How Torii Rowe Built an AI Operating System for His Entire Agency
This episode features Torii Rowe, a Foxwell Founders member and agency operator who has quietly built a near-fully automated media buying operating system using AI. Tori dives in on how he uses Claude to ingest over 9 billion rows of client data to surface creative and media buying insights no human could pull manually. He walks through standardized naming conventions as a prerequisite for querying data across all clients, along with practical advice on where media buyers should spend their time, and the 1 tool all agency owners should be using to talk through their business problems with AI before building anything.
Key Takeaways
How to connect an AI model directly to your data warehouse to unlock insights without writing SQL.
What a true AI operating system for a 30+ clients actually look like
Why this one metric is a better early indicator of creative performance than click-through rate or CPM.
How standardizing their naming conventions unlocked a cross-client creative analysis that no analytics platform currently offers out of the box.
The step-by-step stack Tori uses to go from client call data, to creative brief, to AI-generated ad.
The 1 thing that Weavy (Figma Weave) actually does that other AI image tools cannot.
When to bring in real engineers versus continuing to build with Claude Code yourself.
How to prevent analysis paralysis when you have access to massive amounts of data and insights
Tangible Links:
Snowflake: Data warehouse that has its own Cortex AI model built in and a native MCP snowflake.com ↗
Airbyte: Data pipeline tool used to pull in data from all paid media platforms, Klaviyo, Attentive, and post-purchase surveys airbyte.com ↗
Weavy / Figma Weave: Node-based AI image generation pipeline tool. weave.figma.com ↗
Kai.ai: AI image generation tool used in tandem with Figma for ad creative iteration. kai.ai ↗
Gemini (Google): Used for video tagging because it can watch and analyze video content gemini.google.com ↗
Klaviyo + Attentive: Email and SMS platforms klaviyo.com ↗
WISPR Flow: Voice-to-AI tool wispr.ai ↗
To learn more about Torii Rowe and his team At DREAM Labs Agency head here: https://dreamlabsagency.com/ Https://x.com/ToriiRowe
To Connect With Andrew Foxwell reach him here Andrew@FoxwellDigital.com
To connect with Will Sartorious DM Him Here https://x.com/will_sartorius
To Connect With Thomas Moen DM him Here https://x.com/thomasmoen
To learn More about The Foxwell Founders Community and the conversations, like this one being had go here: www.foxwellfounders.com
Full Transcript
(00:00) If you work in D2C and you use AI and you're wondering what the F is going on every week, this is your podcast, the AI D2C WTF podcast, your home for tactical tips, strategies, and ideas that you can implement right now in your AI workflows to make your brand or agency more money. Welcome to another episode of AI D2C WTF.
(00:25) Let me tell you, I have an incredible episode here for you today with Tori Rowe, someone that... I really respect, Foxwell Founders member, who has essentially with AI built an entire operating system for his agency. I mean, he's got an audit maker. He's got like so much stuff in here. I don't even remember all of the pieces he showed us.
(00:45) So I'm excited. And this episode is definitely going to be worth your time to listen to. Will, what were your takeaways in chatting with him? Yeah, I mean, he's a pretty modest guy, but this, I will say Tori is well ahead of the bell curve in more ways than one. He showed us a dashboard that he's effectively built out automated media buying, ingesting client calls, having a repository for...
(01:11) So much stuff. It was so cool. It's a little bit of... You know, I had a little info overload at one point, and then we sort of dialed it back and he was able to sort of explain everything really succinctly. And I was just sort of flabbergasted by how fast he's moving. Yeah, I completely agree. So check it out.
(01:30) Let us know what you think of the episode and hit me up, andrew at foxwelldigital.com if you have any questions or feedback on it. But, you know, we always are trying to give tactical information to you on these episodes and that you're able to just take and implement and think about and give you some brain food on AI.
(01:46) So that's what we're here for. So let's take it away to the episode. So let's take it away. Tori Rowe, Floridian legend of AI and digital marketing, who I love having here and is an incredible founder member. Welcome to both of you. Appreciate it. We are here to talk about deeply AI tools that we're using, things that we're getting into, making sure that you can walk away with this show with tactical stuff that you can use.
