21:06ready,

21:08but we basically again so we start with

21:11a with a trigger to kind of like fire

21:14the entire workflow. So in this case I I

21:16do it manually inside of Naden as I do

21:18with most of these flows that I

21:20mentioned. So this one is fired based on

21:22a form submission. So here I basically

21:24just enter the brand name. I entered the

21:27website and a product

21:30explanation and then where is

21:33the where's the

21:36keyword? Let's

21:38see.

21:41Oh, that's cuz I have a pin data here.

21:44Let me just unpin this and I'll show

21:45you. Oh, sorry. This is based on the

21:47product explanation. Sorry. So if I just

21:49do let's just do

21:52try

21:55drink.com and then

21:58alcohol

22:00alternate

22:01alternative

22:04beverage. Sorry I have so many flows. I

22:07think this one basically descripes

22:08Reddit for pain points based on that

22:10like on a on a product description or on

22:13a yeah like an uh like for a certain

22:15niche if that makes sense. Totally. Yes.

22:18So, it's running in the background.

22:19Basically, it went to Reddit. It

22:21identified, you know, all of the Reddit

22:23conversations that happened around Yeah.

22:25probably alcohol alternative. I imagine

22:27that's probably like the main keyword

22:28that it ends up pulling. Correct. Um,

22:31pulls that context back into the context

22:33window of the AI and then at that point

22:36then, you know, the writing process

22:37basically starts occurring. You're just

22:38hitting a chat GPT endpoint to actually

22:40do that writing. Um, just again,

22:42narrating's for the the audience that's

22:44only listening.

22:46But the uh also if you're listening only

22:48to this, go to YouTube. It'll probably

22:49be more impactful for you. So then after

22:53it completes that writing, what are

22:54those following steps that are occurring

22:56after that? So yeah, I can I can just

22:58walk through everything. So basically

22:59yeah what we said here is that we we

23:01send everything off to a uh open AI node

23:04first uh which is basically like a chat

23:06GPT node and we tell it that it is a

23:08marketing strategist and Reddit

23:09researcher and based uh on the product

23:12description that we pass through the

23:14form right here. You can see that this

23:15is kind of like a dynamic keyword. So

23:17yeah whatever we add in the in the form

23:19is kind of like what we um research

23:22right here. So we basically yeah now I

23:24remember the entire flow. So we

23:26basically explain our product

23:29uh description and then based on that we

23:31have um the actual AI spit out the

23:34optimal keyword to scrape rider that's

23:37how I built this for I remember now so

23:39basically this one so I should have

23:42basically explained uh this a little bit

23:44different in the in the form but in this

23:47case uh the AI meant that we should

23:50scrape ready for quit drinking that's

23:51the keyword that we wanted to scrape so

23:53you basically just send uh the keyword

23:56quit drinking to the Reddit node and

24:00we're then we're scraping Reddit for

24:02that specific keyword and then topics

24:04and posts around that. Then what what I

24:06want there's a specific Reddit node on

24:08NAN. Sorry, I haven't seen this before.

24:10There is Yeah, there is. It's amazing,

24:12man. Yeah, it's crazy. So, you can do

24:14this or you can kind like do it through

24:16like HTTP request nodes as well

24:17obviously, but yeah, I just like

24:19perplexity or whatever. That's what I've

24:20done previously is like basically a

24:22perplexity call to their API to pull

24:25back the you know relevant Reddit

24:27threads and then we put that whatever's

24:29on there like into the context window.

24:30So yeah. Yeah, for sure. But yeah, they

24:32haven't they have native ones as well.

24:34So this one works just fine. So yeah, I

24:36basically just wanted to scrape 10

24:38posts. We then filter those posts by uh

24:41I guess like by virality or popularity.

24:43So I basically just filter by if the

24:46post uh has two up votes or more and the

24:50the basically if the post has a text. So

24:52it's just not like a like a empty

24:53headline post or whatever, right? So

24:56yeah, you can basically just filter. So

24:58I I pulled just 10 posts right here, but

25:00you could pull like 100 a thousand and

25:03then you can filter based on like

25:04certain amount of uploads or certain

25:06amount of comments just so you know

25:07you're getting like super hot post or

25:09super like viral post. Um, but yeah, you

25:12can customize this however you want. We

25:14then send this over into a different AI

25:17node which basically just like um ranks

25:21if the all of the posts that we scraped

25:23are relevant for our keyword or product.

