Ben Sufiani: Welcome everyone to the Build Grow Repeat Podcast. Today I have with me Markus and Paul from DreamyTales. I know these guys from the cologne startup scene. They are frequent visitors to the vibe coding cologne event where they showcase their physical and digital products One of my favorites is their little vibe TV That always shows my current state of things how many tokens I have left on my cloud subscription and By the way, one of the funniest things that I learned from that is it works based on the codex bar
Paul from DreamyTales: Mm-hmm.
Ben Sufiani: and my last 30 day costs is $4,540 on the 90
Paul from DreamyTales: It's insane, right?
Marcus: What's that? Yeah.
Ben Sufiani: euro subscription. This little thing makes you appreciate the subsidization that these crazy companies do. appreciate even more. So how about you guys give yourself a little intro, tell us what you guys do and how you're separating between yourself, your business.
Marcus: Yeah, first of all, thanks for having us. we are very happy to be here today and to introduce us and our company and our business. maybe a couple of a couple of words to myself. I'm Marcus, I'm thirty-six years old, I am living in Cologne here with my wife in the southern city. And actually I am a mechanical engineer and after my studies I founded my first company in twenty nineteen. And then twenty twenty three I had a lot of luck to be able to sell that company to an investor. In parallel, I was meeting Paul for the first time and we started with our company DreamyTales with our startup. DreamyTales is selling personalized gifts for everyone. And before we continue with that, maybe Paul you can introduce yourself a little bit.
Ben Sufiani: Yep.
Paul from DreamyTales: Yeah,
Ben Sufiani: Thanks mate.
Paul from DreamyTales: hi everyone, my name is Paul. I am the other half of DreamyTales. I am 31 years old, and my professional background is kind of yeah, I'm I'm sort of a jack of all trades. I used to be a event manager and then I studied business psychology. and yeah, during that time I got interested in startups and the startup scene in general. And then had the great luck of meeting Marcus in 2023 and built our first business together. DreamyTales, as Marcus said, is a company that sells personalized gifts. So if you Ben need a gift for your fiance, say for your anniversary, you could turn to DreamyTales to turn your guys' love story into a book or personalized video or personalized song. that's basically our best selling products. And yeah, I'm also based in Cologne here and we've been doing this for the past three years now.
Ben Sufiani: Very cool. So I always found it fascinating how you guys are switching between physical and digital. I remember when we had one of our first sessions together, I was like, what the fuck is this not digital? Who comes up with the idea of doing AI generated images in a story and then actually prints it and ship it out to people? And at that time you were, think, still working a lot like with external developers who were not vibe coding these things yourself. Right? So I'm curious, like, how are you thinking about doing like physical and digital using external developers versus vibe coding it yourself? What are you thinking about that?
Marcus: So yeah, maybe maybe first things first. Actually we sh we started with a physical product. in the very beginning with Dreamy Tails we were selling children's books, personalized children's books with AI. And then we figured out pretty much very very early that this needs to be a physical product because no parent wants to hold a tablet or a smartphone into the face of their child's and it had to be something something printed, you know. So we started this this journey with this physical book, with this personalized book. We then later switched the the audience, the customer audience group to adults. and then Later on we tried to find new new kinds of products we can do with AI and we came up with AI songs and AI videos. And while doing that we figured out that it's much, much easier to scale those digital products into different kinds of markets, for instance.
Ben Sufiani: Mm-hmm.
Marcus: Right now we are selling in
Ben Sufiani: Yeah, yeah, yeah.
Marcus: more than ten countries worldwide. So for instance we are selling in Brazil, in Mexico, in Turkey, and and all those countries. and we started selling in Brazil, for instance, with the physical books and we had to find some local partners who are printing our books and distributing them. And that's very, very hard work, you know. And then
Ben Sufiani: Yeah, crazy. Yeah.
Marcus: in parallel we tried our AI songs there and figured out that it's much, much easier to scale those because you'd simply can can send this song by by a download link via email or whatsapp and that's it. And that's a great great trade-off we have with these digital products
Paul from DreamyTales: So maybe just to add here, so when back in the day when we started, these digital products weren't even thinkable of, right? So the first image models, they were so crappy that we could barely generate images to print in our books. But not to mention,
Ben Sufiani: Mm-hmm.
Paul from DreamyTales: you know, generating whole videos or you know even songs. This was kind of developing over time and we always just kept a close eye on the market and our competitors were these guys were selling. And then over time it became technically possible to build these products and that's when we tried and then discovered that it's actually way easier to scale those instead of building physical products. But I do think that physical products also have like like a bunch of pros, right? So they have a bigger
Ben Sufiani: Yep.
Paul from DreamyTales: mode around them, for example. So we see that with the Vibe TV as well So the digital products, I mean, if Marcus and I we we are able to build those with basically just a laptop, it's pretty fair to assume that other guys with a laptop could rebuild the exact same product, right? So not a real mode around this. But with a physical product, this g this gets a lot more complicated, right? You need a supply chain, logistics, handle returns, shipping, blah blah blah. So that's also some sort of a a mode you build with physical products to protect yourself
Ben Sufiani: I understand.
Paul from DreamyTales: from just random guys with a laptop somewhere.
Ben Sufiani: And I hope you guys recognize that this makes you unique. You are not scared of physical. You started there, you now understand digital and I think this is a, that's probably your biggest edge. that you actually can play both and interact between them and come up with product ideas that... normal just physical or just digital founders wouldn't come up with. so I'm very curious to see what you're building on top. But let's go back to the to the songs. I remember maybe it was maybe a year ago you were you guys were at the meetup and showing the the end it end workflow. and I just I'm sure like this was the initial version of it, but Markus since you're building them, like why end it end? Why not build like the full backend in the beginning? How did you get decent results? Make it reliable?
Marcus: that's a very good question. so first of all maybe to explain My job at DreamyTails is basically the back end. so I'm responsible for everything, what happens, what the customer cannot see directly. and Paul is
Ben Sufiani: Mm.
Marcus: responsible for the front end. So front end online shop, also communications and and marketing. And you said that I I am using N at N, that's right. as I'm just a mechanical engineer, I'm not really a software developer. So to be honest, I n I have no clue how to how to manage and program a complete back end for our shop. So from the very first beginning,
Ben Sufiani: Mm-hmm,
Marcus: we were using no code tools or low code tools. And with the help of ChatGPT, we basically built everything from scratch with these low code tools. And N at N is my best friend right now because it's a good it's a great platform. You have great overviews, you have great debugging. And right now I I would call myself a Vibe backend coder. so that works works quite great.
Ben Sufiani: That's so fun. This is a thing. Okay, questions. Where do you host your NADM and where do you source the audio and video models from, from which API?
Marcus: yeah, that's different. So Yeah, sure.
Ben Sufiani: If you want to talk about it. You don't have to, but I'm curious.
Marcus: so in the very beginning we were not even using N at N. We were using make dot com. It's a quite similar platform but much, much easier, much more basic. And we were using the the hosted programs. Same with N at N until to a certain scaling point where we said, Okay, it's much easier and much much more scalable. hosting our own server, our own virtual machine and deploying the NN instance on there. And that's currently the case right now. for models and and products we are using, this is like like a complete bunch of of of AI models. We are using some open source models on our own servers, but also like the usual ones like Chat GPT, open AI image generation. we are using Sora too for video ads and stuff like that.
Ben Sufiani: You go straight to the provider to use something like OpenRouter in between.
Marcus: it depends. So I think most most of the time we're using FEL AI, which which is quite common. It's a very robust
Ben Sufiani: Same, yep.
Marcus: API for all kinds of all kinds of AI solutions.
Ben Sufiani: And I remember that there was this guy at the meetup who showed us Feli AI. While I didn't know about it until then and he was like building landing pages and I've done so obviously also like for a while, but then he was also doing all of the imagery in the layouts of the landing page. And I was like, God damn it Why didn't you understand like that? This is the obvious fucking thing to do. And then I looked at my website and everything was free of images. And I was like, okay,
Paul from DreamyTales: Mm-hmm.
Ben Sufiani: this looks decently designed, but it doesn't feel as alive. So interesting enough, I used to use a lot of NNN and since I'm very much in the Next.js and the Versel ecosystem. I switch to their workflow SDK, an open source version of like NNN, but it just runs as part of the code base and not like as a third party thing. And then I'm sourcing stuff from, from Thel or whichever platform I get that stuff from. That was an interesting change to me and listen to what three non-developer guys are talking about. What the hell is going on here?
Marcus: Yeah.
Paul from DreamyTales: Exactly. Exactly. I remember
Ben Sufiani: The main discussion is going to be on the marketing madness and we'll get into that, go Paul.
Paul from DreamyTales: I remember when we met and it used to be all like, all right, show me your ad manager and you know,
Ben Sufiani: Yep.
Paul from DreamyTales: what's your CTR and stuff like that. And now we're talking about you know, backend SDKs and Next.js and other frameworks. I mean it's insane to me how much you can learn. Like Marcus and I, we are not developers. You are not a developer then,
Ben Sufiani: Yeah.
Paul from DreamyTales: but we are all like marketing guys, right? That now dig into building the actual products we are trying to sell, which is a great way to do things. I think it's it's it's a great skill to have if you come from e-commerce or in general digital marketing and you know how to sell stuff. Now you can learn how to actually build stuff, the product underneath it and iterate on that. I think it's a very, very good starting point. instead of you know having to know how to build stuff and then now having to learn how to actually sell the stuff you're trying to build.
