Why AI UGC Is a Trap (And What Actually Works) | Cody Plofker

This episode features Cody Plofker, of Jones Road Beauty, and one of the most talked-about voices in the DTC/e-commerce space. The conversation covers how Cody has turned Claude into a full-stack CRO and conversion engine, going 5-for-5 (as he puts it) on winning A/B tests with almost 10% CVR lifts each.  

Cody also gets into his end-to-end workflow and how it all starts with his "customer intelligence" markdown file fed by live MCPs ( a combination of Junip reviews, RichPanel CX tickets, OuterSignal, ListenLabs, GA4, Heatmap API). He also covers how he built Jones Road's entire design system inside a GitHub markdown file so he can now one-shot landing pages at an 80% quality rate in a single session.

Andrew and Will dig in with the harder questions from why human taste still matters, why QC agents alone won't save you and how to decide whether to build a landing page first or an ad first. The episode closes on KPIs and how Cody is raising output expectations across the board and tying AI utilization directly into performance evaluations.

Key Takeaways: 

• How a CRO intelligence process that's pulled from customer reviews, support tickets, surveys, heatmaps, and GA4 can change the quality of your test hypotheses.

• What you're actually leaving on the table when relying on QC agents to catch everything in your AI output.

• Why you should be building your ads & landing pages from a single brief. 

• How your team can go from idea to live within 24 hours to get more testing cycles and improve your overall funnel launch standard.

• How to use this analysis to surface winning persona/angle/format combinations you quietly turned off a year ago.

• Understanding how much of your competitive moat actually lives in the quality of data and documentation you've fed into your AI system.

• When you tell your team AI adoption is mandatory and the approach that creates real behavior change.

• What "best in class" output actually looks like now that AI has raised the ceiling on what a single operator can produce.

Tangible Links

• Granola: AI meeting notes tool - granola.ai

• GitHub: Code repository - github.com

• Google Trends: Trend monitoring - trends.google.com

• Reddit: Community platform - Reddit.com

• ClickUp: Project management software - clickup.com

• Tempo: Recently launched tool mentioned as an example of commoditized software in the design/build space - WithTempo.Ai

To learn more about Cody Plofker and his team at Jones Road Beauty go here: https://x.com/codyplofhttps://www.jonesroadbeauty.com/

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 Foxwell Founders and conversations like this one, go here: www.foxwellfounders.com

Full Transcript

Andrew: Welcome to another episode of AI D2C WTF. So good to be with you again, Will, and today's guest — Cody Ploffer. Legend. Obviously everybody loves this guy. He's doing a ton with AI. The episode was absolutely mental, in my opinion, in terms of the things he talked about. What's your take on this?

Will: Yeah, I think "mental" is an understatement. It seems like he has become his own CRO agency in and of itself, and it's very clearly driving results. But also — how to properly use Claude Code and how to incentivize your employees to use Claude Code. I think he just did a really nice job touching on all these things that we're all sort of stressing out about. But if you're interested in landing pages, this is the one to listen to.

Andrew: Yeah — AI landing pages, the creative process, the way the systems work in his organization, and organizational AI usage. A lot of good topics touched on today. Let's get into the episode.

Tracking CRO Tests at Scale

Andrew: As I was just saying to you, everybody loves hearing from you and loves hearing what you're up to. The first question I want to talk about is what you mentioned to me at the Meta Performance Marketing Summit — showing me how many landing pages you've built. You admitted it was not a great use of your time as a CEO, but you said it had been huge in terms of results. One question I had is: you're building all these pages and going through them. How are you keeping track of the tests in a way that makes sense? Because you're doing so many iterations, and everybody does this a little bit differently.

Cody: Yeah, I built a roadmap tool. That was actually one of the first things I did — and I don't follow it, I don't organize it. I just hired somebody new, which we'll talk about, to run this whole process for me. I'm going to give her everything I built and did. But I'm just terrible at taking the time to do it. Theoretically it would all be in there. Otherwise it's just kind of like intelligent chaos. But yeah, built a lot of pages — a little combo of new landers and then doing a ton of website CRO with Claude as well.

Andrew: So I'll just ask one more follow-up on this. You've gone through all of these and — how have you decided which elements to use? Because a lot of the pages I saw you had built, they have obviously the PDP and then a whole bunch of features under it — different skin tone variations, all these different features that I'm sure you've seen convert before. How do you decide which to put where? Is it based on what you know about the best-performing combinations?

The 5-for-5 CRO Streak — and Why

Cody: Yeah, that's a good question. Those were actually sections I built for our PDP for a new launch we had. For me, the most foundational thing — and I've said this in a lot of tweets — I'm five for five on intelligence tests that I've run end to end in Claude. And by end to end, I mean starting with the insights and the research and the strategy.

What I think is cool is not just that I can design and build stuff, but because you can do that so fast, I actually think the strategy is more important. You only have a certain limited number of testing slots you can take up. So being able to do those to a much higher degree of quality — I think I'm really onto something, because prior to this, we have never had five winning tests in a row. Usually it's like one or two out of ten. And these were some pretty big wins — like five to ten percent lifts.

Andrew: Ten percent lift on CVR?

Cody: Yeah, CVR.

Andrew: That's insane. I mean, relatively speaking — that's crazy.

Cody: Yeah. And the fact that it's multiple in a row. The reason why is this customer intelligence system I trained. I just — I didn't know where it was going to go — but one Sunday I was just like, hey, I feel like Claude could do a pretty good job with CRO. So I started feeding Claude everything I had access to. I gave it our intelligence, MCP, I gave it a Slack channel, I gave it spreadsheets, I gave it transcripts of Looms from a past CRO agency, I gave it notes I always take in Granola when I'm recording my podcast. I've had Dylan Ander on twice — he's the CEO of Heatmap — and I gave it those notes. He later sent me his book as a PDF. So I kind of built these knowledge bases and trained it.

And then what I'll do is just go to Claude and say, "Hey, this is what I want to do. Here's what I have. What else do you need?" And I'm like, "What do you need for UX best practices?" And Claude tells me — Baymard. So I go scan the web and find that. So Claude kind of has all the CRO frameworks. It's trained on the ICE framework. It's trained to look at highest traffic pages above the fold as the highest-leverage areas. There's just a certain testing framework we've picked up over the years from working at different agencies. It's trained on that. So it has the CRO basics. Then it has our data — at the time IntelliGems MCP, now GA4. We have the Heatmap API. So it's pulling in data from multiple sources and analyzing that data in dashboards.

And then where I think is the coolest part is the qualitative stuff. I have this one really big customer intelligence markdown file that analyzes a lot of subjective insights. It has our Junip MCP reviews, it has RichPanel CX tickets API, it has OuterSignal MCP, it has ListenLabs — which is like a customer survey and focus group tool we use for prospects — we can build panels and send surveys, do customer interviews. And then we use Typeform as well for internal customer advisory stuff. So Claude analyzes all of that.

I have skills set up — any of the skills, like my CRO skill or landing page building skill, anytime they run, they actually refresh those APIs and pull new fresh info into that customer intelligence markdown file. So the reason that's important — and it's a long way of saying this — to your question of where do you even start and how do you know what to test: I just ask Claude. It'll analyze all of this. And I think that's been the biggest factor.

Then I can go and design four variations of each test, stack variables, and be able to design and develop features. Claude knows from analyzing that markdown file that shade matching friction is the number one issue. I then have a skill that does competitive intelligence — looks at eight different websites, sees how they handle that friction, what they do on their PDP sections, builds a mockup of it. Once we get it in a good place, it gets that over into Shopify. So that's a little bit of the workflow that's been really cool.

Human in the Loop — Where Cody Gets Involved

Andrew: Yeah, that's awesome. I have a couple of follow-ups on that. The first question is: when you start running this process, at what point are you as the user getting involved? Is it not until the page is fully developed, and then you start providing feedback? Or do you have checkpoints throughout where Claude says, "Hey Cody, here's what I'm thinking — what are your thoughts?"