25:27So we basically stack ranks the uh the

25:30posts or sorry stacks ranks the

25:33information that it scraped against each

25:35other like the ideas basically against

25:37each other of which one's the best based

25:39on our product description. So like for

25:41for an alcohol alternative, it it kind

25:43of like it it analyzes if these posts

25:45are actually like relevant to that

25:47pretty much. Interesting. Okay. So it

25:51does that and then what happens next

25:52after that? Then we pass everything

25:55through to the um the main AI agent

25:58which turns these scrape posts into

26:01market insights. So here we have a

26:03prompt that um tells the agent that it

26:06is a senior marketing strategist helping

26:08DDC brands extract powerful messaging

26:10insights from patch of Reddit post. Then

26:13we basically just send over all the

26:15posts and then we give it the the

26:17product uh description right here and

26:20then we basically have it uh analyze the

26:22overall conversation and return the

26:24following structured insights. Top three

26:26pain points, trigger events, um desired

26:29outcomes, interesting quotes or phrases.

26:31So, like if there are any like really

26:32cool phrases inside of the post that we

26:34could potentially use. Um, ad copy

26:37hooks, marketing insights, and any

26:39trends or opportunities. And then you

26:41can see right here, it pulled everything

26:43out on the on the right right here. So,

26:45top three pain points, realizing alcohol

26:47was limiting their potential as success,

26:49struggling with social situations

26:51without alcohol, um, experiencing

26:53negative physical and mental health

26:55benefits. Um, interesting quotes. I

26:57thought I was high functioning, then I

26:59quit drinking. life can be so much

27:01better than you even imagine. I only

27:02thought that I was high

27:04function. Yeah. Yeah. It's super cool.

27:08And then pretty much I just um yeah, I

27:10send everything out to a Google Sheets

27:13that has everything right here. So yeah,

27:16you have the the pain points, the

27:17trigger events,

27:19outcomes, quotes, content ideas, ad copy

27:23hooks, and then marketing insights, and

27:24then trends. Then you can just literally

27:26run this like on a weekly basis, every

27:29single day if you want based on multiple

27:31keywords. Then you can just come in here

27:33and just see like okay like are there

27:36any like maybe trigger events that we

27:38haven't considered for like our audience

27:40that we could discuss or are there any

27:42like cool outcomes here that that might

27:44be that we could position our product

27:46towards that we haven't already. So, I

27:48see this flow as kind of like a way to

27:50maybe like open up your TAM or kind of

27:52like tap into new or different audiences

27:55that you haven't maybe considered based

27:58on like actual data and post that people

28:00are talking about. Same like with quotes

28:02as well, right? I mean, like obviously

28:05like your customers are your best

28:06advertisers. So, like taking their exact

28:09wording and phrases is is is for sure

28:11going to be an effective marketing

28:13strategy a lot of the time. So, 100%.

28:15Yeah. I mean, you're mirroring the

28:16language that they're using and

28:17describing the pain point that you're

28:19solving, right? And I I I mean, that's

28:21one of the most powerful and the also

28:24the hardest things to do like

28:26historically is like how do I identify

28:27the language that my target customer

28:29uses to describe the thing that, you

28:31know, my product actually does. I see

28:33this with startup startups all the time

28:34where they like they call it something

28:36different than who they're actually

28:37selling it to, right? You know, when

28:39they're still trying to figure out kind

28:41of their their go to market motion. So,

28:43I feel like this is just like a it's

28:45basically a cheat code to to get to to

28:47get to that messaging faster on how do

28:49we even describe the product? How does

28:50it fit into the market? Um, based off

28:52of, you know, the conversations that are

28:53happening. Are you just using Reddit for

28:55this or are you using any other sources?

28:56I know some pe I saw people using Kora

28:58in the past. I saw people trying to do

29:00this with Twitter as well, like d you

29:02know, the conversations or dialogue that

29:04are happening around the platform. I

29:05actually do I do have one for both

29:08Twitter and for I actually just finished

29:10a a Twitter one this morning actually.

29:12So this one is a little bit more basic.