Ben Sufiani: Let see if I can pull this up and hold it to the camera. I got this idea of the sovereign entrepreneur in my mind. And I always felt very dependent on developers to build the product that I was finally able to sell. Here it is. I'm not sure if this is readable. Maybe I'm hiding behind it a little bit. So like you need a founder who builds something, a marketer who sells it and an organizer who administrates all of this stuff. And you get, you get in trouble if you're lacking one of these skills. Like we marketers tend to be like, okay, we, we're founders, we organize shit and then we marketing a crappy product. That is a typical marketer trap. Or we have the opposite. We have a developer that builds a great product but doesn't have enough customers. They have great stuff but they didn't sell enough. And then we have the business types who are often just fucked because they can't build and sell. And now we are developing ourselves into these three functions. And I think that's also the main thing why my podcast here is called Build, Grow, Repeat. It's like build a product, grow it, make it repeatable in a sense that you organize and orchestrate everything in this loop. And you are very very very close to that including the physical world. Well, I'm more tied to the to digital world. But what we do have in common is that we love paid advertising. We love sending money Mark Zuckerberg's away, which is totally
Marcus: Okay.
Ben Sufiani: weird. And but we do it a lot. I don't know how many hundreds of thousands I have spent his way. But let's just, for people who are not that deep in the world, why the fuck would we even consider paid advertising as a valid approach? What makes it interesting for you? Why don't you focus more on organic? What justifies the spend?
Paul from DreamyTales: Great question. So first of all, obviously the obvious question is why would you spend money on meta, right? so
Ben Sufiani: Ahem.
Paul from DreamyTales: if if you can just post organically, we tried this for the past three years and we tried really bad. So you know, there were times where we had in-house teams and all they would do is trying to produce organic content and pump it out on our Instagram channel, trying to reach people. but no matter what we did, it didn't work. So we might have generated reach, yes, and likes and you know what have you, but not to a scale where you would say, All right, this is a you know reliable channel to build a business on. for us this has always been pretty unpredictable. Also with the algorithm changing in the background, like especially on Instagram. you know, interaction such as likes and follows, blah blah blah, get less important, but you know, swiping and hold rates and etc. get more important. all this was an environment where it felt hard for us to navigate in and to actually generate reach. So we tested this, but also simultaneously we obviously ta tested paid advertising, and this worked from the start for us. So This was a channel where we generated insights that we could actually work with. So, you know, you spend a euro and you get insights back on whether is your creative performing or not and why it might or might not be performing. So we generated more data as to why our performance was good or bad. And this gave us a base to iterate on this. And in the end, you know, after three years of doing this, all we can say is that we did what worked for us and that was paid advertising. We tried organic, didn't work for us. But you know, why should we waste time on doing more of the stuff that doesn't work when we can just spend time on doing the stuff that does work, which is paid advertising.
Ben Sufiani: Any opinion on that Marcus?
Marcus: yeah, I mean paid advertising for me, I was starting with that in twenty nineteen and it was back then it it was much easier, I would say, to have a good ROARS on on those platforms like Meta and stuff. and for me it was always this magic little button where w you are like testing an ad and then maybe iterating on that a couple of times. But when you find the the killer creative, the killer ad which performs You just have this little button, you have to press it and spend more money every day. And then it's like a like a printing money machine, right? So you're spending money, but on the other side you're getting money back, you're getting revenue. And this certainly works to a to a certain level. it's not like there's there's a certain move there. But for me it was always this this magic button which I can press. And it worked very, very great for me in the past.
Paul from DreamyTales: Also, like for DreamyTales, DreamyTales is a very seasonal business. So for us, it's very important that our high seasons, such as Christmas and Valentine's Day, that these work, right? So we need a scalable and predictable source of revenue for these high seasons. And organic is just the opposite of this, right? So it might work, but it also might not work. And like Markus said, with paid advertising, you can just scale your campaigns when you know the demand is going to be high. You get more metrics to look at. So it's way more predictable and scalable for us.
Ben Sufiani: I think so too. Like, I'm judging myself a little bit for not truly cracking the organic nut. I
Paul from DreamyTales: Yeah.
Ben Sufiani: have this stupid saying in my brain all the time, traffic is like sex, only losers pay for it. Do
Marcus: Yeah.
Ben Sufiani: you know this one?
Marcus: Well yeah, then then I think we are the biggest losers.
Ben Sufiani: Yeah, we are. to. I absolutely envy people like Daniel Minini, who's so correct, the organic nut. But his target is not a product. His target is ads that he is selling. We are the buyer of his ads, or of his placements. And it's not easy to do organic at all. The only alternative I'm very much willing to test out when I'm on my next B2C rampage, which could be Bridesmaid, the agentic wedding planner, is treating Working with influencers, small to medium sized, on a very standardized contract basis and treating that like performance marketing. This is the reach we get, this is the engagement, this is the clicks we get, this is the coupon code, entries we get, this is ROI. Yes, you can work with me again, if not change or you're going out of the rotation. To me, it's a bit more human hassle than I like. One of the things I love about Meta are Google or TikTok ads. is the, it's just numbers. It's just you click in a dashboard, you click the magic button that Marcus has and it starts cranking. The risk is that you are wasting your money. and this is why I, for example, became such an analytics nerd. It was my way to find out, hey, if I'm losing half of my marketing budget, I would like to know which half. And you have been using triple whale for that for a while and thought about like maybe building that into your own thing. But you have become very focused on what we marketers call ROAS, return of advertising spend. You spent a hundred euro. you want to get out X factor more, let's say three. that makes means gross of 300 % is one euro in three euros out just for the future discussion here that we're going to have. So how was your thinking about like a triple whale expensive? Do I build? Do I buy? What was that process like?
Paul from DreamyTales: I mean you're absolutely right. So when you scale a business, it obviously becomes a little asset heavy over time. So you buy software or you buy I don't know software solutions, you know, monthly subscriptions that just keep on running. And this piled up over time, right? And Triple Way was one of these. So as a classic D2C brand, Triple Way is just considered to be one of the Basic tools to have, right? So it's always the Shopify Triple Way meta stack that you build a D2C brand on, basically.
Ben Sufiani: Yep.
Paul from DreamyTales: So that's where we were all also running with this. And I mean for most of the time this was this was fine, right? I think the base subscription of Triple Way is $300 a month. and triple way basically gives you cross-channel attribution. and the triple pixel on your website, right? So it tells you like a customer has seen your ad on Instagram and then he googled your brand and then he bought via a third channel or whatever, right? this plus a bunch of metrics for profit and you know fixed costs, your cocks, whatever you have. So the analytics dashboard. And yeah over time we always asked ourselves all right can we maybe get rid of A lot of these really expensive software solutions that our business is depending on, right? One of them being triple whale, a bunch of others also being, for example, AI models that we pay per API request, but now we host them on our own virtual machines, right? So we switch to open source alternatives there. And with with AI,
Ben Sufiani: Awesome.
Paul from DreamyTales: this this was made possible, right? So in the performance marketing world, we kicked out triple whale and instead built a blended metric, a blended ROAS metric from our Shopify sales and the meta performance of our campaigns, so to say.
Ben Sufiani: we did dive deep into that for sure. Did you need to do a lot of for that in the backend markers or could you just, did you need to make changes to like Shopify layouts or how do I even think about doing something like that in a Shopify technology stack? Next year's I get it, I don't get it on Shopify.
Marcus: I think no, actually I I did not have to change much in the back end. I think the data is pretty much there already. The only thing triple rays does for you is to collect all this data which is hidden somewhere, puts it all together in a nice dashboard and with a triple pixel showing the the exact same the exact raws you you have on your on your campaigns, right? since we
Ben Sufiani: Yep.
Marcus: are not using Triple Rail anymore, obviously we are not having one hundred percent of the vision anymore. But with the helps of AI and codecs and skills, I think Paul is able to collect and gather as much data as necessary to have at least a clear picture of what is happening. furthermore I think triple well is a very very, very good tool, especially when you're learning, when you're when you're in the experiment experiment phase and learning phase, when you're trying new ads, new angles, new products, all the stuff, then you have to have a very, very clear picture and very good data what is what is working right and what is working not right, And and iterate iterate on those data points. right now with DreamyTales, We we've done a lot of testing, we've done a lot of experiments, and for most of our products we are having already a clear pretty clear picture what channels are great, what ch channels are not working great. And therefore we had the confidence to switch back from from triple whale to our own reporting system within Codecs.
Ben Sufiani: Tell us, what have you built, how have you built it? If you wanna show something, show it, but you can keep that for the VibeCodingCologne event if you like.
Paul from DreamyTales: so maybe to start off, it's important to understand how a Shopify website works, right? So as you said, Ben, you get it on a Next.js sort of project. Shopify works a little different. So you have the Shopify store and you install themes on this Shopify store, right? And a theme runs in Shopify's own template language that's called Liquid. And Liquid comes with a bunch of perks, right? So Liquid doesn't allow you to just you know install whatever SDK on it or in it and place whatever random pixels in it. It works, but it's a little clanky and it's not as straightforward as say you're running an X.js project. So that's why there's Shopify apps. So when you install some kind of Shopify app on your store, this app handles it all for you, right? It installs the necessary scripts in your code base, handles all the tracking, all the pixels, blah blah blah. So when you kick out triple well, you're getting rid of all of this. And now all you're left with is the meta source, right? So what we did was first of all was trying to Pull the data from our ads manager to our local machine, so to my machine, so that I could basically work with that data, right? And we did this for all campaigns via the meta ads MCP. You can do this, but you can also do this via the GraphQL API from Meta. So that's pretty easy. And then The problem is that Meta only shows you, you know, 60, 70% of the truth, right? So meta doesn't track all the sales of a campaign or misattributes them, blah blah blah. That's why solutions.
Ben Sufiani: I do that a lot, yeah.
Paul from DreamyTales: That's why solutions like triple well exist in the first place, right?