Cody: Yeah, that's a good question. Every phase I get involved. Usually there's a plan. I'll do a brief first — like, let's say we're trying to come up with a test or I'm trying to build these sections. There's some component of a brief where I say, "Hey, build me an HTML file, pulling from these sources." And obviously I'll build a skill for building the brief itself. Part of it I like to see, because everything has to be human in the loop. You can't just do an AI swap for everything.

I like to see the brief — it has the insights, the data, the qual and quant, the test plan. I just try to think of it like: what would a world-class CRO or digital product function look like without AI? And like, what would I want? I would want briefs based on user research, based on data, based on competitors. So I just had Claude doing that. I'll review that, give feedback and notes, kick it off. Then we go to the next phase — wireframes, mockups — and it'll do some mockups. I'll give feedback on those. And then once we get those in a good spot, there are probably multiple rounds of revisions. We'll then take that into Shopify. There's a ton of back and forth there.

I want to be super clear: this is a really poor use of my time. I do it at night because it's fun. I do it on weekends for fun. There are multiple, multiple rounds of revisions. This is not a one-shot thing. I can one-shot a page to 80% quality. That final 20% — that takes a long, long time.

Andrew: I certainly hear that. And getting the page into Shopify and ensuring the API is firing correctly can be such a pain in the ass.

Building Ads and Landing Pages Together

Andrew: The other question I have is: it sounds like you're using a lot of social listening tools to help build these pages. It's almost like a chicken-or-the-egg situation — are you thinking, "I'm going to build this page first and then build ads around it?" Or, "I have ads that are performing well, so I'm going to build the page for those ads and then build additional ads on that angle or persona?"

Cody: Almost like thinking of them together. So some of them — for the landing pages I probably have five or six templates. Those are like different pages in Shopify — their templates. Call it a listicle, a performance LP, which is more just your standard LP with a buy box, an enhanced PDP that looks like a PDP but is conversion-optimized, a "10 reasons why" advertorial that I adapted from Zach Stuck — so there are like four or five of those.

Those are the building blocks, and Shopify can duplicate them, you can put copy on them. My goal — and I want to chat with you about this because I'm not there yet on the outside — is to just have an idea and be able to build a funnel. And to me, a funnel is a toolkit of ads — call it 12-plus ads, diversified creative, statics — like a minimum viable test — and have that go to an adjacent landing page.

Andrew: Cool. I like that a lot. Honestly, using social listening insights to do CRO, and then to do ads — I've been doing the opposite, just using insights from ads to generate landing pages. But by that point, the angle may have fatigued out. So for discovering net new angles, that approach makes a ton of sense.

Why QC Agents Aren't Enough

Andrew: The other question I have — people love sharing their CRO agents. And I have a CRO QC agent that I think is pretty good. People ask me all the time: "Can I just have a bunch of QC agents go through this? Why do I have to be involved at all?" Can you talk to why just having QC agents isn't going to solve your problem — and why human judgment and taste matters here?

Cody: Yeah, I mean — so how are people using QC agents? Like, what components of it?

Andrew: For example, like your page should have social proof, a block with all the press articles that mentioned you. If you don't have these 20 things, your page won't perform — that kind of thing.

Cody: Yeah, I don't know. I think about it the same way you'd manage people. I was actually doing something earlier — I was whisper-flowing and giving feedback, and if somebody walked into the room, they would think I was on a call with a person. I was just like, "No, this is the framework we need for this" — we were redoing our KPIs.

I just think about it as managing people. And it's the same way — you can also choose to take junior or mid-level employees and not give them a lot of clarity and direction and not be in the workflow, and see what you get. Or you can give them clear direction upfront, proper context, and help them tweak along the way. I'm not trying to completely remove myself — that's where you get slow output.

At the very least, you have to spend a ton of time building. I put an ungodly amount of time into building our design system — a painstakingly large GitHub markdown file of our design system. That took so long. Now I can one-shot a page and have it be 80% there. But it's still not perfect. AI is not deterministic, and it still messes up some stuff. So you have to be in the loop.

I think building skills and context layers are the highest-leverage things you can do to allow you to be more hands-off later. But it's not just like, "Let me set up this thing." That's all just AI engagement bait on Twitter — "Oh, I built this thing that's launching 1,000 ads." That's where I've made mistakes on ads. On landing page stuff I've been much more detailed. When I've tried to do ads and said, "I'd love to launch 1,000 ads" — I'll get a few good reference ads and try to scale it, and it just becomes slop.

Will: Yeah, I completely agree. I hired someone to sit in Claude Code and generate ads because I tried to do it initially — just "generate me 100 ads, here's social listening, here's ad performance" — and the outputs are just not good.

What Systems You Need Now for Higher-Converting Creative

Andrew: So let's go back to creative overall. I think a lot of folks are trying to solve this with AI — we're going to take over creative and figure it out. Will's putting out a ton of proposals on this from Selfmade. Two questions, and maybe you both can answer: first, what systems need to be in place now to make higher-converting creative? And second, where is the creative workflow going to be by the end of 2026? Because right now it feels piecemeal. Nobody has mastered the workflow from brief to ad — all those pieces, bringing the teams together. And Cody, you've been really spot-on with predictions, so I'm curious what you think.

Cody: I haven't mastered it either. I'm a laser-focus person and all my effort has gone into the CRO stuff. Will is way ahead of me on image gen and that side of things — I want to get a lot better there.

I think it's interesting. I wouldn't want to be a software icon-builder or design tool right now — I think a lot of that is getting commoditized by people who can just build stuff. I consider Claude like the white-collar operating system — like, you never have to leave Claude and you can just do everything from there. Think about it: we have Motion MCP pulling in, we have Polar MCP, we have Meta MCP, we have Shopify — you just have all of your sources. Getting all of that stuff in one place is incredibly valuable.

One of the things we're working on internally is — even if you don't know where it's going to be valuable — getting people to change their workflows and behaviors. If they're used to using Docs or Sheets, and now you're asking them to do it in Claude, it's very different. But I just know that having everything connected in a context layer and talking to each other is going to be massive. I don't even know yet what all the value is going to be. But I think context is the most important thing — the models are great, they're all good, but actually having your context is the key. So the more you can document and orchestrate that stuff, the better.

But also — Meta is probably going to do a lot of it. That's what I think is going to happen.

Andrew: Yeah, I think that's true. We now have like 70 members connected to the Foxwell MCP — our knowledge base — and that's every call transcript we've ever had, everything in Slack, every SOP. What you said about landing pages being incredibly valuable in terms of bringing an answer forward for what's really going to work — I think that's a good point. But yes, Meta is going to do a certain percentage of this. The challenge is really sitting down and looking at where the inefficiencies are in your own workflow and how you're trying to solve them. I don't know, Will — what do you think people can do better now, and what does it look like in the future?

Will: For sure. I think there are sort of three pieces of the equation, at least as it pertains to ads and by association landing pages.

The first one is a gap analysis — you compare and juxtapose yourself to others in your vertical. What formats, personas, angles, and emotions are they testing that you're not? Index all of your ads — Claude's on 4.6, just have it go through all of your ads, look at your personas, angles, emotions — and say, "This is how I compare to my competitors." That's piece one.

Piece two is a time series analysis — you use the Meta API, pull in all the ads you've ever run for the past 365 days or five years or whatever, tag them all with the same taxonomy — persona, angle, emotion, format — and then you can say, "You know what, I turned this ad off a year ago, but I haven't run that persona/angle/format since." So you're able to resurface old winners rather than lose all those compounded learnings. A lot of brands have had so many successful equations in the past that have just been lost to time. It's such a pain to go back through your Meta ad library.

The third piece is social listening — perpetually scraping your reviews, subreddit comments, Facebook comments. If you add those all together and give them proper weights, you have a really good sense of what you need to be creating. Here are my gaps from the gap analysis, here's what's worked in the past, and here's what consumers are saying about me right now.