29:15This one is just pulling

29:18um like just top tweets or viral tweets

29:21and kind of like the the stats like how

29:22many likes and how many how many um yeah

29:25retweets and comments and stuff like

29:26that. But you could easily add in

29:28another step here to kind of like

29:30analyze the the scrape data as well and

29:32then kind like have the output the same

29:34way. But for this for this specific

29:36flow, I'm basically just scraping X

29:38based on a keyword that I set. And then

29:40I just choose if we want to scrape top

29:42or latest post. It scrapes. Uh are you

29:45doing this for your own uh content uh

29:48creation? Yeah. So top automation tweets

29:50from the last week. Uh this influences

29:52the ones that I go and create. Yeah. If

29:54you just if you literally just take a

29:56look here, I I've been using this um

29:59this sheet right here for both not only

30:03just like um ideas or topic ideas to

30:06kind of like create content around, but

30:07also just like how to structure the

30:10post, right? Because like how you

30:12obviously like how you write the post

30:13itself has plays a huge part in in like

30:16the reach and virality. So yeah, I

30:19basically just come in here. I see like,

30:20okay, this post's got 152,000 views and

30:23like 2300 likes. This one got 22,000

30:28likes um and so on. And then I structure

30:31Yeah. Again, kind of like how they write

30:33it. So, yeah, this has definitely been a

30:34huge part of um uh yeah, me growing on

30:38Twitter as well. Just like literally

30:40modeling what I see other people working

30:41here. Totally. So this is my strategy

30:44for Twitter growth like that I've used

30:45over the last two years is just

30:47basically like write a bunch of content

30:49look at what's working go and write more

30:50like the best performing you know all

30:52social media is just that right it's

30:54like figuring out what resonates with

30:56your audience etc. Um, what's

30:58interesting though is like you're do you

31:00could do this both for what's already

31:01gone viral on the platform, your your

31:03internal I mean what it will probably

31:06turn into as well is like you'll start

31:08to see like week overw week trends with

31:10like what is like the idea of that

31:12moment like the zeitgeist of that moment

31:14and you could be basically be like okay

31:16cool like um this you know last week

31:18this idea or you know what are the top

31:20three ideas or what are the common three

31:22ideas from all of this source material

31:23you could basically take all of these

31:25tweets put that into claude and tell it,

31:27you know, ask it to give you the

31:28insights of like the common uh concepts,

31:31right, or formats that are working. And

31:32then based off of that, uh, you know,

31:34that can influence, okay, we're going to

31:36only focus on the ones that are the, you

31:38know, that that are have the most

31:40virality based off of the source data

31:41that we provide. So, right, dude, I I

31:44love this man. Okay, so how are you

31:45pulling from Twitter? Is it just

31:46directly from the API for this data or

31:48are you doing Okay, I'm using a uh tool

31:51called Twitter API.io. So, it's they

31:54have a free account, so I think you can

31:55scrape like a thousand tweets or

31:57something like that. Um, yep. So, you

31:59just literally just sign up, grab their

32:01API key, hook it up to um Yeah. Uh, any,

32:04and then you just run through the entire

32:06flow right here. Totally. Oh, man. It's

32:08so cheap. It's like 15 cents per

32:10thousand tweets is what I'm saying.

32:11Yeah. It's crazy. It's crazy cheap. So,

32:13insane. And you can get a lot of data

32:15for like I mean, literally next to

32:17nothing with a lot of these tools. So it

32:19doesn't even have to be like super

32:21expensive to to run these flows, which

32:22is which is awesome. Totally. We use

32:25Rapid API and Apify a lot for all of our

32:27like kind of third party API scraping.

32:29Um it's crazy what you can scrape with

32:31Ampify, man. I My brother's looking for

32:34a house right now. And so he's just like

32:35using uh like a Zillow scraper API

32:38basically from Appify to go and pull all

32:41the home data for him. And then that's

32:43awesome. Pretty wild. Um uh but anyway,

32:46so okay, so this one amazing. I feel

32:48like the research component makes total

32:50sense like where like how do I aggregate

32:52data, organize it, um understand what's

32:55working based off of uh the impression

32:57data, the engagement data, etc. Okay.

32:59So, once you've gotten that insight,

33:01what what what do you do from there?

33:02Kind of what's your what are some of the

33:03flows that you use to actually create

33:05the the content or the output? So, I

33:08don't really have a full end to-end

33:10workflow that kind of like combines

33:11everything yet because yeah, I think it

33:14I think at least for right now, it

33:16probably makes sense to keep uh the the

33:19flow separate. Yeah, siloed off and then

33:21you can kind of just combine the data. I

33:23did actually create this one float that

33:25was based off of your tweet actually.

33:28Um, let me see if I can find that. Which

33:30is the uh, yeah, the Reddit social

33:32listener to Let me see if I can find the

33:35actual Is it Let me see. Was it the uh,

33:39uh, to create basically like a Yeah, the

33:41comic book style. Yeah, totally.

33:43Exactly. Yeah. So, yeah. So, because I

33:46at that point I I was already using the

33:48the Reddit flow that I had built, right?

33:50So, I basically just tagged on the the

33:52image gen part and then I created the

33:56entire flow. So again, this one we we

33:59scrape Reddit uh based on based on a

34:02keyword that we give it. We again we we

34:04rank if the the post is relevant for our

34:07product description. We then turn those

34:09pain points into messages. We rank the

34:12top 10 messages and then we use those

34:14top 10 messages to create like actual

34:16image prompts that we send to ChatBT's

34:19um OpenAI image den and then we create

34:22those comic book ads. So, if I pull up

34:24my folder right here, I hope I still

34:27have a bunch of them in.