Ben Sufiani: Yeah.
Paul from DreamyTales: so you want to basically cancel out these. So you need a second data source, and that being Shopify. So for us, for example. say we have a campaign that sells a product in Mexico, right? we compare the data that Meta gives us from this campaign, and then Meta, for example, tells us we have a ROWS of 3.5 or whatever, and then we try to calculate the ROS based on the actual Shopify sales that we had in the same time frame in that country for that product, right? Because this is your actual truth, right? So you can pull down how many units of a product of this that this campaign is trying to sell in that country down from Shopify and divide the ad spend by those products to get get the rowers. And by doing so, yes.
Ben Sufiani: question. this is all like all of the Shopify data you get through your custom-built Shopify app. Is that correct? Okay,
Paul from DreamyTales: Exactly. Yeah,
Ben Sufiani: and you don't have to share it on the App Store, you can just have one privately for your own store. What
Paul from DreamyTales: Exactly.
Ben Sufiani: did you call yours?
Paul from DreamyTales: I think it's called the Codex Shopify connector or something. Very creative.
Marcus: Mm-hmm.
Ben Sufiani: Okay, yeah, cool, cool, cool, cool, cool. Yeah, I get it. So you pull these different sources and you're pulling them together in a system and you need to build a dashboard for that. need to save that data. What did you use for that?
Paul from DreamyTales: we do this in Codex. So we run the Codex desktop app that allows you to create automations. And these automations run every whatever, in this case morning. And this automation is built around a couple of skills. And you have a skill that pulls down your data from Meta, then you have another skill that pulls down the data from Shopify, and then you have another skill that merges it all together and builds the dashboard. And inside the codex app you actually get an interactive dashboard, you know, with nice graphs and blah blah blah.
Ben Sufiani: I think this is an in-between step of software that simply didn't exist two years ago. We got these super apps now like Codex and Cloud Code and All we need to do quote unquote is essentially connect this harness with the brain in it to the different data sources. Give it the, I always say like give it the arms and the legs. Yeah. So it can do shit for us. Not only read data, but actually write data and, and campaign setups as well. And then you probably don't even need a database to build up a historic set of performance data. Right? You don't give a.
Paul from DreamyTales: Right. No.
Ben Sufiani: You do this moment by moment analysis, where are we at with the campaigns? What do we want to change based on that? And it keeps it so simple. Damn, I like it. I like it, I like it. I mean, there are opportunities to enhance that and I have some weird thoughts about that that you might be interested in. But, well. Are you happy with this setup right now and what is the benefit for you in your marketing practice of your physical and digital products?
Marcus: maybe so actually actually we are working like this since I would say six or seven months now. and what I mean our work has developed a lot since since the last twelve twelve months I would say. since Codex is available and the AI models are so strong as they are today. We are not really working anymore as operators. We are just the supervisors, supervising a a bunch of AI agents operating our systems. And this is not only for for marketing and and the web shop, but also for my backend, right? One year ago I was working on all those N at N workflows in the back end by myself. I was digging into that and changing notes and and all this stuff. And right now, if some some bug happens, I just give it to Codex and say, hey. Make an analysis for me and then you're providing a fix and do it yourself. Yeah? I'm just a supervisor supervising if the direction he steers is the right one. And that is kind of crazy to me that that this is possible right now because since since we're we are working like this I feel that we have become so much more powerful and efficient and doing so much more work in in less time. and I'm just talking about the back end. I think for the front end, for for all the marketing stuff, it's pretty much the same.
Paul from DreamyTales: Yeah, I one hundred percent agree. And also
Ben Sufiani: Yeah, can you go into that, Paul, like how you... you tell it.
Paul from DreamyTales: Also, I think the biggest change is even that we don't leave this super app anymore. Like you said, Ben. Like
Ben Sufiani: Mm-hmm.
Paul from DreamyTales: twelve months ago we were spending so much time in different UIs, right? You had to learn N8N and then you had to learn this tool that does this, and this tool generates a video, and blah blah blah. And now it's all condensed into this one app that I don't leave all day, right? I just tell these app. To do this and go into that system and now do this and blah blah blah. I think this is one of the biggest changes of the past twelve months. And speaking for our
Ben Sufiani: Amelie Soto
Paul from DreamyTales: basically our own day to day work, I think like Marcus said, it has become less of doing work, but more of building systems that do work.
Ben Sufiani: I try to tell people that like make whichever app you're using, like the codexes and the codes of the world. get really comfortable in it and essentially have a conversation per work stream that you're responsible for. I have a meta-ads work stream, have my A-B testing, my data work stream, I have my security work stream. It's just a conversation and it has a lot of context and when I have a problem in this and that area of responsibility, I can just ask like, can you fix that for me? Or can you also talk to the other conversation? This is what I also really appreciate now that I can literally say, hey, ask the meta ads conversation for input on blah blah blah, then set up the appropriate A-B test in post-hoc on the landing page. Insane. And when people hear this sometimes, think like, this is super advanced and super evolved. No, no, We're using one app.
Marcus: Yeah.
Ben Sufiani: When we talk about agents, we usually mean skills working on a schedule. like roughly, just simplifying it a little bit. And that allows you to give the agent over time with more trust, more more leash. So in terms of like campaign management, where do you currently give it the leash and where do you hold it back?
Paul from DreamyTales: So we in the beginning we developed a rule set that we would use ourselves to scale campaigns or to decrease campaigns, to kill campaigns, or to create campaigns. And this rule set is basically just a long document of me, you know. Talking and you know telling the AI turn this into a nice document. And right now this evolved into skills that where I would say create an create a campaign for me. These are the ads, and it it just goes and does it basically. So this is where I give it the leash. And where I would still Hold its hand and you know, kind of oversealed is with scaling or decreasing budgets on a day-to-day basic
Ben Sufiani: Mm-hmm.
Paul from DreamyTales: basis. Because there's so many rules that I feel like my rule set that I've tried to written down is still not as perfect. It still has, you know, it still has blank spots here and there, but you know, we're getting there. It's a it's a process. You keep iterating on it until you can, you know, let it fly.
Ben Sufiani: and I remember doing this kind of like budget scaling rules years ago. There have always been like this Russian scaling rules where there was... very clear like if ROAS over the last 48 hours is below this and this you need to lower the budget by 10 % per day. If it's above then go up and that creates like this corridor in which the campaign can either scale up or scale down and I even did like weather-based campaign scaling rules and shit like that literally years ago pre-AI but this feels so much more flexible and so much more grounded in the truth because I needed to rely on the meta-ads conversion data being correct, which it was not. So now I also feel much more confident with this kind of setup. Maybe one question before I want to move on to the creative production process. You keep saying codecs, but in the chat GPT desktop app, you have home and you have codecs essentially. In home you have chat and co-work and lots of non-technical people very much prefer to live in co-work. It uses your computer, it does everything, but you don't have to mess around with the code. You seem to be working in codecs a lot. Why?
Paul from DreamyTales: Well, we also have a bunch of repositories we run and that we push code to. So this was so first of all, that's one of the reasons. And second of all, we've been using codex before there was ChatGPT work. and we just got so used to it that it doesn't really make sense for us to now switch back to work because we've been doing everything that work can do now and more with codecs before.
Ben Sufiani: So let's have this explainer session for a second for everyone listening that is also non-technical. Why do we mess around with code bases and repos? What does it give us that, I don't know, a notion knowledge base that I access through on work would not give us? What would you guys say?
Marcus: it's for me it's trust issues because a lot of the times I'm using Codex to s work on my back end stuff. I'm really surprised of how elegant and smooth everything works. The the solutions Codex com comes up with are really, really good. And I figured that if I let do codex all the work, he's making less mistakes than I'm doing. So if I'm doing that myself,
Ben Sufiani: Mm-hmm.
Marcus: I have to debug much more. And Codex is doing that much much better. But sometimes there are these days or these tasks where he simply steers into the the wrong direction, completely wrong. Where I'm getting the feeling of why why are you so stupid right now? Why are you not doing the obvious thing? whereas I would of course choose the the white the white solution. And therefore,
Ben Sufiani: Yeah.
Marcus: therefore I like to keep supervising all the big steps Codex does. And for that I need the full control. I need the code base. I need I need to watch into the the different positions in the code, in the back end. so for me it's clearly those trust issues.
Ben Sufiani: Yeah, on the back end I get it. How would you pull?
Paul from DreamyTales: I mean for the front end it's just two completely different use cases I would say. So when you start out with AI and you have no idea what you're doing, you don't need a code base to keep track of stuff, right? But for example, if you want to build a website, the code that your website is made of has to live somewhere, right? And you have to maintain it somewhere. And that usually is a code base. Could also be, you know, some HTML file on a server somewhere. But you know, you know Ben how easy it is to set up stuff with Wesle, Next.js and GitHub. so I think it's just completely two different use cases. For example, one being a Notion document. we also have automations that write stuff just into Notion, you know, just append, append, append, automation after automation. To keep a lock of something, right? Where it's not it's not that big of a deal if if something breaks or if the if the content of that file is not one hundred percent true. But me working on the front end, you know, if I break the online shop and people can't check out, we lose money. So I need an infrastructure that makes sure that this doesn't happen and that's usually is a code base.
Ben Sufiani: that sometimes feels for me hard to explain to people who don't live in that world already, like justifying why to mess with Gitsch and to remember, shit, I need to commit and push this shit otherwise Marcus will not be able to see
Paul from DreamyTales: Mm.