The cherry on top is using Reddit, TikTok, and Google Trends to see if there's anything trending right now that you can quickly spin up ads for. If you can spin up ads within 24 to 48 hours — I know it's hard, but it's actually not that hard — and sort of add that as the gravy, then you're going to be in a really, really good spot. That's my equation and how I think about things.

Andrew: Quality is important, but strategy is important — getting the right data, the right angles. And just speed to launch. I'm making that a core KPI of my growth team — a 24-hour funnel launch time. We review data on Monday, we have a deck, we're trying to automate that deck, but basically: here's an idea, this angle has taken off, here's an iteration on it. Pre-AI, teams were better or worse at this, but it just took too long. You'd get an idea and it's like, "We'll brief that, put it in Asana, go to design revisions, build a landing page" — it's a month later and you've missed the opportunity. No — let's get that live tomorrow. And even if that's just a V1, you can go brief UGC creators on that and scale it up.

Cody: Can I tell you where some of this comes from? I had a landing page somebody sent me — a new template I hadn't seen before. I sent it to a few people, including Zach Stuck, and Zach said, "I'll tell you how it's doing in 24 hours." And it haunted me. I was so upset. I was like, that's what it takes to be successful. And if I asked my team to launch this, it would have taken an embarrassingly long time.

Will: Yeah, I mean — so much of it is going through and assessing the systems you have and completely adjusting all of this. It's a totally different ball game as it relates to digital marketing now. So much of the AI stuff is focused on ads, but it's not — we all know that launching ads that match to landers in a persona-driven funnel is ultimately what helps scale big brands. And there were way too many tools in the middle. It doesn't need to be that way anymore. Too many people in the middle too.

Cody: Yeah — you have to brief a copywriter, brief a designer, you have no idea what their priorities are or their workload, and then you have too many approvals. Versus now it's like, I have an idea for an ad, I'm going to make some ads. I have an idea for a landing page, I'm going to make a landing page. That's so much of the actual value — you don't have to rely on other people to produce stuff.

Incentivizing Employees and Creator Programs

Andrew: I agree. I also think there's not enough employee alignment on this. There's not enough incentive for them to be faster. This is something I've been working on with a couple of agencies. How do we design systems where, when the employee saves 40% of their time, there's a financial incentive tied to that? Cody, you said before — if you were starting an agency today, one of the three things you would do is offer AI automation as a service to brands, audit and automate, and take a percentage of the savings.

Cody: That was an old tweet — I forgot about that one.

Andrew: We went deep cuts on our research. But my point is, I think a lot of it is employee incentive. Okay, look — if you're going to do this, we're going to save X amount of time. Let's put that into dollars, and that's going to be a bonus for you. That's a big part of it.

Andrew: So when you're utilizing Claude and building all this — don't you ever run out of storage and tokens? How do you functionally do this? Because from my end I'm hitting the wall, and I feel like a lot of people are.

Cody: I have my whole team on Claude Enterprise. Most people are on the Pro plan. Power users either request extra usage and we approve it, or they're upgraded to the Max plan — which I think is around $100 a month — if they're proving it's worthwhile. I'm on a personal $200-a-month Max plan. And as a power user — outside of that one week period where they had all those usage limits that were clearly bugs — I never hit it. I push it really hard. I use Claude 4.7, extra high. I mean, I don't use it that much nine to five because I'm busy in meetings. But mornings, nights, weekends — I almost never hit the limit.

I've thought about it, though. What would I pay somebody to run these CRO tests? I'd pay thousands a month. I'm paying $200. If I had to pay $1,000 a month, it's legitimately building that much value for us. And I've also canceled three tools that'll save us a few thousand a month — not only saving money, but those tools were actually inhibiting our workflow.

Andrew: One thing — you talked about your creator in residence program and shipping daily content. How is AI playing into that?

Cody: We're not there yet. It's been hard to build and just find the right people to lead it. We're doing stuff, but it's been a slow start.

I think as big of an AI bull as I am, I'm actually pretty bearish on AI content. I think static ads are one thing and it's great, but when I see people talk about AI UGC, I'm just like — you're so far off. I've read some good stuff on this lately. The ClickUp founder did a really good tweet after their layoffs — he had some really good points. He was talking about essentially automating as much of the administrative stuff so you can spend more time talking to customers. How can you automate as much of the operations, the backend, the analysis so you can spend more time being creative?

That's what I'd want my creative strategist to do — just think. Where creative strategy used to involve poking into Motion, building reports, spending a lot of time in an ad account — hopefully that can just be handed to somebody on a silver platter: "Here's exactly what the data is from the last seven days, and here's the ad you should make." And then maybe that changes the role. Everyone's talking about the role of the media buyer changing — maybe it changes the role of the creative strategist too. If the analysis can be done by anybody, I actually just want to hand that to a creator who really gets social content.

So that's my goal — we have some dashboards that show our top ad account data but also organic social trends. Hopefully people — this is my vision — we have the studio built out. If you're listening, I just need the person to run it. The idea is to have these dashboards on a big TV in the studio, like AirPlayed: "Hey, here are four videos and the top trending thing — acne-prone skin is trending on TikTok right now, here are the top four viral videos. Make a version of that." So long answer to say: not automating the creative, but automating all the prior stuff — the project management, the insights — and the post-production as well.

Getting Teams to Actually Adopt AI

Andrew: I want to sort of go back to something we were talking about earlier. Similar to you, I'm kind of addicted to Claude Code. I think there are certain people that just find it impossible to pull themselves away from. How do you incentivize your team to have that same level of commitment? I've done certain things internally, but I'm curious — Cody, it probably takes you like two or three hours to do a landing page now. If you were to hand it off to someone on your team, it'd probably take them five days. How do you reconcile that, and what are you doing to teach them or hand things off?

Cody: Yeah, it's a good question. First of all, it kills me because it's so hard — it's unbearably hard and we're working through it right now.

So I have a Director of Data and AI who's leading all our initiatives internally. He's a data analyst by trade who was super passionate about AI, so he's leading everybody. One thing we did is get everybody on a Claude plan, and for a while he did weekly trainings. He'd teach them Claude co-work skills. He did a mandatory one on what a GitHub repository is. He did one this morning with a small group — just like, "I'm going to build something live; who has something they want to build?" Trying to get people excited and show them how it can help.

I would say it definitely helped. Everybody is using Claude daily, some more than others. I would say the highest level is obviously him. One of his goals — and one of my goals for him — was to get our data warehouse into Claude. We're working with Polar to do that so anyone can get insights from it. And we're building a lot.

So our demand planning team built a demand planning software in there. That's replacing an $80,000-a-year software, and doing it better actually. So on one level that's happening. Then I would say like a bell curve — most of the org is like, "I'll use some skills, I'll try this stuff," and then there are laggers. It's really hard. He's even noticing — meeting with people, doing some internal change management work — people are really resistant to changing their workflows. Probably multiple reasons. First of all, they just don't know what it's capable of. Sometimes they're too busy. And they worry that they're automating themselves out of a job. That's definitely a concern.

So it's really, really hard. What we're noticing is — and the ClickUp guy talked about this — you have to create disruption. I have one person who's doing a really good job using it, and she hasn't built stuff herself but she's using it for workflows we've built for her. Part of the reason is she's got additional scope while losing people on her team — one's on maternity leave, one just left. She's had no choice but to use it. The people who are comfortable don't feel the urgency.

I've built five landing pages — and by "built" I mean I populated the templates with content without ever going into Shopify. I used Shopify CLI to do stuff you'd normally do manually, because I'm not going to sit there and upload images one at a time. But I find most people aren't doing that. They're just comfortable in their old workflow. So I don't know what the full answer is. I just know we have to disrupt it. I tried to be like, "Hey guys, this is exciting!" And now it's like — no, you f***ing have to do this. If you want to work here, you have to follow this process. You have to put this information here. It's just a new way of doing things. I've been unimpressed with how people think they're using Claude and they're actually using it to like 10% of its ability.