34:31Uh, no, I think I deleted them. But

34:34basically, you can create those comic

34:36book style ads that um totally that you

34:39were talking about. So, this is probably

34:40the closest that I've come to kind of

34:42like combining Yeah. an end to end

34:44solution. Yeah, exactly. Exactly. But

34:46other than that, like, yeah, I'm keeping

34:48things a little bit more siloed for now,

34:50just for simplicity's sake, to be

34:51honest. Totally. No, that makes sense.

34:54So, so you've done this research, you

34:56found your insights. Um, at that point,

34:58um, like where where are you going? Uh,

35:00do you go and, you know, bulk generate

35:02with the Open AI, uh, API image calls?

35:04Like what flows are you using? And could

35:06you talk through some of those, um, that

35:08you're seeing be effective right now?

35:09Yeah, let me pull up an image gen one.

35:12Let's see.

35:14So, I imagine when you're breaking this

35:15down, it's like research, the actual

35:17like creative creation.

35:19Um, and then maybe like the

35:23the really it's probably like the data

35:25analysis of what's actually working. I

35:27know you talked about or the

35:28conversation we had previously is like

35:30you were pulling all that data in from

35:32the Facebook ads API and then basically

35:34being like, okay, you know, the ones

35:36that are winning, what do they have in

35:37common? That type of thing.

35:38I'm thinking just kind of next or next

35:40flows I'd love to see is the actual, you

35:42know, the creation of the content and

35:44then if there's any of that analysis

35:46part as well. Yeah, let me pull up that

35:48the the the main one right here. Let's

35:50see where is where is that

35:53one. So, no, that's not that one. Let's

35:56see. I have so many flows in here right

35:59now. I need to organize things. No,

36:01you're all good, man. You're all good.

36:02This is how you know you're like a real

36:04practitioner. It's cuz you have a

36:06thousand of these that are just like

36:08either half baked or or fully done that

36:10you just like we're just we're just

36:12cooking in a lab at this point. So I

36:14love it, man. This one is pretty cool.

36:16So, so yeah, I mean, well, first of all,

36:19just to be like completely transparent,

36:21like I would say tools like triple whale

36:23do this a lot better than what what you

36:26can do inside of edit end because just

36:28to be completely frank like working with

36:30and I mean we discussed this on our last

36:31call like working with Facebook uh

36:33Facebook's API right now is a headache

36:36to be honest. So literally just p just

36:38pulling data from ads manager and kind

36:40like performance is is pretty difficult.

36:42I had to actually bring in a like an

36:45actual expert to help me complete this

36:46flow because I couldn't figure it out.

36:48Yeah, you just have to dig through like

36:49so much documentation and kind of like

36:51just really dig through things. So, um

36:55for for for paid ads uh analysis

36:58specifically, I would say that that's

37:00probably the the the last thing I'm

37:02using Nate for right now. But I can work

37:05you walk you through this one which

37:06basically just shows that it is actually

37:08possible uh inside I think that'd be

37:10super interesting just to kind of like

37:12you know plant the seed of what what

37:13this is capable of. Um are you using is

37:16it a pre-built node that's in it in to

37:18pull the Facebook ads uh data or you

37:20doing an HTTP like a a call. It's a it's

37:23a mix of both. It's a mix of both. Yeah.

37:25Okay. Cool. So, so basically this flow

37:28in a in a high level we we pull

37:30performance data from ads manager on a

37:333-day basis. So every 3 days we pull the

37:36data. So for this for yeah for for

37:39example for an ecom brand like we're

37:41pulling spend purchases um yeah CPA rows

37:45all all the main numbers that we're that

37:47we that we're uh optimizing and

37:50analyzing. Uh we basically just all of

37:53these steps are for like pulling out the

37:55individual kind of like metrics and also

37:58combining them with the with the

38:00creative ID. So like the actual

38:01creatives that are generating those

38:03results. We pull all of this together.

38:06Um and then we basically send this over

38:09into we have four different agents here.

38:12We kind of like have the the leader

38:13agent which is kind of like the

38:14orchestra and then we have three sub

38:17agents that kind of like report to the

38:19main one. We have a performance analyzer

38:22agent. We have a deep research agent.