Ben Sufiani: the changes. Christina and I, we do this now every day and it has become second nature and I'm still asking myself, What's the final interface going to be? Is it going to look more like co-work and just the code stuff is abstracted away because I literally couldn't care less. about it and I feel like I'm carrying around 50 years of IT baggage because of that but I also gain powers. Again, Marcos, when you said trust issues, bet you would have said this simply seeing what was deleted and what was added, especially when I work on my skills, yeah,
Paul from DreamyTales: Mm.
Ben Sufiani: and really important pieces of documentation like... I don't want you to rewrite a chapter of my book, like, no. No, no, no. I need to know. Like, I'm literally writing the Pirates Codex book in the code base. It's on pirateskids.com live. And I need to know. I have trust issues. And it just happens that code turns out to be this amazing tool, this amazing documentation and collaboration piece. It just feels a little bit overly complicated. I'm very curious where this is going. What do you guys think?
Paul from DreamyTales: Yeah. I
Ben Sufiani: Where's the final interface?
Paul from DreamyTales: That's that's
Marcus: I I
Paul from DreamyTales: a really interesting question. I completely agree because I remember when I first had to learn what a pull request is, I had such a hard time understanding what it is and what it does, man.
Ben Sufiani: Ahem.
Paul from DreamyTales: So I feel like I also feel like this is a technology of the past that now evolves into the AI era. And you know, stupid guys like us actually now get to mess around with a code base, right? But you would actually maybe you need something like Code but but but for dummies, right? So for people that
Ben Sufiani: you
Paul from DreamyTales: have that didn't exist pre-AI era or that didn't have to do anything with code before AI. I think we will see a lot of products that don't have a user interface at all. So you only have your agent and your agent talks to whatever product it is, but it doesn't even have a user interface anymore. So you don't get to see the actual product and click around. But your agent does. So it's just more or less an API and that is it. I think that's gonna be one of the things we we will start seeing more and more. as for code bases and man, what comes next? I am having a hard time on understanding even like the basics of it, so I'll I'll leave that to the to the pros. Yeah, it still works.
Ben Sufiani: But it's still working. It's still all working. That is kind of the magical part that you can abstract away the stuff that you don't understand yet. But if you take the time, if you sit down and ask the agent, hey, can you explain this to me? Walk me through this? Why is this necessary? Blah, blah. It's all good. And this whole PR madness, I had a time when I was like, no. And then the more I move into these little bit more factory software factory style working where and let's say my agent reads the Vercel log every hour season error. Do I do I just wanted to to merge it back into main is live?
Marcus: Has been
Ben Sufiani: Or or do you want to have like
Paul from DreamyTales: Okay.
Ben Sufiani: create the PR? I'll take a look friend. I'll take a look. I don't know how much I can take a look but at least That's a question that I still have. How much leash, how little leash do I want to give? And then when multiple agents work together, I think this workflow enables that very, very, very much. we don't need to see it anymore. Markus, anything on that before we move on to the creative production?
Marcus: Yeah, maybe just a couple of words, how I see things. I think we are right now in a transition phase, as as OpenAI and Anthropic are preparing with these tools like Codex work and and Cloud Co-Work, the step into the broader usage of their products, you know. So everyone in the future that the the end game the end game will be that everyone working with computers will be working with Codex or Claud. And they're not working on on their systems anymore because Codex has all the MCPs, all the connections to the systems. So just talking to an assistant, your your to your agent, doing stuff for you. And for most of the people in the real world all this code based stuff, they they they don't want to learn that. They cannot learn it or don't want to. And in the in the future I think they don't have to because the AI is getting so powerful of maintaining that all by them themselves. And if something breaks in the in the in the back end where the customers are not seeing that, they will fix it themselves. I think that's the end game. And right now we are the early adapters but we're just in the transition phase to this final stage
Ben Sufiani: It feels so early, doesn't it? Like it's so clanky still and we still, we already have superpowers. So one of the superpowers we now get as marketers, we get huge support in creative generation. I'm not talking about stupid AI slop, but actually images, videos, copy that converts. What is your content creation process look now?
Marcus: So maybe okay, Paul, it's our topic, you can proceed.
Paul from DreamyTales: So it so parts of it have changed with AI and parts of it have been the way they have always been. So in the beginning, you always have to get an idea or inspiration. What does the creative actually is is going to look like that I want to run, right? So you can get inspiration via AI, obviously. You can just ask ChatGPT or Claude. I want to sell this product, you know. What kind of creatives could I run? What's the angles? What's the hooks I could test? Blah blah blah. this can be one approach. The better approach that we do most of the time is take a look at your competitors and especially the meta ed library and try to scout creatives that have been running for a long time, you know, from brands that try to sell the same stuff that you are trying to sell. And then you try to get like three to four concepts out of this process before actually try to build the actual creative, right? So say you found this one ad from your one competitor in the meta-ad library and you know it has been running for three years or whatever, and you're like, man, this is printing. I have I have to build an ad that looks exactly like this. there's basically two cases. Like one case A it's a static, case B it's a video. If it's a static, you are very lucky because you can just pump it into GPT image two and tell it to, you know, rebrand it to your brand guide and put your product on it, whatever. Because this image model is just, you know, so good that this will work. from our experience, static images don't work as good as video ads. So What we are trying to do is to gr create video ads with AI.
Ben Sufiani: Is it your experience that videos are converting better than statics? It's
Paul from DreamyTales: Yes.
Ben Sufiani: not always mine, it's mixed project by project basis. What is it for you right now?
Marcus: Yeah, I think I think it really depends on the product. so for us, for our products, I think it's very very much important to explain a lot of things to the customers, how is it working? to get their own like personalized gift. and also it's a lot about emotions. It's a lot about connections between people and this is the feeling and the atmosphere you want to deliver to the customer. And this for us at least works better by by videos. I think we we never had a real success with images at least on a scale base.
Ben Sufiani: Interesting.
Marcus: Just a little bit, but yeah, video works best.
Paul from DreamyTales: Yeah, I think so for DreamyTales this is true. with Vibe TV we are seeing mixed results, but we're still in early testing phase there, so not sure where this will go. But yeah, for Dreamy Tails, we usually then try to scout like some kind of ad. And then we try to first of all try to digest this app. So we as humans take a look at this app because we At this ad because we think that AI is it's pretty good at describing the you know details of a video. It can tell you, alright, there's a lamp in the background, the girl is around this and that, age old, and you know, has blonde hair, whatever. But it's really bad at describing the aesthetics of it. So is why is this probably performing great, right? Is it because it
Ben Sufiani: Mm-hmm.
Paul from DreamyTales: looks Weird in a certain way? Is it because it's very long or very short or what is it? So we try to look at the ads and try to come up with hypotheses why this ad might work well, right? So it could be a certain hook or an angle or a lighting or the person or the music or the voiceover or whatever. And then we try to break this ad down into these different hypotheses. And from there we first of all try to generate prompts that would generate videos. So we have this engine build that we for example put in an idea for an ad and it turns our idea into a very detailed long prompt for a video scene, right? So that's the first step.
Ben Sufiani: Mm-hmm.
Paul from DreamyTales: And then you from that prompt you try to generate The actual video with the model of your choice. For us, you know, currently we are using Sora 2, but we are also switching these models. like underwear basically always depends on the use case and what's what's the state of the art model right now. And what you wanna do is you wanna you don't wanna create just one video, but you wanna create like on bulk, right? So one from one prompt you would then create, I don't know, 10 videos or whatever and choose the one that you think fits best. You know, because still it can be there can be weird stuff happening in the background or the lighting is too perfect or too imperfect or whatever. So then again you take a look at it and basically choose the one that you like best.
Ben Sufiani: Now, is it one video or is it sequence steps that you still need to put together?
Paul from DreamyTales: So it's sequences that you stitch together in the end. And
Ben Sufiani: Hmm
Paul from DreamyTales: you would it kind of depends on the angle and on the ad. There's also ads that are very short, you can do them in just one video. But most of the ads that we have are multiple scenes stitched together.
Ben Sufiani: And what do you use for editing?
Paul from DreamyTales: So we to put it all together we use a tool that's called Creatomate. ever heard of it?
Ben Sufiani: No. What is it?
Paul from DreamyTales: CreatorMate is a tool that's basically a video editor that allows you to edit a video just as you would in CapCut or whatever. But CreatorMate then lets you export this template of a video as JSON. And then you can replace every little piece of that video with another piece.
Ben Sufiani: Awesome.
Paul from DreamyTales: So say you have a template for a video and the voiceover starts exactly at you know second. Three, Creator Mate lets you switch out that voiceover just by you know feeding that JSON into into your AI agent of your choice.
Ben Sufiani: Again, with codecs at the center, managing all of these tool connections.
Paul from DreamyTales: So for this, we actually built a simple HTML file that speaks to different N8N workflows. So you have one you Yeah. So you have
Ben Sufiani: Yep. okay. Good idea.
Paul from DreamyTales: one input field in this HTML file that speaks to an N8N workflow that generates the prompt for the video. Then you have another input field that speaks to the N8N workflow that generates three videos from that prompt. Then you have another input field, you know, that makes headlines, voiceover, puts it all together. Just to save on contacts, to save on tokens and just because it's so repeatable, basically. But then you can also just hand this to Codex and tell and tell it, hey, you know, look at this.
Marcus: So this is one way of doing it. It works very, very great if you're having always the same template which which works for your ads. But if you wanna experiment on on cutting times on voiceovers and stuff like that, we've also made very, very good experience with yeah, with with codex skills. So they're different kind of skills. There's a lot of tutorials how to do that. So that codex handles all for you. They have the tool base, they are basically rendering the videos on your on your device. putting putting voiceovers to it, putting like simple captions and stuff like that. so you can basically with one prompt create a complete ad which is cut it and stitched together.