Will: Yeah, I think a lot of it is — whenever you have employee stuff and you're trying to get them motivated, the number one thing to look at is: what are they scared of? They're scared of automating themselves out of a job. So discussing it and saying, "Look, this is not something I intend to replace you for. I need you to know I'm doing this because I'm trying to make your life easier and your work more efficient." Doubling down on that.

Also, giving them a roadmap of what you hope it looks like — even if it's not what it'll ultimately be — and refining that every month, going through and saying, "This is what we're trying to do, these are the things in my head, so you have more transparency." So much of it as an employee is just guessing what the leader is trying to do. The more transparently you can talk about that, the better.

The third one we talked about is incentives — absolutely incentivize people. If and when you're adopting this, if you can prove you're saving money, there are financial incentives aligned with that. This is just like when Toyota came up with the Kaizen process in 1980 — they gave people a proportional amount of money based on the percentage of time saved. That's how everybody became obsessed with incremental efficiency improvement. That screwdriver is positioned too high — let's move it down here. It was incremental, not huge. And I think that resonates with people, because a lot of people sit down to do this and it's like sitting down to write a book — you don't know where to start. The answer is: start small. It's brick by brick. Audit your time, understand where AI can help, start there.

And ultimately — show people the possibility. You're going to keep working here, you're a great employee. If you want more time with your family, this is one way to do it. This is a way to make your work more efficient. Incentivizing them that way too, not just financially.

Cody: First of all, you're just so much better at that than me. I just try to brute force it. You're so much more thoughtful and empathetic — like, "Oh, here's what people are actually thinking." I'm like, I just want people to use it.

But how do you feel about just making it mandatory? Do you think that's a good idea or a bad idea?

Will: Yeah. As a business owner and as the leader of any organization, you have the right to make anything mandatory. This is your organization. The issue with mandatory statements is — if you come in and just say, "This is mandatory, period," that's a problem. What typically happens in corporate settings is they'll send a paragraph from the leader, and people are like, "What is this?" There's no runthrough of the why.

If instead you go through and say, "Look, here's the situation. We're making AI mandatory in 40% of what you're doing. I want to make sure you're all there. Here's why it's important — it's going to allow us to scale, it's going to allow you to spend more time with your family, more time to think strategically." That is a massive leverage — being an emotional leader and saying, "The fear I have is X, Y, and Z. The hope I have is X, Y, and Z." Speaking about it honestly. Checking back in on the mandate every month. And also saying, "I may not make it mandatory in the future if I feel like it's a bad idea — but right now, it's absolutely something that has to happen."

And allowing people to disagree. "If you vastly disagree with me on this, you can come to me — you're not going to be fired for telling me if you think this is stupid or could be better. I want you to know you're empowered to come to me and say this is a problem." So much of getting people motivated is showing them the why and what you're hoping the future looks like — not just a dogmatic statement.

Cody: Makes sense. Very thoughtful. Much more thoughtful than me.

Will: Well, this is like so much of what I do now — helping organizations and agencies that are in the founders community, coaching on employee stuff. AI is obviously a big topic because everyone's trying to figure it out.

KPIs in the Age of AI

Andrew: Good episode. I have one more question — do you have time, Cody?

Cody: Yeah, yeah.

Andrew: The question is about KPIs. You mentioned you're redoing your KPIs now. There's a lot of not focusing on in-platform metrics, as you've said, and focusing more on incrementality. How are you reshaping KPIs in the age of AI to make better decisions in the short and long term?

Cody: Expectations are just higher. Expect that people can do more, get more done, move faster. For us, KPIs are tied to bonuses — to their overall rating, but also bonuses. If I'm going to be paying out based on it, it's gotta be pretty high expectations — what I think people are capable of and what best in class looks like. I don't think we're there yet, but I'm definitely going to set expectations there.

So when it comes to the volume of output — number of ads launched, funnels launched — definitely trying to pair those hand in hand with it. You can just get more done. You can just get more done.

Andrew: It's interesting. I'm curious about your philosophy. It's not KPIs per se, but one of the components of our performance evaluation is AI usage and utilization. Last year we did a bounty — whoever did the most with it won a prize. This year it's like, all right, cool, that was fun. Now we're making it mandatory. It's just our general expectation — the same way you're using a computer and Slack and email, this is a tool. We will train you on it and invest in it, and here's why we're doing it. So it is part of people's evaluations, reviewed monthly.

Cody: Is that just like token usage?

Andrew: No, because I don't want to incentivize that. They kind of have to prove how they're using it — how they're improving their role and their function with it.

Cody: Yeah, that makes sense.

Andrew: That's interesting. Cody, thank you for being on. I appreciate it. Have a great rest of your day.

Cody: Awesome. Thanks, guys. Thanks for having me.

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.

Stop Starting From Scratch

Andrew Foxwell & Will Sartorious open with a quick roundup of what's new in AI before diving into the main interview. They cover GPT-4o's image generation model beating Gemini in a 15-ad cookoff, a new Alibaba animation model called Happy Horse for animating static ads, Claude's new design tool, a Canva feature that separates layers on AI-generated ads for editing, and a joint venture between Anthropic, Blackstone, and Goldman Sachs that signals a shift toward AI-augmented service businesses.

The interview with Grant Hushek focuses on practical AI infrastructure for teams. Grant is an AI consultant specializing in education, adoption, and implementation. The bulk of the conversation covers skill files and markdown files, what they are, how to build them, how to share them across teams, and how to iterate on them. Grant explains how he connected Fathom (his note-taker) to HubSpot to auto-update CRM records every two hours, and he shares his three-iteration framework for building high-quality skill files. The episode wraps with a debate on Cowork vs. Claude Code and a practical screen-share walkthrough Grant does live on the call.

Key Takeaways

• Why this one file type is the single most important building block for getting repeatable, high-quality AI output in your business.

• Why you need to be sharing AI skill files with your team so that everyone produces the same level of output.

• The first four markdown files every company should create when getting started with Claude.

• Why should you be talking to Claude instead of typing when building your "About Me" and voice & tone files.

• How this three-iteration framework turns a bad first skill file into one you'd actually be proud to use every day.

• The real difference between Cowork and Claude Code, and the one that is actually right for your team.

• The Process Grant used to automatically update every CRM contact record after every call.

• Why giving Claude a PDF template tanks the quality of the output it produces and the file type you should use instead.

Tangible Links from the Show: 

• GPT-4o (GPT Image Gen / "GPT2") - Image generation model https://openai.com/chatgpt

• Happy Horse (Alibaba)Animates static AI-generated ads https://github.com/ali-vilab/MAGI-1

• Seed Dance (TikTok)Previous go-to for ad animation https://seedance.tiktok.com

• Claude / Claude DesignUsed to build a new website over a weekend; Claude Design mentioned for one-off projectshttps://claude.ai

• Canva New layer-separation feature for AI-generated ads allows element editing https://canva.com

• Fathom AI note-taker connected to HubSpot to auto-update CRM records every two hours  https://fathom.video

• HubSpot CRM platform that receives Fathom transcript data via automationhttps://hubspot.com

• Whisper Flow Voice-to-text tool recommended for capturing voice & tone authentically when building markdown fileshttps://whisperflow.app

• Vercel Deployment platform for apps built with Claude Codehttps://vercel.com

• GitHub Used alongside Claude Code for spinning up landing pages https://github.com

• Notion / ClickUp https://notion.so / https://clickup.com

To learn more about Grant and his team At Grant Bot head here https://x.com/GrantHushekgrantbot.co

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 converstations, like this one being had go here: www.foxwellfounders.com

Full Transcript

(00:03) 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. Well, welcome to the AI D2C WTF podcast.

(00:26) And let me just say by starting, this podcast was really for helping you get tactical information about how to use AI better in your D2C e-commerce agency or brand, right? I mean, that's basically what we're trying to do here, isn't it, Will? Yeah. I mean, very simple. How to make your stack, whether it's generating ads or just your stack in terms of your operations better through AI.