(02:14) You know, Tori, you're somebody who's an incredible digital marketer that I really look up to and respect in your thoughtfulness and the approach and the completeness of the way that you go through thinking through the funnel. Right now, what are you absolutely like creating on AI that you're loving that's really changing results for you? Yeah, I mean, there's so much.
(02:35) First off, appreciate the kind words. There is so much. I think the biggest thing that I'm like nerding out about is probably the reporting side, which is drastically different than where everybody else is going. I feel like everybody's so zoned in or honed in on like the creative front. There's so many tools.
(02:51) There's so much cool stuff going on there. But like connecting an MCP to like a data warehouse is something incredible that I've never made an MCU for. So like one of the big things we're doing right now is mapping all the orders from as far as we can reach back, first click all the way to the back, basically the very last order of someone and running an MCP through it.
(03:10) Ours is Claude. We run it through the Snowflake database and basically pull out as many insights as we can. Right now we have like 9 billion rows of data. There's no way a human can actually sort through this. So us going through and trying to find basically anything we can out of this and say like piece this together, show me some insights that I wouldn't be able to pull on my own, stuff like that is super incredible.
(03:33) Yeah, there's a lot on the reporting side that I'm just like going really deep on. Still the hypnotic creative side a lot, but the reporting is where I'm spending a lot of time the last few weeks. So let me just jump in for a sec. So you have all of your clients' data separately, obviously, like in separate warehouses.
(03:51) And you're going through and you're building report. Like what was your reporting before versus what is it now? And how has it helped you make better decisions? Yeah, as far as like reporting before, standardized spreadsheet stuff, super metrics, you know, BigQuery kind of stuff, pulling it out, thrown into a spreadsheet.
(04:08) As good as AI is, you still need like the granularity in like a spreadsheet. Sometimes I feel like you still can't beat it, right? So the client facing reporting is still like spreadsheet focused. But for example, if we standardize our naming convention across everything and have the exact same UT and parameters across every single ad that goes across every single ad account, we could say, hey, tell me how UGC performs to females across every single client in the month of March.
(04:35) And it would tell us exactly what happened, right? Tell me how it performs to men. Tell me the age of the model, right? So if we name the ads, you know, angle, model, age range, hair color, hook, length, right? And just keep going. Now, if that's standardized, we could say, hey, pull format six. Tell me, you know, the like comparisons across all of our clients throughout March and what has the best row as, right? Or what has the lowest CPA? And then what we're taking on top of that is we're tying this to customer lifetime
(05:07) value with identity mapping. So we'll be able to tell you what the customer lifetime value is instantly. So as soon as someone finishes an order, we push that customer lifetime value back and we end up tagging that in the warehouse. So we're able to say, hey, this is generating the highest customer lifetime value customers.
(05:24) UGC, blonde hair model, 25 to 34 with this hook. Nice. I have a follow up there. So naming conventions, a lot of folks use to ensure that they stay organized. A lot of these larger platforms, Motion, you know, they rely partly on naming conventions to keep organized. Say I'm a creative agency, you know, for example.
(05:48) And there are a lot of different media buyers out there who have their own naming conventions. Have you explored something like using, you know, I don't know, let's say Claude Sonnet to review the assets rather than using a naming convention? And if so, like how does that compare to like sort of the old fashioned way? Like you're outlining UTM parameters plus naming.
(06:11) Yeah. So we're using Gemini because Gemini can watch videos. I'm not sure about Claude Sonnet, right? And so Gemini, we're trying to figure out if Gemini can tag these videos, very similar to how Motion tags the videos. The issue is with it, the granularity that we're trying to pull as we get deeper and deeper is where things get difficult, right? And so like there's also not like a QA point of this of like somebody has to go through and QA all this.
(06:35) And when you have 10,000 ads going live, so it's easier for us to just dump it onto the creative team who already has the naming convention. And so they're going through and actually naming it while they know what the file actually is. So instead of our media buyers watching it refile, instead of us QAing over some AI source, I think it'll get there, to be honest.
(06:53) I just don't see it right now. It's something we are working on, but it's a tough one to unlock, I feel like still. Motion surprises me how far they've gone with it. I mean, they're incredible. Totally. And I think that like honestly speaks to sometimes like the old fashioned way is still the best. Not everything needs to be totally optimized in that sense.