38:24And then we have a new ad creation

38:27agent. So in a nutshell, what this agent

38:29does is that it again it scans um ads

38:32manager performance on a 3-day basis. So

38:35every 3 days you you this flow runs uh

38:38it scans uh campaign adset and ad

38:42performance. It pulls all of the running

38:45ads. it and then it sends all of that

38:48all of that data over to the performance

38:50analyzer agent. Uh once that's done, we

38:53pass it over to the deep research agent

38:56which basically analyzes the content of

39:00the winning ad. So like the ad copy um

39:03the headlines like the the if it's an

39:05image or if it's a video and uses uh

39:08perplexities deep research to kind of

39:10like analyze why those specific ads are

39:13working or not. So kind of like yeah

39:15desires uh trigger events pain points

39:18etc in the copy and kind of like

39:19analyzes why it's working. It then sends

39:22all of that over to the third uh sub

39:26agent which writes new ad variations

39:28based on all the data and then we spit

39:32the final output to a slack message. So

39:35I can share with you what that looks

39:38like as well here. So, if I just see,

39:42are you able to see my Slack now or no?

39:45Uh, no. I can't. You might need to

39:46reshare. Yeah. Yeah. Yeah. Let's see.

39:49Let's do screen. Let me just do Yeah.

39:54So, this is kind of like how we we did

39:55the output. So, we have

39:57like creative insights drop April 16.

40:00Um, we have the top performing ads, we

40:02have the ad ID sent, we have the format,

40:04we have the headline, the body copy, and

40:07then we have like the the main kind of

40:08like KPI. So like we have spend,

40:10purchase, CPA, CTR, and then rorowaz for

40:13the top three ads. And then we analyze

40:16what's working across these ads. So like

40:17the the messaging, the voice, the social

40:20proof hooks. uh we analyze kind of like

40:22trigger events and behavioral insights

40:24with uh again yeah as I mentioned

40:26perplexity uh the core desires uh

40:30psychological framing of like why these

40:32ads convert and then the agent also

40:35spits out three new variations to test

40:38and then call it kind of like a like a

40:40yeah task mockup of like what you should

40:43consider doing based on all the data

40:45right here. That's amazing man. Yeah.

40:48Yeah.

40:49Yeah. So you can go like I mean you can

40:52go super deep with this if you want.

40:53Yeah. Totally. Totally. Totally. I feel

40:55like the data analysis part is the

40:56hardest part of this. Yeah. Yeah. It's

40:59that that's still the human. Totally.

41:01Yeah. Um and even like using tools like

41:03like triple whale now as well like even

41:05they are kind like struggling a little

41:07bit because like it's it's tough to kind

41:09like especially for like for an ads

41:11manager like understanding the nuance

41:13between like what a top performing ad

41:15looks like. So, for example, like let's

41:17say we have a top performing campaign

41:19that's scaling up, but the CPA is a

41:21little bit higher than on like a

41:22retargeting campaign, right? The the LMM

41:25might rank the retargeting campaign as

41:27better performing because it has lower

41:28CPA, but it doesn't really understand

41:30that it's a retargeting campaign and you

41:33can't really scale it up, right? So,

41:34it's really tough to kind of like insert

41:36those small nuances into into yeah,

41:39these workflows. So yeah, I think I mean

41:42obviously it's it's coming, but for now

41:44it it's probably better to kind of like

41:45use it for like an like more analysis

41:48and then you like take the analysis and

41:51data and execute on that instead of like

41:53having the the the agents or the LMS

41:55kind of like suggest what execution you

41:57should do based on the data if that

41:59makes sense. No, totally. Yeah, I I feel

42:01like it's that human in the loop

42:03component is still like a key aspect of

42:05this. 100% 100%. Um okay, cool. uh we

42:10have a little bit of time left. So,

42:11wanted to uh dive into the actual um uh

42:14like media creation uh process or the

42:16the flows that you're doing to to to

42:18make Facebook ads, you know, make video

42:19ads, any of those that you're seeing for

42:21sure. Which ones are you using uh like

42:24most right now or that you're seeing the

42:25most success with currently? So, I mean,

42:28I'm using all just um Open AI's image

42:31gen at the moment for AI uh production.

42:35Uh, in terms of like video, it's still

42:38not there yet to be completely honest. I

42:39mean, sure, like with VO3, but it's

42:42super expensive, but in terms of video,

42:44it's it's just not there yet from what

42:45I'm seeing. Um, so that's still

42:48something that isn't quite there yet.

42:51But in terms of uh image production, I

42:54do have a flow that I can pull up and

42:56show you. Just got to find the right

42:59one. Yeah, here we go.