(00:54) So one of the things that we have is we're interviewing someone for most of the episodes that we have, and we have a bunch of episodes we've already recorded. So be on the lookout for those as they come. This is the first one, but we're going to try to, as much as we can as well, talk about what's new in AI at the time that we've recorded it.

(01:11) There's just changes by the hour, right? In the recent weeks that we're talking about. I mean, certainly we could talk about the new model on ChatGPT or what are we calling it? GPT2 or whatever it's called for image generation. So what do you think about that, Will? And then what are some other things that have happened recently? Yeah.

(01:30) I was sort of bearish on GPT's image modeling and just GPT in general for quite some time, but I think it's sort of twofold. One, the GPT2 image model sort of blows Nano Banana, the Gemini Google model out of the water in almost every stat. I ran a bit of a cookoff in 15 different ad types and GPT2 won all of them. The caveats are it's a bit more expensive, but ultimately, you know, in terms of product fidelity, in terms of matching your fonts, in terms of matching your colors, it's pretty unstoppable.

(02:08) And obviously we're going to continue seeing these components changing, but I would say bar none, this is the model to be using if you're going to be generating static ads. And then subsequently, if you want to be animating those ads, the newest model from Alibaba, Happy Horse, I think is what it's called.

(02:28) You know, who comes up with these names? God only knows. But it is, it's really quite good. You know, Seed Dance from TikTok was sort of reign supreme for a while, but now sort of the Happy Horse model is great for taking static ads and animating them. I think so that's good to know. That's a that's a change that we've seen. What other new announcements have you seen that have sort of hit you and been like, oh, wow, that's that's something that's really cool.

(02:53) I mean, one that I saw that I thought was really cool is sort of the operating system. It seems that Jacob has put out that a lot of people are curious and using and it allows you to spool up agents. I'm not really sure what it does 100%. Maybe you can help me understand that. That seems to be something a lot of people are talking about.

(03:11) Yeah, I'm honestly not too, too familiar with Jacob's setup. Yeah. So he put out this thing that's it's called HQ and it allows you to share knowledge, skills, workflows and API access across the organization. So everyone works at the same foundation. Nice. OK. I mean, this sounds super promising. I would say something we struggle with internally and I get on client and consulting calls and all the time is where should I be putting on my markdown files or where should I be putting all my agents? You know, right now they just sort of live locally on my computer.

(03:42) How do I share them with my team? This is definitely a pretty massive solve. I can't say that I played with it, but it sounds pretty promising, I would say. Awesome. Yeah, I think it's interesting. I mean, I think a lot of that is like the next phase of AI seems to be the organization of where we put all this.

(03:59) And so that clearly makes a lot of sense, which is pretty cool. Outside of that, Claude, our best friend lately, anything that Claude has has put out that you've been like absolutely loving or utilizing recently? Yeah, I would say sort of twofold. Obviously, the Claude design came out. You'll hit your limits pretty quickly.

(04:17) The limits very fast. Yeah, very fast, very fast. But we built our new website on Claude design. It took me a weekend to do totally nailed brand guidelines. I would say for one off projects like a website, it's great for iterative processes. So you prefer to use Claude code, especially with CRO landing pages.

(04:38) Claude code is still bar none the best. If you combine Claude code for Cell and GitHub, you can spin up landing pages really rather easily. The other thing that is, and I'm sure most folks listening to this may or may not know about, is on Canva, they dropped a tool that allows you to upload AI generated ads. And what it will do is it will take each layer of the ad you created and separate them.

(05:06) So you can move elements around, you can change your copy. Folks have had issues for ages, you know, editing AI images. And so like that has been a massive unlock for us, you know, for, you know, for allowing our editors to, to clean those minor things. Be more flexible with it. Yeah, nice. Cool. Totally. And then I think, you know, the most exciting announcement probably unrelated to marketing was sort of the announcement of the JV with Anthropic, you know, with Blackstone, Goldman, et cetera, sort of saying that it's going to be spinning up this

(05:39) service oriented consulting service to ultimately implement AI within enterprise businesses. And I think folks have been talking about this for a while and saying that service augmented with AI is going to lead the pack in terms of company types. More, we had SaaS for a while, but now it's service plus AI. And this almost sort of confirms that hypothesis that there is a lot of demand for this and both Anthropic and OpenAI, you know, OpenAI is about to announce it as well.

(06:16) Doing this almost confirms that, like, if you are able to effectively better augment your agency or your brand with custom tooling with AI, you are going to be at the forefront of this, of this bubble, if you'll, if you want to call it, even call it that. Yeah, it's an, it's an exciting time. I agree that, you know, you going through this as a person that's learning and listening to this podcast and soaking up new knowledge is, you're at the leading edge.

(06:43) And so good for you for, for continuing to follow along and yeah, seeing the big player validated is huge, right? So the more that you can help people navigate this, the more that you can experiment yourself and be able to augment the service offering that you have is also a huge with this. So it's a, it's a fun times.

(07:00) Our first guest that we have on the podcast is Grant Holschick, who is really interesting. I mean, we talked about a ton of Claude stuff in this episode. We talk about a ton of the structural ways that he's building things. Thomas is also in this episode outside of a cafe in Prague, drinking a glass of champagne, which is just like, so you can tell by his vibe, the questions are quite good.

(07:22) So let us know what you think, Andrew at foxwelldigital.com. I'm always here to receive anything from you and these episodes are going to keep coming, right? So if there's people you want us to interview, there's people you want us to talk to, we're happy to, and we're really excited to, to dive into it.

(07:36) We go into these episodes with a lot of curiosity and that's our goal. So without further ado, let's take it onto the interview. Thanks for coming on the show. We appreciate it. Yeah. Thanks so much for having me. I thought you were going to try to say my last name and I was like, let's see how this goes. Nope.

(07:54) I thought about it and then I didn't. And then I didn't. So I'm just going to leave it for now. How do you pronounce your last name? Let's just see. Hoshik. Hoshik. Okay. I always thought there was an L. So this is why this is the problem that I have because I've said it wrong to other people. Hoshik.

(08:12) So Grant Hoshik on the podcast and we are glad to have you. By the way, so just like a 30 second brief intro, what do you do with AI now? Yeah. I'm an AI consultant, education, adoption, implementation. And what that means is that I am actively processing transcripts, processing client needs, processing my team's communication into something coherent.

(08:37) Right. So that could be making skills for other people, making skills for myself, making scheduled tasks. It runs the absolute gamut. Well, so just, you know, you see a lot of implementations in business. And my first question is, what are some of the biggest unlocks you've seen for folks in like the last 30 days that you're that, you know, you feel like every time you show the super person, they're like, oh, that's an insane idea.

(09:02) I think it's a lot less sexy than people will hope to hear. But really breaking down the fact of like, what a skill file is, when you think about repeatability and leverage in an AI world, it all starts with a simple markdown file. So like, I've spent 90 whole minutes explaining what a skill file is and how to make them and how to think about them to companies.

(09:28) That has been the breakthrough moment. But the more fun one is that I've connected Fathom, my note taker to HubSpot, and every two hours, it updates every single record that of people I talked to. And that that's like my happy place. I'm so happy. What might be helpful is I think folks maybe get a little confused about the difference between like a skill versus a markdown file.

(09:48) You know, what are these different components? Like, as I understand it, and I'm, you know, playing dumb for a reason, they're just like a block of text, right? Like, what exactly are these things? And why would you not just use Claude Chat directly? And what is the purpose of even using them? Okay, so a skill file for someone listening is it's essentially like an SOP, right? It's a standard operating procedure, but for AI.

(10:12) And so there's two main parts to it. There's the description, which is formatted in YAML, which you can just ask Claude to be like, write this in YAML. And then the real meat and potatoes of it is markdown. And markdown is the coding language, the formatting language that Notion runs on, ClickUp runs on.