(07:13) Again, I agree. Like if you do UTM parameters and naming conventions, right? Like you're going to have a far more systematic and like thoughtful database than if like you're just, you know, either having Sonnet or Gemini sort of guess at what you're trying to build. Yeah, the way I've been like explaining it to our team is basically imagine if Triple Whale, obviously they have like 4,000 people on it, right? We're nowhere near this size.
(07:35) But imagine if Triple Whale could standardize naming convention across every single client they have on the platform, what type of data they would have. That's what we're doing internally. They should do that. It would be insane. It would be insane what they could pull out of that, right? Yeah. I mean, I don't know if you guys have access to the Motion MCP, but there's some of that that's new.
(07:58) I mean, that's coming that if you don't have access to it, maybe just email them and ask if you can get access to it or whatever. There's a ton of crazy stuff that you can pull out of there. I think that this analysis, one is, you know, yes, this is what it looks like. It's specifically this type of ad, this type of thing we're putting out there that really is going to bring in the highest profitable customer.
(08:20) I think the future is totally that connection between the financial metrics of the business and understanding, you know, how you, what your inputs are on the marketing side and how you bring in higher, more high value cohorts. So you're building this out. This is your, the data warehouse. You're looking at this.
(08:38) Like, what are some other tools that you've used in building this that you talked about a couple of options, but what are some other connectors or anything else to mention there? Yeah, it's pretty simple right now. It's literally all the connections, right? All the top funnel paid media kind of platforms across the board.
(08:55) And also Klaviyo Connects and Attentive, all that kind of stuff. Post-purchase surveys. We're building APIs over from post-purchase surveys to feed that data in. Everything's pulled in through Airbyte. Airbyte goes over to Snowflake, stores into the Snowflake database. And then Snowflake database can kind of push anywhere, but MCPs over to Snowflake.
(09:12) They also have their own MCP on there, but we're feeding it back into Claude. Claude has so much information already about us because we've been so deep on it, right? Of kind of where the business stands and all that kind of stuff. So we're feeding it back over there and then basically running anything we want on top of Claude.
(09:29) But you're trying to... Our goal as an agency is trying to cut down on the actual media buying part for our media buyers. I don't want people uploading creatives 40 hours a week. Like, that's not the goal here, right? If I can get them to spend 50% of the time in data and asking questions, and like, this is not a knock on media buyers.
(09:51) This is a knock on myself too, obviously, in the media buyer. We are not always the top data scientist, most data analytic person, right, across the board. So if we can just ask simple questions and start to understand this data more, you're basically just empowering people to be able to make better decisions without having the guess of what they think is working or trying to...
(10:13) Hey, this has click-through rate. Here's frequency. Here's this. Here's this, right? We're just like, no, this is the ad that is driving the most people for the first touch point. This is the ad that is converting people. What are the, just as an aside, but like, what are the metrics that you use most often that are actually correlated to results in growing a company that aren't the typical ones that you hear people talk about? Everybody's on CPMR and stuff like that right now, which is just funny.
(10:45) It's always switching. And the one, we actually ran an analysis on everybody last year and the number one indicator of a creative that ended up scaling was cost per ad to cart, which makes sense. It's a very simple, you know, creative, but it's, you can get ad to carts early and not get purchases early, right? And so it's just a higher top of funnel metric.
(11:05) If that is kind of in line with everything else, everything else seems to fall in line. There was not a click-through rate analysis that ended up showing that that was it or CPMs, nothing like that. It was always cost per ad to cart was the number one indicator for a good creative. Nice. I sort of want to, if you don't mind, sort of go back to the database AI component, because I think that's really interesting.
(11:25) And I guess my question is sort of twofold. Like, how did you sort of land on each of those different touch points? Like, why did you pick Snowflake, for example? Was it just like a cloud recommendation? Or did you see like an advantage there? And two, I think what folks maybe get confused or lost about it's like, okay, I have all this great information.
(11:44) Like, I know what's performing now. Like, how do they sort of take that to the next step to actually generate ads that will work, right? Like, having the information is great. But like, what is that human next step to either iterate or build new concepts that will actually perform? Yeah. So great question. As far as like, I mean, we went through a few steps here.