Ben Sufiani: I've been messing around with the re-motion and the hyperplex tools and accompanying skills. Is that what you're talking about Markus?
Marcus: Yeah,
Ben Sufiani: Which one of the two?
Marcus: To be honest, I have no idea. I I I told Codex took
Ben Sufiani: Yeah, it's okay. It's okay. It doesn't matter.
Marcus: Codex to make a research what is the best one, what is the state of the art, and he he he he proposed something and I said, Okay, go for it. don't let me hear about it, just do it.
Ben Sufiani: one of the reasons we are doing this open exchange is because we have nothing to sell. Like no Paul, Markus, you have nothing to sell here except for your wonderful B2C products, which here is not the primary audience, but because we're looking for others to share with us as well. Their tools, what they're using. Like I'm sure I'm going to be busy exploring Creatomate afterwards. for example. And maybe there's this or that idea that you explore and we need more people like us who have been AI-pilled but who have this performance-marketing mindset. Yeah, there are not so many, I feel. And I would just love to in touch, get in the comments, connect with Markus Paul, with me and talk about this stuff. And just to share a little bit about my process, I have... this all in Cloud Code in skills and references file. I also do this analysis of what makes it work, what makes an ad work. But I have everything documented by Cloud in docs files essentially from our design language. to the tone of voice that we're using, different formats of ads from single images to carousels to shorts to infographics and it's always a consistent feedback process. Every batch of creatives that you are creating is an opportunity for you to nag on what is not good enough yet. And if you look at pirateskills.com on the insides, yeah, where all of the stuff we are blogging, we are creating custom images for every blog post and animate them. And that is such a smooth process now. As Marcus said, when you have a repeatable template, it is very, very, very nice. And for exploration, for Paul's idea, something I was thinking about is... This is new, no not new, but a popular Go-Dev product called YouTube Download, YTDL. And I
Paul from DreamyTales: Mm.
Ben Sufiani: was like, we could let our agent research the best YouTube channels and just have a daily or weekly... routine that pulls all of the highest virality videos from these channels that we are observing or the niches that we are observing, then slicing it and seeing what was the hook, what was the general content, but as marketers we know how important the hook is and the format, the idea of the creative and then just pitch us that based on validated data. Do think that could work?
Paul from DreamyTales: I mean, in the end, it's all testing, right? So I think this might work, might as well not work. Surely depends on the product you're trying to sell with it. So one hook and one creative might work specifically for that one use case in that market for that audience. Might not work for another product, for another market, for another audience. But I think that it surely makes sense to build a system that tests this, right? So I think the real superpower isn't in lying the i isn't in finding the the secret ingredient, but is in building a system that can test so many things that you know you will inevitably find a winner just because you tested so much.
Ben Sufiani: Yes, and I still notice how I am limited in this idea. have this campaign idea for some product. I create the creators, I set up the campaign, I monetize and optimize them, blah, blah, end of line. Instead of thinking of it as a closed loop. Yeah, like have you ever come across this auto research repo from Kapathi? The idea is that you let an agentic system in a loop. Do the creatives first batch with you, set up the campaign, see the results, but then continue the process. Yeah. And create more creatives and go left and right. See what is working and as long as early stage indicators like higher click-through rate and in e-commerce, you have a very short cycle time to conversions ROAS. Yeah. Is positive, go more in that direction. Yeah. Explore to a certain degree and it just goes round and round and round. And this idea with a YouTube downloader or Instagram downloader or whatever is how can I, what Marcus said, pull myself a little bit out of that interactive loop and thinking about it as an assembly line that starts from A to B, but build this. Have you guys already tried that or is that something we should think about together after the pot? What do you think?
Paul from DreamyTales: Yeah, one hundred I think this is absolutely the right way to think about this process, right? As e-commerce guys, we are heavily dependent on paid ads and to automate as much as it as possible. and optimize is absolutely the right way to do. What we do is we do it in a different way. So we are not trying to optimize the creative, but we try to optimize basically our Our broad campaigns, right? So we run campaigns in I don't know five to ten countries depending on season. And the system that we build allows us to test angles in one market, and if they work, we can replicate exactly the same ad in a different language for another market, right? So on the second. I mean I'm not talking about like roughly the same, but I'm talking the exact same ad, just you know optimized for that other market. So in that kind case, I think this is some some variation of that loop that you are talking about, but a primitive one. I think there there's a way to go and especially like with testing the creative and optimizing the creative, that is definitely a a road to to explore.
Ben Sufiani: So maybe to wrap things up slowly, Markus, from you I got the impression that your views at the moment, lots of people can now build stuff. Very few can also distribute and market that. Can you expand on that a little bit?
Marcus: yeah, I think I think that pretty much comes from the fact that right now you can build an eighty percent product pretty, pretty quickly. So if you're having a an idea, you can with the help of an AI j agent, doing doing eighty percent of the work and you're having something solid, you're having a prototype or something. But the the the rest of the one hundred percent, the twenty percent, this is the tough nut crack, you know. and this is where where also marketing comes in. into place. I think y y you're still rely relying on on this those this human input, this human feeling, the atmosphere, the the knowledge, the the experience you have as a marketer. and I hope this stays like this. I mean I'm not sure but it can be that in twelve or twenty-four months this twenty percent has been shrinked to like five percent or even lower
Ben Sufiani: Mm-hmm.
Marcus: and then Everyone can do what we we are doing, yeah. but yeah, I think nobody knows what's what's going to happen that in that terms.
Ben Sufiani: But it's worth staying at the edge of what's possible.
Marcus: Absolutely. If you're not if you're not staying at the edge somebody else's and then yeah, you're losing.
Ben Sufiani: such a new opportunity space and pre-AI, probably like since I'm 16, I have this image of my mind of like the Neanderthals. Like, I was always fascinated how slow progress was. Why the fuck did it take 30,000 years to go from a round stone to a sharp stone? And then 15,000 years to put it on a stick like for the first spear. And for me fundamentally, the image in my mind was... This is one monkey sits down on his ass and is thinking about round stone. What can I do with round stone? Yeah. And it's up to the monkey that sits down on his ass and thinks, how can we improve our quality of life? How can we now build a little piece of that Iron Man suit that we're able to build for ourselves? And we got a whole new toy put in our lab and it's down to us and it's not just the developers anymore that get to play but everyone. Yeah that is very fascinating for me. I hope we have inspired a couple of people to join us, think about... maybe regularly getting in touch with Marcus, Paul and myself to build a little mastermind, nothing paid, where we really want the practitioners of performance marketing combined with AI in what is left and right for that. Analytics, creative, the whole landing page, customer conversion experience. would be very very welcome to hear. Are there any last words or asks that you have for the audience before we sign off?
Paul from DreamyTales: thanks again just for having us, Ben. It's been a lot of fun and you know, our eyes widen when we talk about this stuff because we play with it every day, just like you said. So I think it's a great initiative to try to get more people together to come together and play and see what we can build.
Ben Sufiani: Not cuz you.
Marcus: Yeah, also from my side. Thank you, thank you so much for having us. it was very, very fun. and hope hope to be one of your guests for for another time.
Ben Sufiani: For sure, for sure. And we see each other on September 2nd at the Vibe Coding Cologne event where you actually open a laptop, show stuff, as well. So I'm very much looking forward to that. And I would say have a good day. Bye.
Paul from DreamyTales: Have good day.
Marcus: Thanks.
Paul from DreamyTales: Bye Ben.
Paul from DreamyTales: Mm-hmm.
Ben Sufiani: and my last 30 day costs is $4,540 on the 90
Paul from DreamyTales: It's insane, right?
Marcus: What's that? Yeah.
Ben Sufiani: euro subscription. This little thing makes you appreciate the subsidization that these crazy companies do. appreciate even more. So how about you guys give yourself a little intro, tell us what you guys do and how you're separating between yourself, your business.
Marcus: Yeah, first of all, thanks for having us. we are very happy to be here today and to introduce us and our company and our business. maybe a couple of a couple of words to myself. I'm Marcus, I'm thirty-six years old, I am living in Cologne here with my wife in the southern city. And actually I am a mechanical engineer and after my studies I founded my first company in twenty nineteen. And then twenty twenty three I had a lot of luck to be able to sell that company to an investor. In parallel, I was meeting Paul for the first time and we started with our company DreamyTales with our startup. DreamyTales is selling personalized gifts for everyone. And before we continue with that, maybe Paul you can introduce yourself a little bit.
Ben Sufiani: Yep.
Paul from DreamyTales: Yeah,
Ben Sufiani: Thanks mate.
Paul from DreamyTales: hi everyone, my name is Paul. I am the other half of DreamyTales. I am 31 years old, and my professional background is kind of yeah, I'm I'm sort of a jack of all trades. I used to be a event manager and then I studied business psychology. and yeah, during that time I got interested in startups and the startup scene in general. And then had the great luck of meeting Marcus in 2023 and built our first business together. DreamyTales, as Marcus said, is a company that sells personalized gifts. So if you Ben need a gift for your fiance, say for your anniversary, you could turn to DreamyTales to turn your guys' love story into a book or personalized video or personalized song. that's basically our best selling products. And yeah, I'm also based in Cologne here and we've been doing this for the past three years now.
Ben Sufiani: Very cool. So I always found it fascinating how you guys are switching between physical and digital. I remember when we had one of our first sessions together, I was like, what the fuck is this not digital? Who comes up with the idea of doing AI generated images in a story and then actually prints it and ship it out to people? And at that time you were, think, still working a lot like with external developers who were not vibe coding these things yourself. Right? So I'm curious, like, how are you thinking about doing like physical and digital using external developers versus vibe coding it yourself? What are you thinking about that?