(10:30) Anytime you do like a pound symbol, and then the text gets bigger, that's markdown. Okay. And so what Claude does, it takes your process or your SOP, the way that you do things, and turns it into this file that it's going to follow step by step when you request that output or that skill. Okay. And so most people in their role, they have a collection of tasks and deliverables that they're responsible for.

(10:58) And you can make a skill file for each of those deliverables. Skill files can be updated in any chat. You can ask a skill to run. It can make an Excel file, it can make a document, it can make a PowerPoint, an HTML file. And you can say, okay, I didn't love how you did this part, just update that part of the skill. And then next time it'll get better.

(11:18) And so why do people care about this is that when you think about Claude and co-work and Claude code, I talked about leverage and repeatability. So if you have a skill file that accurately and consistently gets you to a high quality output in a 10th of the time, then you can use that programmatically through a scheduled task or just in your chat, and you will get the same high quality outputs more consistently using less time.

(11:45) And so the skill file is kind of the key that unlocks AI leverage in each individual role and ultimately in a couple. And for me, Grant, talking to you in Lisbon a couple of months or like a month ago, one of my biggest takeaways is that one thing is the skills files and using it for yourself to, you know, make yourself more productive.

(12:06) But you talked about sharing skills with the team. And for me, this was like, I haven't really thought about it that way before, but I've now started doing it and it's like a superpower unlock. But could you talk a little bit about how that works and what that is and maybe some really practical tips on how to do that? Yeah, absolutely.

(12:27) So why do we share skill files? Well, for one, if I'm making, let's say I make a lot of branded documents. So branded documents could mean SOWs, scopes of work. It could mean a prompt that I want to share with one of our clients. So I make a nice document explaining what is the prompt, how to install the prompt, how to use it and what you should get from it.

(12:48) I want those to look nice. And so this branded document skill is something that I had been using personally for months. And then I realized that my engineers are also making deliverables and things that memos, briefs, etc, for our clients to read. And so I was like, okay, well, we need to get you guys the same level of output that I get.

(13:06) But I don't need to teach you what our brand colors are, you don't need to necessarily know what our hex codes are, because all that stuff is inside the skill file. So just by downloading that skill file to my computer, even though I had Claude write it, I've iterated on it probably five to 10 times, I downloaded it.

(13:23) And then I gave it to someone through Slack, I can give it to them over email and say, Hey, go into Claude, go to the customize tab, click on skills doing this all like I know exactly where all the things are. Go to skills, press add new skill, upload skill, bag the file into that little box that appears.

(13:43) And it'll upload to your, your Claude database warehouse, whatever you want to call it. Now, if you're on Claude teams, you can also go into that customize tab, click on skills, and then share a skill with your team. If you are an admin, or if your admin has enabled the team members to share those skills. And the last thing is that if you are an admin, and you have a team account, you can also create org level skills that you would then upload to the organization skills that exists in your organizational settings. That was a lot of ticky tacky

(14:13) stuff. But you asked for the practical. So I'm giving you all the all the clicks. It's good. I mean, I think what the this is exactly what we want to get into and give people the idea of how they do this. One, I have a question about Claude code. I don't think a lot of people have done a lot with Claude code.

(14:30) Or, you know, it's like people have built skills potentially, right? Or they're working on markdown files. But it's like, what are what are some of the like, top couple things that you've seen you've done, or you've seen other people do in Claude code that you feel like are massive unlocks for organizations? So I'll say two things on this one, I have a cool Claude code use case.

(14:49) So I'll answer that part first, which is, I think the the ability to, for product people, engineers, marketers, the ability for you to visualize in an app format, what's in your brain, like, you don't need to storyboard anymore, you don't need to go and like, sit on the phone with a designer or product person for an hour and say, No, the button should be like here, and it should look like an arrow.

(15:11) You don't need to do that anymore. You can go to Claude code and describe what you want. And it'll write the app. And it's very easy to now deploy those apps to a platform like Vercel. And I say that, even though I'm an AI consultant, like I took three coding classes through my university time, hated all of them, barely passed, I didn't even pass some of them, still graduated point being me and code are like oil and water, we don't get along.

(15:35) But I understand how the things connect. And so I'm actually more proficient in a tool like co work that has the same schedule task features that has the same skill features. So for someone that never really got code, co work still does a lot of good for me. But when I want to get a thought out, that's much more interface driven, like I want to click around, I want to see the thing lovable became really popular.

(15:59) But like, I canceled my lovable, because I know that I could accomplish the same thing in cloud code. So for marketers, salespeople, product people, if you felt like code was too far away from you to actually be able to do anything useful, I can tell you that if I was able to just describe what I wanted, and it made something I think that was relatively aesthetic, let's say, you can take a screenshot of an aesthetic website that you like and give that to cloud code, and it'll make something look pretty, right. So just knowing that

(16:28) that's possible, you can describe it and create an app. I think that that is that is a use case that anyone should know that they have in their back. Will, I'm sure you have other notes on cloud code, too. Yeah, I mean, I think it's also worth calling out that like, creating a markdown file, or creating a skill is as easy as going into your cloud chat and saying, you know, maybe you go into a cloud chat where you just had a very long conversation about a client, and you were going back and forth, you know, maybe cloud chat was

(16:53) helping you brief out an asset. And you're like, Oh, God, I just put so much good information in this chat. Like, what do I do now? The solution is as easy as saying, Hey, Claude, can you write a markdown file for everything we just discussed, and make it generic, so I can use it for any client, right? So like, maybe I was using a cloud chat for going about creating emails.

(17:17) And like, I thought like, Oh, this is such a great way to unlock, you know, open rates or whatever it was. Now I want to apply that to any client. So I can create this generic markdown file based on the conversation we have, we had. So Claude is always sort of coming in with that prior knowledge, right? And if you're too scared to use Claude code, you know, one, get over it, it's not that scary.

(17:37) But two, you know, your alternative is you can use cloud chat, you can take all these markdown files, put them in your projects in cloud chat. And like that works just as well. Because I know cloud code can be prohibitively expensive, because it's not offered on the pro plan anymore, you have to be on max. So if you want to just do everything that Grant sort of outlining in cloud chat, you know, that, that works, that works really well, you know, as well.

(18:00) So I guess, Grant, you know, I guess my question for you is, you know, you're chatting with a lot of companies that probably are going from zero to one, like, what is the first markdown file you tell them to create for their company? Branded docs, branded PowerPoint, about me, and voice and tone, those and about my company.

(18:22) So I'll break up. I love decks, PowerPoints. It's like, in business, for some reason, we've been obsessed with like, what the 16.9 aspect ratio, like, God bless it. Everyone loves 16.9. I don't know why. But you want to have PDF. Love it. Everybody does. One time I was in a meeting with a massive ad agency, I was working and I was a consultant with the brand.

(18:44) And they had a consultant, like a massive agency that came in and did this, I think it was 300 slide deck. And it was an all day like presentation thing. And I found out later that they were charging them $75,000 a month. And all they were doing was just like putting this strategy in this deck. I'll never forget that.

(19:02) I haven't forgotten it. Obviously, this was like six years ago. I mean, 75 grand for a deck. So there you go. There's a whole nother business line. Anyway, I'm going to reevaluate what I'm doing for a profession. Anyway, the like people really love decks. So understanding how to make a really good branding skill, I think is really helpful.

(19:23) I talked about branded documents, I make more document files and I do decks. But I got on a call with someone literally yesterday, who's in the sales team. And she's like, our sales team is consistently pulling in different features that we need to sell to different enterprises. But we need to have them all look the same while they sub out different slides, like section one versus section three.

(19:45) And I was like, okay, well, we need to see what your template looks like. And she showed me that she had uploaded the template deck as a PowerPoint, or sorry, as a PDF, instead of as a skill. And so AI, when you give it a PDF, it has to use like optical vision, instead of understanding what's in the code, because PowerPoint, Excel, Doc, those are all code based things, which is why Claude can make them with high accuracy.