(12:04) First, we tried and not trying to knock other companies, but we tried Supermetrics. It was too slow. And it couldn't pull as granular data as we wanted. So we went from Supermetrics, found Airbyte. We have a bunch of engineers we work with. So they were able to basically tell us, hey, Airbyte's a good one. Then databases, we were kind of going through.
(12:23) And Snowflake, they have Cortex, which is their own AI model, right, on top of everything. So we were able to use the LLM on that, which was, you know, just something that we wanted. Because I wasn't sure how MCPs would end up pulling in. So that was kind of like the first step of just finding what worked. Also, as weird as it sounds, we tried to go to the ones that like the big guys use.
(12:43) Because we knew if, you know, an IPO company was pulling in 500 billion rows of data, we'd probably be relatively cheap to pull in five to nine billion rows of data was my first thought process. I was like, we'd be the small guy over here. So I kind of wanted that as well, because there'd be tools that, you know, obviously they're using and we're not using.
(13:01) As far as like the action points on the data, there's a few ways to kind of go about this. One, you already have this creative feedback loop, right? One thing we've always looked at is like, does the model work? Yes or no? Because we'll take the same exact script and give that to 15 models. And it'll always come back to like, this model works or this model works, right? So that's one way of just like starting to bucket the data.
(13:27) Into basically digestible places for you to say, hey, here's the top model. Here's the top hook. Here's the top. For example, we run a jewelry company. This is the top background. They switched from all white backgrounds to all black backgrounds in February because we found out that worked better. All of their top creators are black backgrounds right now.
(13:44) So like simple things like this, if you like start to bucket things, I think you can have like process by analysis and go too deep here for sure. But the end all be all here is probably feed this analysis over to creative briefing, have a creative brief, do the analysis on or take that analysis and then write the briefs for you.
(14:03) Your creative strategist on top of that, have them kind of check it, but just basically speed up the workflow across the board. Nothing's 100 percent. Things probably 70 to 80 percent with that last 20 to 30 percent being the human touch point. But just trying to get it through that pipeline a little quicker.
(14:18) I think it's interesting that the idea of of taking what I've heard a lot is people talking about taking a lot of data and putting it in by client and making agents right to be able to query, which is essentially kind of what you're talking about. But for you, it's even more. It's bigger. It has more implications for the entire business.
(14:35) I could definitely even see a situation where it's like you have access to not just their Shopify data, but like their financial metrics. You know what I'm saying? Like as a marketer, like because then you're able to look at obviously highest margin categories and then you're able to say, OK, when we do these things in succession or, you know, let's combine them.
(14:55) You know, the black background thing is a simple example. But the more that we're able to start to connect these lines, the better off we're going to be, because I think thus far what's been happening to a lot of people is like that's not good enough. If you look at like Zach Stuck, for example, right, who's one of my close friends, like what he does is and with his team is basically he'll go and say, that's not good enough.
(15:14) Let's do a better offer. That's not good enough. We're going to do a better offer. And then they're going to they keep designing the flow. So the ad creative, they're testing and designing and then making sure that the lander matches it. Right. And that's how they're continuing to expand on persona development.
(15:27) So like with this, the more that you give it, obviously, the more powerful it is in moving, moving forward. And it's powerful for your team because then you're higher leveraging them, as you said. Yeah. So like to the persona thing, it was like we can sit there and tell you, hey, 25 to 34 year old women coming from Instagram reels convert at X percentage and have a higher customer lifetime value when they go to these pages.
(15:48) You know, pretty quickly. Right. I can tell you that in probably 10 seconds just by a simple question over to clock. Right. So stuff like that of like we can dive deeper into this. Like I said, I think the biggest thing is trying to stay focused and not go too deep at the moment. Right. Of like we're not trying to.
(16:06) Yeah. We're not trying to go reinvent the wheel. We're trying to do make the wheel a little faster. Right. Now, will we go down that path? For sure. At some point. Like let's get the basics down and do incorrectly and the rest of this will fall into place. Nice. Yeah. That makes sense. It is so easy to probably. And I think folks listening probably can relate to this just like analysis paralysis.