Marcus: So yeah, maybe maybe first things first. Actually we sh we started with a physical product. in the very beginning with Dreamy Tails we were selling children's books, personalized children's books with AI. And then we figured out pretty much very very early that this needs to be a physical product because no parent wants to hold a tablet or a smartphone into the face of their child's and it had to be something something printed, you know. So we started this this journey with this physical book, with this personalized book. We then later switched the the audience, the customer audience group to adults. and then Later on we tried to find new new kinds of products we can do with AI and we came up with AI songs and AI videos. And while doing that we figured out that it's much, much easier to scale those digital products into different kinds of markets, for instance.
Ben Sufiani: Mm-hmm.
Marcus: Right now we are selling in
Ben Sufiani: Yeah, yeah, yeah.
Marcus: more than ten countries worldwide. So for instance we are selling in Brazil, in Mexico, in Turkey, and and all those countries. and we started selling in Brazil, for instance, with the physical books and we had to find some local partners who are printing our books and distributing them. And that's very, very hard work, you know. And then
Ben Sufiani: Yeah, crazy. Yeah.
Marcus: in parallel we tried our AI songs there and figured out that it's much, much easier to scale those because you'd simply can can send this song by by a download link via email or whatsapp and that's it. And that's a great great trade-off we have with these digital products
Paul from DreamyTales: So maybe just to add here, so when back in the day when we started, these digital products weren't even thinkable of, right? So the first image models, they were so crappy that we could barely generate images to print in our books. But not to mention,
Ben Sufiani: Mm-hmm.
Paul from DreamyTales: you know, generating whole videos or you know even songs. This was kind of developing over time and we always just kept a close eye on the market and our competitors were these guys were selling. And then over time it became technically possible to build these products and that's when we tried and then discovered that it's actually way easier to scale those instead of building physical products. But I do think that physical products also have like like a bunch of pros, right? So they have a bigger
Ben Sufiani: Yep.
Paul from DreamyTales: mode around them, for example. So we see that with the Vibe TV as well So the digital products, I mean, if Marcus and I we we are able to build those with basically just a laptop, it's pretty fair to assume that other guys with a laptop could rebuild the exact same product, right? So not a real mode around this. But with a physical product, this g this gets a lot more complicated, right? You need a supply chain, logistics, handle returns, shipping, blah blah blah. So that's also some sort of a a mode you build with physical products to protect yourself
Ben Sufiani: I understand.
Paul from DreamyTales: from just random guys with a laptop somewhere.
Ben Sufiani: And I hope you guys recognize that this makes you unique. You are not scared of physical. You started there, you now understand digital and I think this is a, that's probably your biggest edge. that you actually can play both and interact between them and come up with product ideas that... normal just physical or just digital founders wouldn't come up with. so I'm very curious to see what you're building on top. But let's go back to the to the songs. I remember maybe it was maybe a year ago you were you guys were at the meetup and showing the the end it end workflow. and I just I'm sure like this was the initial version of it, but Markus since you're building them, like why end it end? Why not build like the full backend in the beginning? How did you get decent results? Make it reliable?
Marcus: that's a very good question. so first of all maybe to explain My job at DreamyTails is basically the back end. so I'm responsible for everything, what happens, what the customer cannot see directly. and Paul is
Ben Sufiani: Mm.
Marcus: responsible for the front end. So front end online shop, also communications and and marketing. And you said that I I am using N at N, that's right. as I'm just a mechanical engineer, I'm not really a software developer. So to be honest, I n I have no clue how to how to manage and program a complete back end for our shop. So from the very first beginning,
Ben Sufiani: Mm-hmm,
Marcus: we were using no code tools or low code tools. And with the help of ChatGPT, we basically built everything from scratch with these low code tools. And N at N is my best friend right now because it's a good it's a great platform. You have great overviews, you have great debugging. And right now I I would call myself a Vibe backend coder. so that works works quite great.
Ben Sufiani: That's so fun. This is a thing. Okay, questions. Where do you host your NADM and where do you source the audio and video models from, from which API?
Marcus: yeah, that's different. So Yeah, sure.
Ben Sufiani: If you want to talk about it. You don't have to, but I'm curious.
Marcus: so in the very beginning we were not even using N at N. We were using make dot com. It's a quite similar platform but much, much easier, much more basic. And we were using the the hosted programs. Same with N at N until to a certain scaling point where we said, Okay, it's much easier and much much more scalable. hosting our own server, our own virtual machine and deploying the NN instance on there. And that's currently the case right now. for models and and products we are using, this is like like a complete bunch of of of AI models. We are using some open source models on our own servers, but also like the usual ones like Chat GPT, open AI image generation. we are using Sora too for video ads and stuff like that.
Ben Sufiani: You go straight to the provider to use something like OpenRouter in between.
Marcus: it depends. So I think most most of the time we're using FEL AI, which which is quite common. It's a very robust
Ben Sufiani: Same, yep.
Marcus: API for all kinds of all kinds of AI solutions.
Ben Sufiani: And I remember that there was this guy at the meetup who showed us Feli AI. While I didn't know about it until then and he was like building landing pages and I've done so obviously also like for a while, but then he was also doing all of the imagery in the layouts of the landing page. And I was like, God damn it Why didn't you understand like that? This is the obvious fucking thing to do. And then I looked at my website and everything was free of images. And I was like, okay,
Paul from DreamyTales: Mm-hmm.
Ben Sufiani: this looks decently designed, but it doesn't feel as alive. So interesting enough, I used to use a lot of NNN and since I'm very much in the Next.js and the Versel ecosystem. I switch to their workflow SDK, an open source version of like NNN, but it just runs as part of the code base and not like as a third party thing. And then I'm sourcing stuff from, from Thel or whichever platform I get that stuff from. That was an interesting change to me and listen to what three non-developer guys are talking about. What the hell is going on here?
Marcus: Yeah.
Paul from DreamyTales: Exactly. Exactly. I remember
Ben Sufiani: The main discussion is going to be on the marketing madness and we'll get into that, go Paul.
Paul from DreamyTales: I remember when we met and it used to be all like, all right, show me your ad manager and you know,
Ben Sufiani: Yep.
Paul from DreamyTales: what's your CTR and stuff like that. And now we're talking about you know, backend SDKs and Next.js and other frameworks. I mean it's insane to me how much you can learn. Like Marcus and I, we are not developers. You are not a developer then,
Ben Sufiani: Yeah.
Paul from DreamyTales: but we are all like marketing guys, right? That now dig into building the actual products we are trying to sell, which is a great way to do things. I think it's it's it's a great skill to have if you come from e-commerce or in general digital marketing and you know how to sell stuff. Now you can learn how to actually build stuff, the product underneath it and iterate on that. I think it's a very, very good starting point. instead of you know having to know how to build stuff and then now having to learn how to actually sell the stuff you're trying to build.
Ben Sufiani: Let see if I can pull this up and hold it to the camera. I got this idea of the sovereign entrepreneur in my mind. And I always felt very dependent on developers to build the product that I was finally able to sell. Here it is. I'm not sure if this is readable. Maybe I'm hiding behind it a little bit. So like you need a founder who builds something, a marketer who sells it and an organizer who administrates all of this stuff. And you get, you get in trouble if you're lacking one of these skills. Like we marketers tend to be like, okay, we, we're founders, we organize shit and then we marketing a crappy product. That is a typical marketer trap. Or we have the opposite. We have a developer that builds a great product but doesn't have enough customers. They have great stuff but they didn't sell enough. And then we have the business types who are often just fucked because they can't build and sell. And now we are developing ourselves into these three functions. And I think that's also the main thing why my podcast here is called Build, Grow, Repeat. It's like build a product, grow it, make it repeatable in a sense that you organize and orchestrate everything in this loop. And you are very very very close to that including the physical world. Well, I'm more tied to the to digital world. But what we do have in common is that we love paid advertising. We love sending money Mark Zuckerberg's away, which is totally
Marcus: Okay.
Ben Sufiani: weird. And but we do it a lot. I don't know how many hundreds of thousands I have spent his way. But let's just, for people who are not that deep in the world, why the fuck would we even consider paid advertising as a valid approach? What makes it interesting for you? Why don't you focus more on organic? What justifies the spend?
Paul from DreamyTales: Great question. So first of all, obviously the obvious question is why would you spend money on meta, right? so
Ben Sufiani: Ahem.
Paul from DreamyTales: if if you can just post organically, we tried this for the past three years and we tried really bad. So you know, there were times where we had in-house teams and all they would do is trying to produce organic content and pump it out on our Instagram channel, trying to reach people. but no matter what we did, it didn't work. So we might have generated reach, yes, and likes and you know what have you, but not to a scale where you would say, All right, this is a you know reliable channel to build a business on. for us this has always been pretty unpredictable. Also with the algorithm changing in the background, like especially on Instagram. you know, interaction such as likes and follows, blah blah blah, get less important, but you know, swiping and hold rates and etc. get more important. all this was an environment where it felt hard for us to navigate in and to actually generate reach. So we tested this, but also simultaneously we obviously ta tested paid advertising, and this worked from the start for us. So This was a channel where we generated insights that we could actually work with. So, you know, you spend a euro and you get insights back on whether is your creative performing or not and why it might or might not be performing. So we generated more data as to why our performance was good or bad. And this gave us a base to iterate on this. And in the end, you know, after three years of doing this, all we can say is that we did what worked for us and that was paid advertising. We tried organic, didn't work for us. But you know, why should we waste time on doing more of the stuff that doesn't work when we can just spend time on doing the stuff that does work, which is paid advertising.
Ben Sufiani: Any opinion on that Marcus?