(20:08) So if you give it a PDF, you're going to lose a lot of quality between what your what your reference is, and what's actually going to come out. So branded decks, really popular. And then just general branding. So where your hex colors, your fonts, etc, good to have just across the board, separate to the deck.

(20:28) And then I talked about a few personal ones. So your voice and your tone, how do you write? How do you actually talk to build out that skill or that markdown file, highly encourage you to get a tool called whisper flow. Because when you're talking, not typing, you actually maintain your authenticity and your voice. And then you have your about me file.

(20:49) Great, just a you can, I've got a prompt if you want to email me, but like, you can go and just say, Hey, Claude, interview me to understand who I am, what I do, how I like to work. And Claude will interview you and you can respond. And that'll create an about me file to give context about who Claude is working with. And the last one is about my company.

(21:07) And this is one that should be shared across the org. So if you want to have consistency, share your about my company file. And that way, everyone will have the same reference for Claude. I think that's great examples. So let's get super practical into this. I now understand I need to create these different documents.

(21:28) I understand I can chat to Claude and he might do something for me, but could you like, break it down super simple? What's the process of creating these documents in as simple way as possible? So the about me, the voice and the my company will be more straightforward, because you're not using it.

(21:47) Like your reference is your is what's in your head, right? So if you start just to jump in there, Grant. So when it's in your head, then you would recommend whisper flow versus writing, because then it's more you use more words, it's more colorful, it's more in depth, right? Yep, exactly. So in those situations, you can start a new chat or a new task and co-work.

(22:08) And just explain what you're trying to achieve. Like I want to create a markdown file that you Claude will reference when I start a new chat or task. And this is specifically like, this is really much more useful for co-work than it is for chat, because the chat is limited in the knowledge base of a project, or it's limited to your global instructions.

(22:31) Co-work can actually reference files like Claude Code. So this is when we talk about file creation, we're really thinking, okay, well, how can we get leverage on co-work and Claude Code? So stating your goal, I want an about me file that you can reference when you're talking to me, interview me, right? I also really like to use a tool that co-work will often recommend, but I asked for it specifically, which is called Ask User Questions.

(22:56) And so what this does is it just pops up a little multiple choice select. And I like that for those kind of quick, okay, well, how long do you want the file to be? Or in what circumstances will I be referencing this file? Those types of things, like it'll suggest answers that you can just select as multiple, multiple or single select and move on.

(23:14) But then you want to start getting into the whisper flow and really speaking what's on your mind. And you can go for two minutes of just yapping. And that's great context for Claude to have. So don't be shy, really let it all out. And then it'll make sense of that unstructured text.

(23:30) And so you can do that for your voice, your about me, and your my company. Those those three are going to be really good for the self interview. I've done, I think the my about me, I spent literally like an hour back and forth until it felt good. I had a really good starting prompt to work from that was like probably too comprehensive. But, you know, let it take 10 minutes, let it take 15 minutes or 30.

(23:54) Like it's a it's time well spent, because that's really strong leverage for when you're working with Claude. So just to follow ups from that. So one thing is you created the file, you start using it. But as I understand it, you can also kind of ask it to update itself. And and it's it should be like a living document. Am I right? Yeah, definitely.

(24:13) All I think all of your files, all of your markdown files can be considered living documents. Because at the end of the day, this is what I learned from automation, like edge cases exist in life, just generally, right? So if I'm making an SOW, and we have these core offers, and then someone comes to me and says, all right, well, I'll pay you triple your rates if you do it completely differently.

(24:33) And I'm like, Okay, well, I want to start from this template that I have. But let's work in this new option, right? I might want to use this new service again. So you can then say, go to my SOW skill and update it with this new this new path, potentially. And so even when you think about your about company, like if we another great example, we're growing, we're hiring, new person joins the team, part of the about company tells you what the team roster is and what each person's role title is and what they're responsible for. So if you hire

(25:01) a new person, great idea to go and update that file, simply by starting a new chat thread and saying, update my, my company file, because I've just hired Bennett, Bennett runs media buying, senior media buyer, he's going to run these five accounts. Great, that's gonna be super helpful for Claude to know that. That's great.

(25:20) And then my second follow up would be, so that's kind of the the markdown documents. So what about like the designs and getting it to make the correct presentations and all that stuff? How do you go about that? Yeah, so skill files, for the most part, are going to be just single markdown files, single text files.

(25:38) However, when we want to have reference material, like a PowerPoint template or an Excel file as a template, we might want to have our skill reference that material. This is a great example for our decks, right? And so if we are in the process of creating a new skill for branded decks, then we can provide that PowerPoint file to Claude.

(25:59) Claude will read the file, understand what's there, give it as a PPTX, not a PDF. And then when we make a skill file, we can tell Claude, include this PowerPoint file as reference material. And the big difference here is that at the end of its process of asking questions and getting answers and writing that skill file, there's going to be multiple files that it spits out.

(26:20) And you're going to want to click download all because then it's going to zip those files together and you can move the entire zip folder into Claude. And when you upload that, all the material is there. It's referenceable and it's a larger body of work for that skill to work off of. And that's where you get really specific and granular skill files.

(26:44) Great for branding, great for visual things. One thing I've always wondered that I haven't had the guts to ask someone is why and do you recommend setting up like a dedicated computer for all of this? Like, I mean, because I got a thing, you know, dispatch or whatever it is on Claude that was like, hey, you can do this from your phone and like, it's still running on another computer.

(27:09) And I don't, you know, apparently people just have computers sitting around. Like, I don't have that kind of money. Just like having a computer sitting around. But what do you recommend doing that as you continue to sort of build on this? I guess I'm just curious of your opinion about the infrastructure side of it. I'm getting into that much more right now.

(27:26) I kind of went down the rabbit hole of trying to push co-work as far as I could with this idea of like a helpful employee companion. And so what I did was for the longest time, I've had scheduled tasks working in co-work that read my transcripts, Slack, email, look up and identify open loops, things that I said I would do, but I'm not doing in the call. And it logs those.

(27:48) And I found out that I had six different areas in my computer that that information was being logged. So that's a no, no, don't do that. But then I was like, okay, well, we got to start cleaning this stuff up. And so I thought, okay, well, if I had co-work go and index that list, try to fully execute a task. And if it couldn't give it up to someone on my team and say, I need information from you.

(28:09) But an actual employee would first go try to see if the task had already been done, because maybe I do it on a call, but the transcript doesn't say that. Or maybe I sent the email, but I never confirmed with Claude. So you need to go look for proof of completion. So where I'm going with all this is that co-work has limits.

(28:24) Scheduled tasks only work on time. They don't work on tool. They're not tool-based triggers. So for example, difference is when Fathom completes a transcript, it is not going to start a scheduled task. A scheduled task can run every two hours and look for calls that finished in the last two hours, but Fathom can't start the workflow.

(28:43) Okay. So from an infrastructure perspective, I am now taking a playbook, a note out of a play out of Thomas's playbook. I've got a Mac mini that I just turned on today. And I am currently moving all of the edge pushing, the boundary pushing that I've done in co-work. I'm moving that over to my Mac mini. I'm setting up OpenClaw and I'm setting up the Gary Tan infrastructure for file storage.

(29:09) And so I'm combining all these things now into this external thing. Main reason why my computer is on fire all day. And I cycle through the battery. It is a less than one year old MacBook Pro. And I cycle through the battery five times a day. That cannot be good for me. Like I just doesn't stand to reason that that's a good idea.

(29:30) So I moved it to a new computer. I definitely, I definitely identify with this. Uh, this is why I asked because like I asked, it was doing some really complicated shit the other day. And like at the same time was also building a, um, monkey riding a toilet cart game for Nora. By the way, it's pretty dope game.

(29:47) If you want to play, let me know. You got a lot. What's the context though? Uh, she, I asked her to come up with a, um, a funny game and she came up like an idea. So I'm trying to teach her like game design and design thinking and stuff. And she designed basically this Western themed game where a monkey rides a toilet around space. It's very fun.