(16:29) Right. Like we have all of this incredible information. Like what the fuck do I do next? Right. So you were saying the database helps you effectively come up with raw scripts. And at first when you're talking about models, I was like, I thought you're talking about different AI models. But then I was like, oh, it models, creators in the wrong headspace.
(16:48) But so what you're saying is you have effectively you have the model say this is what is working well and maybe generate like an initial brief for it. And then that brief goes to a human being that, you know, cleans the brief up. Then that ultimately goes to the creator. Is that sort of the flow? Correct. Yeah.
(17:04) Like it's basically, I mean, if you want to go deep on it, like we have from the very top of like what we build with AI. We now have a creator platform that we cold email creators on Instagram. They can go to this platform. They sign up. It stores into a backend database for us. So our team can filter through the creators.
(17:21) Right. So they filter through the creators. We actually push out product to those creators. Claude writes the brief to those creators. Now that's like the top of funnel. This filters back in after those ads run back into the data warehouse. And that cycle continues. Right. And so we say, hey, it's a, you know, brunette model.
(17:38) It's a guy who's 45 years old. Sweet. Go back to the creator database. And eventually this will all be connected. You just tell Claude, hey, go find these creators. Now it goes by. And so it's just hopefully cyclical at one point where all of this is just connected. Anything we build with AI, we try to take the next step from what we last left.
(17:57) And I think I found this out pretty early for myself of like, I would go over here and build something. I'd go over here and build something. Right. Instead of like piecing things together and eventually just kind of connecting them all. I would go do two things on the opposite side of the room, which don't benefit you as often.
(18:13) Yeah. That definitely makes sense. Yeah. And I think the impetus behind, or for many of us rather, is just like, why would I have Claude just do this one use case when I can have it do 10? And like, you know, why would I just make this internal when I can turn this into a tool for everyone to use? It's just like, there's such a rabbit hole to sort of go down.
(18:33) So how do you sort of prevent yourself from going down that rabbit hole? And then even more so, like, what, at what point do you realize, like, I've built something that is sufficient for my use case? Like, do you, are you sort of coming up with a concept beforehand? Like, here is what my end goal is. And once I hit that, I'm not going to move any further.
(18:53) Or is it more so you just like intrinsically realize when you've hit that? It's a really good question. I will say, like, I'm in the rabbit hole. I'm climbing out of it, probably. So, like, I've been down it quite a bit. I think the easiest way for me is, like, I'm trying to solve a problem when I go to this, right? And so I go to my entire team.
(19:11) And, like, we sit down. We sat down for, like, two hours one day. And it was early. And I'm sorry that my team had to go through it. And I was like, tell me all your problems. And I just sat down, recorded the whole meeting. And then basically was like, okay, what problems do we need to solve here? And what problems are adjacent problems that I can go fix? And so that was, like, the first thing of, like, the internal side of the business was, like, fixing all of this, right? Now, on the external side that you hit
(19:36) on, as far as, like, SaaS platforms and people kind of pushing things out, one of the things we wanted was, like, bulk ads launcher, an email writer, like, all this stuff of, like, a platform that we could use to media buy emails, read images, do all of our reporting all in one singular place instead of clicking.
(19:54) I mean, we have, you know, 30-plus clients that were, like, clicking through ad account to ad account, right, having to check everything. So I wanted it for us. We built that, got it probably 50% of the way there. And then we hired three engineers, two AI engineers and a product engineer. And they came in and they're overhauling it.
(20:12) And so we wanted to make sure it was done 100% accurate, as clear as we possibly could. And that is a SaaS platform that we're launching. And that comes out in four weeks because we went to the rabbit hole. And I was like, hey, if we want this, everybody else will want it. Right now we have, like, 250 signups waiting for it.
(20:28) What point, that's really interesting. At what point did you say, like, we need to sort of bring in, like, the big dogs here? Like, was it just, like, you found yourself spending way too much time in cloud code and you were just, like, you know, bug, bug, bug? Or is it more so just, like, you know, I've reached an impasse? Yeah, it's, it's, everything was functional.