Marcus: yeah, I mean paid advertising for me, I was starting with that in twenty nineteen and it was back then it it was much easier, I would say, to have a good ROARS on on those platforms like Meta and stuff. and for me it was always this magic little button where w you are like testing an ad and then maybe iterating on that a couple of times. But when you find the the killer creative, the killer ad which performs You just have this little button, you have to press it and spend more money every day. And then it's like a like a printing money machine, right? So you're spending money, but on the other side you're getting money back, you're getting revenue. And this certainly works to a to a certain level. it's not like there's there's a certain move there. But for me it was always this this magic button which I can press. And it worked very, very great for me in the past.
Paul from DreamyTales: Also, like for DreamyTales, DreamyTales is a very seasonal business. So for us, it's very important that our high seasons, such as Christmas and Valentine's Day, that these work, right? So we need a scalable and predictable source of revenue for these high seasons. And organic is just the opposite of this, right? So it might work, but it also might not work. And like Markus said, with paid advertising, you can just scale your campaigns when you know the demand is going to be high. You get more metrics to look at. So it's way more predictable and scalable for us.
Ben Sufiani: I think so too. Like, I'm judging myself a little bit for not truly cracking the organic nut. I
Paul from DreamyTales: Yeah.
Ben Sufiani: have this stupid saying in my brain all the time, traffic is like sex, only losers pay for it. Do
Marcus: Yeah.
Ben Sufiani: you know this one?
Marcus: Well yeah, then then I think we are the biggest losers.
Ben Sufiani: Yeah, we are. to. I absolutely envy people like Daniel Minini, who's so correct, the organic nut. But his target is not a product. His target is ads that he is selling. We are the buyer of his ads, or of his placements. And it's not easy to do organic at all. The only alternative I'm very much willing to test out when I'm on my next B2C rampage, which could be Bridesmaid, the agentic wedding planner, is treating Working with influencers, small to medium sized, on a very standardized contract basis and treating that like performance marketing. This is the reach we get, this is the engagement, this is the clicks we get, this is the coupon code, entries we get, this is ROI. Yes, you can work with me again, if not change or you're going out of the rotation. To me, it's a bit more human hassle than I like. One of the things I love about Meta are Google or TikTok ads. is the, it's just numbers. It's just you click in a dashboard, you click the magic button that Marcus has and it starts cranking. The risk is that you are wasting your money. and this is why I, for example, became such an analytics nerd. It was my way to find out, hey, if I'm losing half of my marketing budget, I would like to know which half. And you have been using triple whale for that for a while and thought about like maybe building that into your own thing. But you have become very focused on what we marketers call ROAS, return of advertising spend. You spent a hundred euro. you want to get out X factor more, let's say three. that makes means gross of 300 % is one euro in three euros out just for the future discussion here that we're going to have. So how was your thinking about like a triple whale expensive? Do I build? Do I buy? What was that process like?
Paul from DreamyTales: I mean you're absolutely right. So when you scale a business, it obviously becomes a little asset heavy over time. So you buy software or you buy I don't know software solutions, you know, monthly subscriptions that just keep on running. And this piled up over time, right? And Triple Way was one of these. So as a classic D2C brand, Triple Way is just considered to be one of the Basic tools to have, right? So it's always the Shopify Triple Way meta stack that you build a D2C brand on, basically.
Ben Sufiani: Yep.
Paul from DreamyTales: So that's where we were all also running with this. And I mean for most of the time this was this was fine, right? I think the base subscription of Triple Way is $300 a month. and triple way basically gives you cross-channel attribution. and the triple pixel on your website, right? So it tells you like a customer has seen your ad on Instagram and then he googled your brand and then he bought via a third channel or whatever, right? this plus a bunch of metrics for profit and you know fixed costs, your cocks, whatever you have. So the analytics dashboard. And yeah over time we always asked ourselves all right can we maybe get rid of A lot of these really expensive software solutions that our business is depending on, right? One of them being triple whale, a bunch of others also being, for example, AI models that we pay per API request, but now we host them on our own virtual machines, right? So we switch to open source alternatives there. And with with AI,
Ben Sufiani: Awesome.
Paul from DreamyTales: this this was made possible, right? So in the performance marketing world, we kicked out triple whale and instead built a blended metric, a blended ROAS metric from our Shopify sales and the meta performance of our campaigns, so to say.
Ben Sufiani: we did dive deep into that for sure. Did you need to do a lot of for that in the backend markers or could you just, did you need to make changes to like Shopify layouts or how do I even think about doing something like that in a Shopify technology stack? Next year's I get it, I don't get it on Shopify.
Marcus: I think no, actually I I did not have to change much in the back end. I think the data is pretty much there already. The only thing triple rays does for you is to collect all this data which is hidden somewhere, puts it all together in a nice dashboard and with a triple pixel showing the the exact same the exact raws you you have on your on your campaigns, right? since we
Ben Sufiani: Yep.
Marcus: are not using Triple Rail anymore, obviously we are not having one hundred percent of the vision anymore. But with the helps of AI and codecs and skills, I think Paul is able to collect and gather as much data as necessary to have at least a clear picture of what is happening. furthermore I think triple well is a very very, very good tool, especially when you're learning, when you're when you're in the experiment experiment phase and learning phase, when you're trying new ads, new angles, new products, all the stuff, then you have to have a very, very clear picture and very good data what is what is working right and what is working not right, And and iterate iterate on those data points. right now with DreamyTales, We we've done a lot of testing, we've done a lot of experiments, and for most of our products we are having already a clear pretty clear picture what channels are great, what ch channels are not working great. And therefore we had the confidence to switch back from from triple whale to our own reporting system within Codecs.
Ben Sufiani: Tell us, what have you built, how have you built it? If you wanna show something, show it, but you can keep that for the VibeCodingCologne event if you like.
Paul from DreamyTales: so maybe to start off, it's important to understand how a Shopify website works, right? So as you said, Ben, you get it on a Next.js sort of project. Shopify works a little different. So you have the Shopify store and you install themes on this Shopify store, right? And a theme runs in Shopify's own template language that's called Liquid. And Liquid comes with a bunch of perks, right? So Liquid doesn't allow you to just you know install whatever SDK on it or in it and place whatever random pixels in it. It works, but it's a little clanky and it's not as straightforward as say you're running an X.js project. So that's why there's Shopify apps. So when you install some kind of Shopify app on your store, this app handles it all for you, right? It installs the necessary scripts in your code base, handles all the tracking, all the pixels, blah blah blah. So when you kick out triple well, you're getting rid of all of this. And now all you're left with is the meta source, right? So what we did was first of all was trying to Pull the data from our ads manager to our local machine, so to my machine, so that I could basically work with that data, right? And we did this for all campaigns via the meta ads MCP. You can do this, but you can also do this via the GraphQL API from Meta. So that's pretty easy. And then The problem is that Meta only shows you, you know, 60, 70% of the truth, right? So meta doesn't track all the sales of a campaign or misattributes them, blah blah blah. That's why solutions.
Ben Sufiani: I do that a lot, yeah.
Paul from DreamyTales: That's why solutions like triple well exist in the first place, right?
Ben Sufiani: Yeah.
Paul from DreamyTales: so you want to basically cancel out these. So you need a second data source, and that being Shopify. So for us, for example. say we have a campaign that sells a product in Mexico, right? we compare the data that Meta gives us from this campaign, and then Meta, for example, tells us we have a ROWS of 3.5 or whatever, and then we try to calculate the ROS based on the actual Shopify sales that we had in the same time frame in that country for that product, right? Because this is your actual truth, right? So you can pull down how many units of a product of this that this campaign is trying to sell in that country down from Shopify and divide the ad spend by those products to get get the rowers. And by doing so, yes.
Ben Sufiani: question. this is all like all of the Shopify data you get through your custom-built Shopify app. Is that correct? Okay,
Paul from DreamyTales: Exactly. Yeah,
Ben Sufiani: and you don't have to share it on the App Store, you can just have one privately for your own store. What
Paul from DreamyTales: Exactly.
Ben Sufiani: did you call yours?
Paul from DreamyTales: I think it's called the Codex Shopify connector or something. Very creative.
Marcus: Mm-hmm.
Ben Sufiani: Okay, yeah, cool, cool, cool, cool, cool. Yeah, I get it. So you pull these different sources and you're pulling them together in a system and you need to build a dashboard for that. need to save that data. What did you use for that?
Paul from DreamyTales: we do this in Codex. So we run the Codex desktop app that allows you to create automations. And these automations run every whatever, in this case morning. And this automation is built around a couple of skills. And you have a skill that pulls down your data from Meta, then you have another skill that pulls down the data from Shopify, and then you have another skill that merges it all together and builds the dashboard. And inside the codex app you actually get an interactive dashboard, you know, with nice graphs and blah blah blah.
Ben Sufiani: I think this is an in-between step of software that simply didn't exist two years ago. We got these super apps now like Codex and Cloud Code and All we need to do quote unquote is essentially connect this harness with the brain in it to the different data sources. Give it the, I always say like give it the arms and the legs. Yeah. So it can do shit for us. Not only read data, but actually write data and, and campaign setups as well. And then you probably don't even need a database to build up a historic set of performance data. Right? You don't give a.
Paul from DreamyTales: Right. No.
Ben Sufiani: You do this moment by moment analysis, where are we at with the campaigns? What do we want to change based on that? And it keeps it so simple. Damn, I like it. I like it, I like it. I mean, there are opportunities to enhance that and I have some weird thoughts about that that you might be interested in. But, well. Are you happy with this setup right now and what is the benefit for you in your marketing practice of your physical and digital products?