(30:06) And it's, uh, you have to visit different like Western themed bathroom stops. The monkey does. So she's sick. That's amazing. I'll send it to you guys. It's pretty dope game. But anyway, so, so we're like, I've got this and it's like running this other, like two other things that are like quite complicated.

(30:26) And I mean, it was, it took like the, all of it was taking like 20, 30 minutes and it was comical because my, it was, my computer was flaming hot. I'm like, I think I'm reaching this point. And so I didn't like, I, that's, that's why I asked the question. But anyway, yeah. Just a comment, Grant. So first of all, welcome.

(30:42) And it's, it's great to have you in the, in the cloud community. Uh, what I think going back to the start of your point of like building, building your skill bases and your MD files, right? So what's amazing with, um, AI going forward, for example, with the open clause where you can install it on your Mac is that I use a shared library via Dropbox.

(31:03) So everything that co-work or cloud code or cloud chat is learning and updating these MD files of the company about me. I'm also putting that into my open clause agents who are also reading and writing from it. So all the infrastructure that you have created with cloud, you can just directly put into kind of, uh, the infrastructure can keep working in, in another platform.

(31:28) And I think that's the magical part about MD files and skill files. Yeah. I'm setting up that sync tonight, late, late in the evening hours. And then just to, and just to add onto that, like, so you're new to this ecosystem. One thing I'm curious about you guys thoughts on is just like co-work. Like I don't use co-work at all.

(31:45) I just use cloud code for everything. Is there like any advantage at all in using co-work, uh, versus using code? Uh, that's my first question. Second is just like, sort of like a piece of advice for our listeners here is like your first mark dial markdown file or your first skill or whatever the fuck you want to call it will probably suck.

(32:03) Like it may not be perfect. And so what you need to be doing is iterating. Right. And so like internally we have marked on files for every different type of asset. So if I wanted to create like a before and after ad and I say, Hey, use this markdown file to create a before and after ad for me, for this brand.

(32:23) And it does, it generates a prompt for me. And I need to put that prompt into nano banana, GPT two, whatever it is. And I get an output. What I need to do is I need to take that output and I need to put it back into code. And I need to tell code, these are all the issues with this ad. You need to go back or maybe more politely, please go back and update all of the respective markdown files that you use to get to this point.

(32:45) So we don't get this mistake again. So the idea is like you're creating an iterative loop. And I will say, I may be, you know, I may be on an Island in this group here. I will say having a human in the loop doing that is far more beneficial than having a QC agent do this work. Like I've tried having a QC agent just like iterate and iterate and iterate.

(33:05) It's fine, but it doesn't necessarily get there. The one thing that humans have as an advantage, especially in the marketing space is taste. Like, you know, and we know on this call, you know, what maybe like a good ad looks like or what a good landing page looks like, right? Like we can optimize as much as possible, but like there's, there's, there is an innately human intrinsic thing about taste.

(33:25) And I think that's why, you know, having a human in the loop do this thing. But so anyway, that was just a bit of a tangent and an aside. I want to go back to the cowork versus code thing, because maybe I'm missing something here. And maybe our listeners are missing something here because I think there's a code seems to solve any problem that you could ever do.

(33:42) Can I, can I go first? I have, I have feelings about this. So I've been experimenting with chat, org, cloud, and open cloud. And it's, it's a, it's a strange harmony and there's a lot of feelings around it. But for me, what I landed on is that a cowork is my product manager that connect. It's easier to connect to all the different connectors.

(34:06) It's easier than to just make sure it has all the files connected. Cloud code for me is too complicated. It looks cool when my wife looks behind my back on what I'm doing, but in terms of just making it do consistent, good quality work for me, cowork is better. Especially in terms of making sure all the connectors and everything is working because on cloud code, I have a hard time making sure it remembers that it has to do this connector or the liner connector or whatever.

(34:36) So for me, it is my project manager that I talk about when I'm going to develop different things. And then I get that to give me the prompt I need to do in cloud code to develop. So that's my take on it. Got it. That makes sense. Maybe it's gotten better since the last time I used it, but I've just found that using APIs and code is just like so much more efficient than the connectors and cowork, because I feel like the connectors, and again, it's been a few weeks since I've done it last, but I felt like the connectors just kept

(35:05) breaking down. And then I have to go reconnect them, reconnect them, reconnect them. I was like, this is not a repeatable system. There's a very known bug about this connectors dropping down with the authentication thing going, which I think will be resolved soon. But you're correct. Some of them are breaking down and it is much better to connect directly to the APIs.

(35:28) But I think for the normals, the normies who are just using this every day, I think it's easier to just understand you have cowork, you put in all the connectors and it will just work for you. And of course, they will prompt you if you need to reconnect or re-authenticate things. But it depends on the nerdiness level, I think, of the whole delivery.

(35:47) But for me, showing cowork to different clients, which is like small e-commerce businesses in Norway, they get their mind get blown by just understanding how cowork works. And, you know, applying all the things Grant have been talking about today and just giving them MD files and creating schedules and stuff like that.

(36:09) You are like the coolest person in the office. You're like a superhero. You're like the the peak AI person just applying these simple tactics, I think. And I'll throw in on that completely agree. The percentage of people that are going to enjoy code, cloud code is disproportionately smaller than people that are going to jump in and immediately get value from cowork.

(36:29) So if you're listening to this and you're like, I don't know what to do here. Go for cowork, have a ball. Right. And to your point, Will, is this a video podcast or is this audio? We have video component as well on our YouTube channel. Yeah. Fantastic. Do you mind if I screen share something? No, let's do it. Okay.

(36:46) So this is what I was looking off to the side and in eager anticipation. So this is something that we teach a lot in our workshops. Super colorful. Effectively, to get a good skill, I tell people it takes three iterations. But when we say iterate, it's like, okay, well, what does that mean? Right? Like really tactically, brass tacks, how do I do that? So what I tell people is take, you need to have three inputs that are very similar.

(37:12) So let's say we're trying to make a Excel file of financial report, right? Or we're trying to do some sort of manipulation. You can have input one and you're going to prompt and you're going to say, all right, I really want this to actually be a cleaned up Excel file with my brand colors. And I want the tabs to be really nice.

(37:26) The first tab needs to say, start here. And then everything else needs to be logical from there. Give it your best shot. Cloud's going to make an output. You're going to say, you missed the mark about 80% of the time. And then when you revise it, you say, okay, great. I've revised what you worked on. Make me a skill to do this, like my revision next time.

(37:45) You take that and you go to input two and you say, great. Hey, here's input two. Use the skill I just created. That's going to make another cloud output. You're going to revise that because it's going to miss something, but it's going to be much closer, much better. And then you get the two of that skill. And then lastly, you take input three, which is unique from the first two, apply skill v2 to it.

(38:06) You get an output. Maybe you need to revise it, but after you revise it and you say, great, update the skill to my revision, that gives you v3. And v3 is something that I'm proud of because I've taken the time to actually iterate with it, tell it what it got right, tell it what it got wrong. And it knows I'm crystallizing my information and my process.

(38:26) That way I don't have to keep re-explaining myself. And I think this is the aha moment for people when they realize that they can go from zero to their finished product, especially in the reporting world where reporting takes freaking hours. Like some people's jobs is just reporting, right? And fine, but I want to make your life easier so that you can actually go build more relationships, focus on strategy, be creative, or do quality assurance.

(38:50) Those are the four main categories that people should be spending their time on. This is how you get that. This is how you remove copy and paste from your life. So hopefully that's a little bit more of a, like, take these steps to make a really high quality skill. This is what people pay us to do. Love it.

(39:08) Well, Grant, thank you very much for being on here, sharing with us. If anybody wants to connect with Grant, he is on all the platforms, but I'm happy to connect you with him as well. You know where to find me, Andrew, at foxwell.com. Until next time, though, folks, keep on building. Bye. All right. Bye.