43:02So in this flow we

43:06basically this is for creating like uh

43:10iterations of a winning ad that you

43:13have. Right? So again we basically just

43:16open up or fire off the the the flow

43:20with a form trigger where I add in the

43:22the brand name, the brand website and

43:26kind of like the the winning ad, right,

43:28that we want to create new variations

43:30of. we we fire up the flow. The first

43:33thing that we do is basically just

43:34upload the the the winning ad. That's

43:37just so we can easily reference it later

43:38in the flow because uh yeah, a little

43:40bit technical, but right now open it or

43:43sorry, nod isn't great at kind of like

43:45pulling images from previous nodes in a

43:48later step. So that way we just uploaded

43:50it to a Google Drive just so we can

43:52always reference the the the winning ad

43:54or the the reference image whenever we

43:56we need to. So it's like local storage

43:59is is what we've seen. So we we've

44:01been using a a um a Google uh script

44:05function. So basically a Google app

44:07script as the way to to um do that that

44:10pass. It's it's been it's been effective

44:12for us to kind of like bypass the

44:14limitations of NAD. So like leaning on

44:16Google Cloud basically as a way to do

44:17it. So anyway, just a random learning

44:19for you. Love it. Yeah, that's great.

44:20That's great. I mean yeah that's that's

44:22a great workound. So yeah I mean yeah

44:25this is like super simple. So you just

44:27basically you just upload the image ad

44:28and then you just redownload it. So we

44:30just have it handy for reference. The

44:33the next thing we do is we basically

44:34just have a uh open AI node that uh

44:37pulls a uh description of the actual ad.

44:40So we basically just tell it to describe

44:42the visual style, subject matter,

44:45composition of this image. Is it a

44:47lifestyle image, product shot or a

44:48combination? Include lighting style and

44:50camera angle if possible. So we

44:52basically just pull that through. Next

44:54we pull branding data. So we basically

44:57just um have the uh the next node uh act

45:02as a visual brand strategist and art

45:05director uh and that we have it analyze

45:07the brand website and focus only on kind

45:09of like the brand's visual aesthetics.

45:10So like color pletes uh photography

45:14style imagery themes mood and so on. And

45:17then we just again pass the brand

45:18website and the brand name. So the

45:20entire flow has that.

45:23Then we start sending all of this over

45:25to the main agent which will be creating

45:28the actual prompts. So here we basically

45:30just again we pass through all of the

45:32data from the form. So we pass the the

45:34brand name, the website, the image

45:36description and the visual style

45:38overview which we pulled from the two

45:40previous notes. So it has all of that.

45:43And then in the bottom one, this is

45:45basically Sorry, my dog's going crazy

45:47here. You're all good, man. We this note

45:49we basically have like the the main

45:51rules for kind of like the the output

45:54that we have the entire flow build,

45:56right? So yeah, we basically tell it to

45:58generate 10 um uh tightly related visual

46:02variations of a reference ad, not new

46:05concepts. So yeah, I I won't read

46:07through the entire thing, but basically

46:09this is kind of like where you can play

46:11around with the with the prompts to kind

46:12of like get the output that you want.

46:14So, let's say that you instead of

46:15wanting to create um just uh small

46:18variations of an image ad, you want to

46:20create kind of like a like a net new ad

46:23or kind of like a totally different

46:24concept. This is kind like where you

46:26want to play around and test like what

46:28the prompt engineering if you will to

46:31kind of like get the perfect output or

46:33the output as close to perfect as as

46:35possible for now at least.

46:38Interesting. Um so, you're doing

46:40variations of the winner rather than

46:42like totally new. Yeah. Okay, that makes

46:44sense. Yeah, for this one, this flow

46:46specifically. Yes. So, then we basically

46:48get 10 prompts uh based on the the

46:52reference image and the bra uh the

46:54branding data and the uh description of

46:56the winning ad. We download the first

46:59winning ad and then we send all of that

47:01over into OpenAI's image genen node

47:05which then has the the prompt that we

47:07pulled from the the agent right here the

47:10branding data the image description and

47:12the reference ad which we um downloaded

47:15here and then we just run through this

47:17entire flow as many times as as you

47:19want. So in my case I had it generate 10

47:21variations. So then it just runs through

47:23the entire flow 10 times. Um, I can open

47:26up the fold right here and see if I have

47:28anything in.

47:33Yeah. Amazing, man. Yeah. So, I'm trying

47:36to find just like the the original ad if

47:39or the original uploaded image, but I

47:42mean, you get the gist of it. So, like

47:43this was probably Yeah. one that I

47:45uploaded. And then if you just go

47:46through, you'll see that

47:50you can just play around. Yeah. All

47:52different variations possible. So, yeah.

47:54here. Here's for like obviously new net

47:56new concepts, but it's still the same

47:59thing, right? So yeah, just depending on

48:02again how you kind of like prompt

48:04engineer these agents will depend a lot

48:07on kind of like the output that you get.

48:08So here you can kind of like see you

48:10have a lot more similar variations where

48:12it's like it's in the bathroom. Um yeah,

48:15before and after style. Uh yeah, kind of

48:17like uh yeah, in nature and so on. Yep.