(20:48) Even something as simple as, like, the bulk ads launcher, right? Like, it takes so much time to launch creative. Of, like, I could get it working. It's going through and everything like that. But, like, think about how good an actual engineer or a developer with cloud code can be, right? And so they're going to take it to the next level.
(21:04) And so that was kind of the goal of, like, if I can get it here, they can get it there, right? And I want to make sure that it's as efficient as possible for everyone in the room. So that was kind of the next thing is I knew that I'm not an expert and I just wanted someone to make sure, hey, you know, dot the I's and cross the T's and say, hey, this is good enough for everybody.
(21:22) Yeah, it's that kind of line of taking a step back and understanding what's going to be useful to you versus the team is huge. And, yeah, it's interesting. And good job, by the way, on launching a SaaS. I feel like if you don't have if you're messing around with, you know, AI and you don't have a SaaS that you're launching, it means that you're you're not trying hard enough right now.
(21:39) I feel like I was like every day I'm like, I could launch that. And then I'm just and then I'm like, no, this is garbage. Nobody's going to pay for this shit. But it's but but it is like it's a sign of like, oh, wow, this is actually really useful. This is super interesting and it helps you kind of connect dots.
(21:53) So under that, what what other things are you have you been doing with AI lately that outside of this that have just been absolutely mind blowing for you that have been awesome? Weavey is for sure the top thing that I'm like wowed by right now. So how do you spell this? W-E-A-V-Y. Let me just. Yeah. Pretty. Yeah.
(22:14) Yeah. Weavey. Yeah. Yeah. Weavey. It's like Weavey Figma or something like that is what it's called. Everybody like I mean, we've been working on AI creatives for, I don't know, a year and a half. Right. I think everybody's always been since ChatGPT came out, everybody's trying to been like finagling with it. And I'll say like they're pretty good, but it's still like a 60, 70 percent hit rate.
(22:36) Right. Like they're not all perfect. You throw a lot in the trash and stuff like that. This is 100 percent every single time exactly what you're looking for. It has been absolutely incredible. Basically, what it is, is you're building a pipeline with prompts, images and all of this kind of like basically framework to get to your end result.
(22:54) So you could the guy who came over to us, we just hired him. He came from BarkBox and something for like BarkBox is like they really struggle to go do a photo shoot with a bunch of animals and get all their toys inside with this. Right. It's not easy for them. So what they did is they ended up building this thing through Weavey, this pipeline.
(23:12) It basically says like, hey, pick the dog that we want. You know, like Golden Retriever. And then they're like, OK, pick the plush toy. They picked a plush toy and it basically has like a 3D rendering of the plush toy. And then it's basically feeding this into a prompt. They're describing the background. Then it feeds to another prompt.
(23:27) They're describing where they want the dog to lay. Then they're feeding it where they want the toy to lay. And then it feeds into an image and that image comes out after you run this. And you can basically just do this on a pipeline nonstop. So basically getting Claude Code to basically build something to feed the briefs and the prompts in here and just kind of put it on a flywheel.
(23:48) Cool. It looks like it accesses pretty much any model. Right. Yeah. It's insane. Will, you're going to love it. Yeah, I'm stoked about this. You're using it. So like you're doing this with clients now. And it's who I mean, like what are the verticals of the clients you're using it with? We use it all over. I mean, it's honestly like we haven't found a vertical we can't use it in, to be frank.
(24:11) Like I think the biggest thing is founders who are open to it. That has been the biggest thing that we've kind of ran into is like some people are like, hey, it's not authentic. Right. And I get that. I get both sides of the table here. So it's like hard for me to be like, hey, you should use AI or you shouldn't.
(24:26) So I mean, supplements is a very simple one, kind of easy across the board. But also like we've used it in cowboy boots. We've used it in CPG quite a bit. What else we got? Apparel is super simple. We'll also use it just to do like backgrounds for emails. Right. Like we like really cohesive backgrounds for emails of like generating a background that is still branded and then kind of overlaying the template on top of it.
(24:49) Like just simple stuff that would normally take a designer. I don't know, an hour, hour and a half to kind of create something like this. This is done in five minutes. Very cool. Sorry. I'm just like lost in there. I'm lost in their flow. So would you find this is sort of give sort of like, you know, what is like the one use case that folks can sort of use it for? Like, like it's because it seems like very manual in the sense like you need to sort of pick your model.