48:21So you're basically so just to talk

48:23through again this whole process from

48:24top to bottom um research on the open

48:27web find the concepts that make sense

48:30look at the data of like what ad formats

48:32are currently working or it's really

48:34like ad templates y take those ad

48:36templates provide the different concepts

48:39bulk generate 10 variations you know for

48:41each of the based off of the insights

48:43based off of the best performing ads and

48:45then look at the you know top performers

48:47to understand okay what's working not

48:49repeat that cycle over and over again is

48:51kind of the whole strategy here. So,

48:53yep. 100%. And you could you could

48:55easily have those all as like their own

48:58individual flows. So, you could have

48:59like one like this one would be for

49:01creating like like again iterations of a

49:03winning ad. Another flow could be like

49:06creating net new concepts or just like

49:07UGC style ads and so on. And then you

49:09just like have a ton of different like

49:10siloed up flows that are all kind of

49:12like yeah working on separate pieces of

49:16the entire creative production.

49:19That's so interesting, man. This is

49:21crazy. I I feel like we could spend

49:24probably the next three hours talking

49:25through all these, but I know I know

49:27we're coming up to time, so I want to

49:28respect uh respect that. But uh man, it

49:31was been so good to host you. Where

49:32where can people find you if they want

49:34to learn more? Please go follow Jonathan

49:36on Twitter and on YouTube. I know you're

49:38posting there as well. We'll include it

49:39within the show notes, but do you have a

49:41preference if people want to reach out

49:42or want to hire you or want to work with

49:43you? Yeah, I mean, just hitting me up on

49:45um on X is cool. Uh yeah, I think that's

49:48probably the best place to to reach me

49:50right now. Amazing, man. Amazing. This

49:52is so

49:53valuable. I there's so much I'm thinking

49:56about. I like can't even like put it

49:57into words of like what I The last last

49:59thing I have a question for you with

50:01actually before we jump though is so I

50:04mean we're finding this sometimes right

50:05now where it's like creating the N8 end

50:08flow is like can be a timeconsuming. I

50:10mean we're we're speeding this up,

50:11right? So what we're doing currently is

50:12we go we use perplexity to basically

50:15write JSON that we then take um and you

50:18know basically paste into N to like

50:21speed up that workflow building. But I

50:23can show you something for please I

50:25would love I would love to see what

50:26you're using to build these out quickly

50:28because the challenge we're facing is

50:29like do we do a workflow or do we just

50:31go straight to code with what we're

50:33doing right like with our engineers. So

50:35I'm curious actually I had that I had

50:37that tap pulled up but I forgot. So let

50:39me pull that up right here. So I'm using

50:42Claude for this a lot and it's working

50:43really good actually. So good. Yeah.

50:45Yep. Yeah. So inside of Claude, if you

50:48have a paid plan now, you can have

50:49what's called projects. Um and and um

50:53yeah, just just full disclosure, I did

50:56not think of this. I pulled this from uh

50:58Mark Chef's YouTube channel. I'll I'll

50:59send you the link for him as well. He he

51:01also shares a ton of amazing flows. But

51:03you can basically um up. So what you

51:06basically do is inside of cloud

51:08projects, you just create a new project.

51:10Uh let's just call this N test

51:143. You can add like what's called

51:17product knowledge or instructions. So

51:19what you do is you and I'll show you the

51:21prompts and everything. You basically

51:22just upload like best practices of

51:25building N templates and like you pull

51:28from the actual N documentation of like

51:30these are the actual nodes that are

51:31possible. This is like the like the tech

51:33behind it. You can you upload that you

51:36upload a ton of like example workflows

51:38like JSON workflows and then you can

51:40basically just have Claude spit out like

51:43I would say like 70 to 80% completed

51:46workflows. So if I just open this up

51:48right here, you can see that I have like

51:50the JSON uh instructions for like um

51:54yeah prompting cloud. So I want you to

51:56create fully functional internet

51:58workflows. No images or screenshots,

51:59only valid JSON follows and instructions

52:02precisely. Then you just like have all

52:04these rules right here. You upload a ton

52:06of like reference JSON builds that might

52:09be relevant. And then you also have um

52:12this like cheat sheet and tips and

52:15tricks that you upload. Then you can

52:17literally come in and say

52:19like, I want

52:21to

52:23build a workflow that scrapes Twitter

52:28for popular

52:32posts and uses that as a database to

52:38create new Twitter posts for me.