(25:15) You need to pick the prompt. Like, what would you say is like the one sort of like start starting point for maybe someone to jump off on for? Yeah, I think you. Yeah. Just start your flow. Right. If like there's some like very simple. Take your hero product and start the flow. Right. Start to build that prompt and stuff like that.
(25:32) But you can start to basically put these in a drop down of like you have 50 prompts that are already there and you're just selecting it. Right. And so you can start to kind of mix and match and do what you want to do. So like, for example, back to like Artbox, they're just selecting the dog they want and they have 50 images of that dog that they've already like uploaded into it.
(25:52) So you're starting to get these variations instead of like going to Higgs field or Kai.ai or whatever. And you're actually like manually uploading these things. It's basically storing it. So, yeah, it's a pain in the ass the first time. But once you kind of have that template, then you're just copying it, duplicating that template, changing out the products, changing out your prompts and saying, OK, this is hero product number two.
(26:12) Right. And then going back to like what you and I have talked about before, well, the iteration process is the easy part. The real I think the real winner is a net new concept from AI is the extremely difficult thing that I haven't been able to find that is like mind blowing until now. Hmm. And so you're saying you have like a winning ad.
(26:34) Like what would be your process in terms of iteration there? I mean, we have it from the Figma board. We built a plug in from Figma to Kai.ai. And so we can say go make 10 iterations and we can select change product placement or change the background. And so we can do either of those and we just select how many numbers we want.
(26:55) Very easy for us is we do nine by 16 with a four by five safe zone. That's kind of like the limits to stay within this. We just go to Figma, select eight, nine, ten, whatever we want. Press run. It comes back, gives us exactly what we want. No prompts. No, no. And so it'll just kind of run through that. And then as we spoke about again, well, it's like then that kind of duplicates and we do the exact same thing with the headline iteration.
(27:16) So the H1s up top and then we just say, hey, 10 new H1s. It runs back to Claude, runs through the Claude project for that company, that brand, runs to the Claude project. Claude project is connected to Notion, runs to the Notion dashboard all the way back, right? And then it just generates the H1 copy across.
(27:33) That's crazy. So the H1 copy is unique, written by Claude every time is what you're saying for every iteration. Every single time. Yeah. I mean, you could do, I think we have it up to like 25 right now, but it probably takes five seconds for it to do it. It's so quick. Yeah. Okay. So I want to sort of drill down on one thing you mentioned.
(27:52) One of the issues with these models, I think we can all sort of agree like Nano Banana 2 is probably the best model for static generation, is like getting your text to look right. My process is like building, you know, building spec cards and having it reference the spec cards, but still it's maybe like 80, 85% success rate.
(28:10) You mentioned something the other day that I thought was fascinating. It's like in this Figma thing, you are manually creating the initial one in Figma, right? With the actual copy, with the actual font file. And then through your iteration process, you're changing that through Claude. So, you know, maybe like at a fifth grade level, can you sort of like explain exactly what's going on there? Yeah.
(28:30) So we'll go generate, let's say you have your image, right? You're like, hey, this is the image I want. That lives in Figma. The designer comes in is like, hey, this is where I want the H1 copy. I think not only is like the font hard to find, but it's also like the placement when you kind of use AI of like getting that placement perfect.
(28:48) It really looks aesthetically pleasing is really difficult, I feel like. So basically all we're doing is we built a plugin from Claude code to run back to Claude, right? And so all they're doing is they're basically designer comes in, makes the first iteration. Like that's basically the only thing they have to do is come in, make that first H1 copy.
(29:05) And then we just hit the plugin. It runs back to Claude. And then it just duplicates the image across and just puts the H1 copy in the exact same place with the new H1. Got it. I think, yeah, I think that like what you're saying is like so valuable because the amount of DMs and questions I get about it's like my font's not rendering correctly.
(29:23) Like this is like the easiest solve that people just like don't realize. It takes like one extra second of effort in Figma to just like actually write out the copy and then like iterating becomes so much easier. Yeah, 100%. The thing why we started doing this, one is I believe it was, I don't know if it was the CMO at Claude or someone, but there was someone at Claude who spoke about doing this with Figma in a different capacity.
(29:47) And I was like, I watched that video and I was like, we're just going to change it and do the H1 copy. Because the biggest thing that pissed me off is we'd go generate an image and the image would look great and the font screwed up the whole image and start from scratch again. Right. And I was like, I don't want to keep doing this.
(30:01) Right. The image is right or the font's wrong or the font's right and the image is wrong. I'm going to do them separately. And so the first thing I did was go make the images and then I'd come back and upload that image and try to get the font right. And then it still wouldn't be right. And so I was like, okay.
(30:15) And a lot of it was placement of where I wanted it. Like even just trying to slide it over, like a quarter of a centimeter on 9x16 matters, right? Like it's a huge difference. And so just having them place it and then you just name it H1 and then Claude just knows that this is the H1 text box and it just runs it across.
(30:35) Great. Very cool. I mean, we would sort of be remiss not to mention that Claude design came out yesterday. Have you played with it at all? I mean, I blocked off some time tomorrow to spend on it, but it seems like, you know, Figma stock is down. People are, you know, talking about this a lot. I don't know if you've gotten in there yet.
(30:53) I haven't touched it. I know it's out. I mean, we had Opus come out, right? Everything like this. But I have not touched either to that capacity. I've been in meetings to meetings the last two days, but that is this weekend. The fiancee will be pissed when I sit on the couch all weekend playing with Claude design.
(31:09) Yeah. What used to be a Friday night of just like drinking beer and watching sports is now just like drinking NA beers and playing with Claude code. Exactly. It's a life. It's a life I dream. Sounds awesome. Yeah, it sounds awesome. So, you know, what do you think, Tori, just as a way of, you know, getting people ahead that the way that you think about AI and how you're, you know, integrating in what you're doing, setting aside time.
(31:37) And where do you think you recommend media buyers and agency owners to spend their time moving forward with AI? Like, is it around the creative optimization? Is it around thinking about the whole funnel? Is it both of those things? Like, how do you sort of allocate your time in terms of really seeing client results? Because I think a lot of us struggle with like, all this stuff's cool.
(31:57) I don't know what to do. And so it's really a matter of where can you focus that time? Yeah, I mean, I think the biggest thing for me, and like, obviously, like, I'm a founder. And so like, I have a team is I wanted to provide more value to two people, my team and my clients. That was like the only thing I really cared about.
(32:15) Like my time, listen, I'm going to work 100 hours a week anyways, because I'm a psychopath. So I'm going to figure out a way to like work 100 hours. It doesn't matter. I wanted to let them get more done within their work week or provide more value on the other side. The biggest thing I would tell people and how we kind of got started and how I tell my team to get started is like whisper flow.
(32:32) If you know what that is, WISPR, you know, you just hold a button on your keyboard and you talk to it. It's a very easy way to talk to an LLM and actually like get good results. And you can sit there and kind of process and think, double tap the key and sit there for 10 minutes, literally 10 minutes, I would say.
(32:48) Like, don't go any shorter and just go describe everything in your business and all the problems you have. Like, hey, you know, we're not the best at reporting. We need to fix this. Our spreadsheets are manual. This is how I want to do this, etc. And let Claude, Chow GPT, Perplexity, whatever it is, prompt you back and tell you, hey, these are the things you can do.
(33:08) Everybody's business is so different. Some people are really strong on the creative side. Some people are really strong on the analytic side. So I wouldn't say like, hey, there's a one size fits all year. I think just go talk to it. Describe where your issues are and let's see what it spits out so you can start to focus on what's going to move your business forward the most, not what I'm going to recommend because everybody's a little different here.
(33:29) I think that's a good call, right? In terms of sitting down and just talking to it first and starting there. That's something that it's easy to forget when you're getting a lot of information pushed to you about like, these are the 10 things I solved, etc. Well, Tori, appreciate you hopping on with us. Appreciate you sharing some just absolutely golden nuggets.
(33:49) If people have questions or want to contact Tori, you can find him on X or you can email me, andrew at foxwelldigital.com and I'm happy to connect to you. And Tori will share some resources in the notes here, but I appreciate you hopping on. Awesome. Thank you, guys. Appreciate it.