52:53So right now Claude basically is going

52:55it's using your uh documentation that

52:58you've provided it and it's going to go

52:59and write this JSON node. Um yeah and it

53:02it's literally writing the code right

53:04now. So in a in a second right here um

53:08I'll try and copy and paste it and we'll

53:10see if it worked. it usually works

53:12pretty well. And again, it's it's not

53:14it's not 100% there obviously, but like

53:17how I see this is that instead of kind

53:19of like trying to, you know, like mindm

53:21map out these workflows or kind of like

53:23just uh map them out on paper or

53:25whatever first, I I literally just tell

53:28Cloud what I want to build and then it

53:29maps it out for me. And then you kind of

53:31like have a canvas that is like 60 70

53:3480% there. 80% of the way there

53:37complexity of the flow and then you can

53:39just like run with that instead.

53:41Totally. We're doing the same thing but

53:43with perplexity because we found it's

53:45really good at hitting like API

53:47endpoints. So if I'm going to go hit a

53:48rapid API endpoint, um I can include

53:51that within the steps. So like you know

53:53basically it's like write nad JSON for

53:55the following you know flow. Yeah. Um,

53:58and then list out like what I'm wanting

54:00the thing to do, include the API

54:02endpoints that I'm wanting to hit. Um,

54:04it's going to go and find that

54:05documentation. Um, and then what we'll

54:07find is that like most of the times, so

54:09we're using perplexity plow with claude

54:114 like as the as the model. Most of the

54:14times it'll oneshot it. If it doesn't

54:15oneshot it, we'll then take that to

54:17claude and be like uh you know it

54:20basically if it won't import into n

54:22correctly, we just take it to right

54:23cloud floor to um have it uh uh fix the

54:27code if it made mistakes and it will fix

54:29the code within like two prompts

54:31typically to actually get the output

54:32that will go into it. So anyway, it's

54:34just random thing we're seeing work. I

54:36think you have to to show me that then

54:38because I mean that sounds that sounds

54:39amazing. Yeah, I need to make a YouTube

54:41video about it. basically like showing

54:43the process. One that we did was like we

54:45were trying to generate a bunch of

54:46vector images quickly. So I had it

54:49basically like list of keywords from a a

54:51Google sheet. Um it goes through that

54:54list of keywords. It generates an image

54:56with the chat GPT uh open a um uh or the

55:00chat GPT image API. Yeah. Pulls takes

55:02that image back. It then goes and uses a

55:06AI uh uh background remover to remove

55:10the background of the image and then

55:12takes that um the remove background

55:14version and sends it to a vectorzer to

55:16then vectorize that image. And so then

55:18basically like make the you know the SVG

55:20at scale from you know a list of keyword

55:22phrases. So um but anyway that that

55:24whole that whole flow was like written

55:26entirely by

55:29um a perplexity

55:32you know a couple perplexity queries. So

55:34I mean I mean that's what's so cool like

55:35you like sure you need to understand

55:39like I guess again like the logic behind

55:41and kind of like have a sur at least a

55:43surface level understanding but you

55:44don't even need to go like that deep

55:47especially now like with these tools

55:48like literally it it it oneshotted this

55:51thing pretty good. So like you have it's

55:53based based on a chat message. You can

55:56see that we have a couple of tools down

55:58here and it it like it even gives you

55:59like these sticky notes kind of like

56:00explain how the like what tools are uh

56:03available, how the workflow works and it

56:05even like creates these subworkflows

56:07that extracts everything. So I mean

56:09again like you can see right here like

56:10it just created this node and like these

56:12down at the bottom here don't really

56:13make any sense but you already have like

56:15I mean 70 I'd say 60 to 70% of the flow

56:18is already done. You just need to like

56:20hook in the your APIs. Um yeah, clean up

56:23the data, make sure like everything's

56:24pulling correctly and I mean you're

56:26you're good to go. So yeah, it's totally

56:28totally crazy. No, so like as an

56:31example, just for the audience, so we

56:32talked about that it was a it was like

56:34Twitter API.io or something like that.

56:37Yeah, you could basically just reference

56:39their API like endpoints within that

56:41perplexity call and it's going to go and

56:43create that um that node within NAN for

56:46you automatically. So just so like

56:48people understand how it affects crazy.

56:50You definitely need to You need to

56:51record that video and send that send

56:52that to me. Send it over to

56:55Awesome, man. Yo, thank you again for

56:58coming on. I appreciate it, man. Um I

57:00know my audience is going to be uh super

57:02stoked to listen to this and actually

57:04see the applications that you're doing.

57:06Um and again, we'll uh we'll have to

57:08have you back in like, you know, six

57:09months uh to to tell us kind of the

57:11state of the union of what's going on in

57:13AI marketing workflows. So for sure,

57:15dude. I love it. Yeah, I appreciate you

57:16taking the time to to have me on. It was

57:18a blast, man. Awesome, man. Awesome.

57:20We'll talk to you soon. Thanks, brother.

57:21All right, man. Later.