Marcus: maybe so actually actually we are working like this since I would say six or seven months now. and what I mean our work has developed a lot since since the last twelve twelve months I would say. since Codex is available and the AI models are so strong as they are today. We are not really working anymore as operators. We are just the supervisors, supervising a a bunch of AI agents operating our systems. And this is not only for for marketing and and the web shop, but also for my backend, right? One year ago I was working on all those N at N workflows in the back end by myself. I was digging into that and changing notes and and all this stuff. And right now, if some some bug happens, I just give it to Codex and say, hey. Make an analysis for me and then you're providing a fix and do it yourself. Yeah? I'm just a supervisor supervising if the direction he steers is the right one. And that is kind of crazy to me that that this is possible right now because since since we're we are working like this I feel that we have become so much more powerful and efficient and doing so much more work in in less time. and I'm just talking about the back end. I think for the front end, for for all the marketing stuff, it's pretty much the same.
Paul from DreamyTales: Yeah, I one hundred percent agree. And also
Ben Sufiani: Yeah, can you go into that, Paul, like how you... you tell it.
Paul from DreamyTales: Also, I think the biggest change is even that we don't leave this super app anymore. Like you said, Ben. Like
Ben Sufiani: Mm-hmm.
Paul from DreamyTales: twelve months ago we were spending so much time in different UIs, right? You had to learn N8N and then you had to learn this tool that does this, and this tool generates a video, and blah blah blah. And now it's all condensed into this one app that I don't leave all day, right? I just tell these app. To do this and go into that system and now do this and blah blah blah. I think this is one of the biggest changes of the past twelve months. And speaking for our
Ben Sufiani: Amelie Soto
Paul from DreamyTales: basically our own day to day work, I think like Marcus said, it has become less of doing work, but more of building systems that do work.
Ben Sufiani: I try to tell people that like make whichever app you're using, like the codexes and the codes of the world. get really comfortable in it and essentially have a conversation per work stream that you're responsible for. I have a meta-ads work stream, have my A-B testing, my data work stream, I have my security work stream. It's just a conversation and it has a lot of context and when I have a problem in this and that area of responsibility, I can just ask like, can you fix that for me? Or can you also talk to the other conversation? This is what I also really appreciate now that I can literally say, hey, ask the meta ads conversation for input on blah blah blah, then set up the appropriate A-B test in post-hoc on the landing page. Insane. And when people hear this sometimes, think like, this is super advanced and super evolved. No, no, We're using one app.
Marcus: Yeah.
Ben Sufiani: When we talk about agents, we usually mean skills working on a schedule. like roughly, just simplifying it a little bit. And that allows you to give the agent over time with more trust, more more leash. So in terms of like campaign management, where do you currently give it the leash and where do you hold it back?
Paul from DreamyTales: So we in the beginning we developed a rule set that we would use ourselves to scale campaigns or to decrease campaigns, to kill campaigns, or to create campaigns. And this rule set is basically just a long document of me, you know. Talking and you know telling the AI turn this into a nice document. And right now this evolved into skills that where I would say create an create a campaign for me. These are the ads, and it it just goes and does it basically. So this is where I give it the leash. And where I would still Hold its hand and you know, kind of oversealed is with scaling or decreasing budgets on a day-to-day basic
Ben Sufiani: Mm-hmm.
Paul from DreamyTales: basis. Because there's so many rules that I feel like my rule set that I've tried to written down is still not as perfect. It still has, you know, it still has blank spots here and there, but you know, we're getting there. It's a it's a process. You keep iterating on it until you can, you know, let it fly.
Ben Sufiani: and I remember doing this kind of like budget scaling rules years ago. There have always been like this Russian scaling rules where there was... very clear like if ROAS over the last 48 hours is below this and this you need to lower the budget by 10 % per day. If it's above then go up and that creates like this corridor in which the campaign can either scale up or scale down and I even did like weather-based campaign scaling rules and shit like that literally years ago pre-AI but this feels so much more flexible and so much more grounded in the truth because I needed to rely on the meta-ads conversion data being correct, which it was not. So now I also feel much more confident with this kind of setup. Maybe one question before I want to move on to the creative production process. You keep saying codecs, but in the chat GPT desktop app, you have home and you have codecs essentially. In home you have chat and co-work and lots of non-technical people very much prefer to live in co-work. It uses your computer, it does everything, but you don't have to mess around with the code. You seem to be working in codecs a lot. Why?
Paul from DreamyTales: Well, we also have a bunch of repositories we run and that we push code to. So this was so first of all, that's one of the reasons. And second of all, we've been using codex before there was ChatGPT work. and we just got so used to it that it doesn't really make sense for us to now switch back to work because we've been doing everything that work can do now and more with codecs before.
Ben Sufiani: So let's have this explainer session for a second for everyone listening that is also non-technical. Why do we mess around with code bases and repos? What does it give us that, I don't know, a notion knowledge base that I access through on work would not give us? What would you guys say?
Marcus: it's for me it's trust issues because a lot of the times I'm using Codex to s work on my back end stuff. I'm really surprised of how elegant and smooth everything works. The the solutions Codex com comes up with are really, really good. And I figured that if I let do codex all the work, he's making less mistakes than I'm doing. So if I'm doing that myself,
Ben Sufiani: Mm-hmm.
Marcus: I have to debug much more. And Codex is doing that much much better. But sometimes there are these days or these tasks where he simply steers into the the wrong direction, completely wrong. Where I'm getting the feeling of why why are you so stupid right now? Why are you not doing the obvious thing? whereas I would of course choose the the white the white solution. And therefore,
Ben Sufiani: Yeah.
Marcus: therefore I like to keep supervising all the big steps Codex does. And for that I need the full control. I need the code base. I need I need to watch into the the different positions in the code, in the back end. so for me it's clearly those trust issues.
Ben Sufiani: Yeah, on the back end I get it. How would you pull?
Paul from DreamyTales: I mean for the front end it's just two completely different use cases I would say. So when you start out with AI and you have no idea what you're doing, you don't need a code base to keep track of stuff, right? But for example, if you want to build a website, the code that your website is made of has to live somewhere, right? And you have to maintain it somewhere. And that usually is a code base. Could also be, you know, some HTML file on a server somewhere. But you know, you know Ben how easy it is to set up stuff with Wesle, Next.js and GitHub. so I think it's just completely two different use cases. For example, one being a Notion document. we also have automations that write stuff just into Notion, you know, just append, append, append, automation after automation. To keep a lock of something, right? Where it's not it's not that big of a deal if if something breaks or if the if the content of that file is not one hundred percent true. But me working on the front end, you know, if I break the online shop and people can't check out, we lose money. So I need an infrastructure that makes sure that this doesn't happen and that's usually is a code base.
Ben Sufiani: that sometimes feels for me hard to explain to people who don't live in that world already, like justifying why to mess with Gitsch and to remember, shit, I need to commit and push this shit otherwise Marcus will not be able to see
Paul from DreamyTales: Mm.
Ben Sufiani: the changes. Christina and I, we do this now every day and it has become second nature and I'm still asking myself, What's the final interface going to be? Is it going to look more like co-work and just the code stuff is abstracted away because I literally couldn't care less. about it and I feel like I'm carrying around 50 years of IT baggage because of that but I also gain powers. Again, Marcos, when you said trust issues, bet you would have said this simply seeing what was deleted and what was added, especially when I work on my skills, yeah,
Paul from DreamyTales: Mm.
Ben Sufiani: and really important pieces of documentation like... I don't want you to rewrite a chapter of my book, like, no. No, no, no. I need to know. Like, I'm literally writing the Pirates Codex book in the code base. It's on pirateskids.com live. And I need to know. I have trust issues. And it just happens that code turns out to be this amazing tool, this amazing documentation and collaboration piece. It just feels a little bit overly complicated. I'm very curious where this is going. What do you guys think?
Paul from DreamyTales: Yeah. I
Ben Sufiani: Where's the final interface?
Paul from DreamyTales: That's that's
Marcus: I I
Paul from DreamyTales: a really interesting question. I completely agree because I remember when I first had to learn what a pull request is, I had such a hard time understanding what it is and what it does, man.
Ben Sufiani: Ahem.
Paul from DreamyTales: So I feel like I also feel like this is a technology of the past that now evolves into the AI era. And you know, stupid guys like us actually now get to mess around with a code base, right? But you would actually maybe you need something like Code but but but for dummies, right? So for people that
Ben Sufiani: you
Paul from DreamyTales: have that didn't exist pre-AI era or that didn't have to do anything with code before AI. I think we will see a lot of products that don't have a user interface at all. So you only have your agent and your agent talks to whatever product it is, but it doesn't even have a user interface anymore. So you don't get to see the actual product and click around. But your agent does. So it's just more or less an API and that is it. I think that's gonna be one of the things we we will start seeing more and more. as for code bases and man, what comes next? I am having a hard time on understanding even like the basics of it, so I'll I'll leave that to the to the pros. Yeah, it still works.
Ben Sufiani: But it's still working. It's still all working. That is kind of the magical part that you can abstract away the stuff that you don't understand yet. But if you take the time, if you sit down and ask the agent, hey, can you explain this to me? Walk me through this? Why is this necessary? Blah, blah. It's all good. And this whole PR madness, I had a time when I was like, no. And then the more I move into these little bit more factory software factory style working where and let's say my agent reads the Vercel log every hour season error. Do I do I just wanted to to merge it back into main is live?
Marcus: Has been
Ben Sufiani: Or or do you want to have like
Paul from DreamyTales: Okay.
Ben Sufiani: create the PR? I'll take a look friend. I'll take a look. I don't know how much I can take a look but at least That's a question that I still have. How much leash, how little leash do I want to give? And then when multiple agents work together, I think this workflow enables that very, very, very much. we don't need to see it anymore. Markus, anything on that before we move on to the creative production?
Marcus: Yeah, maybe just a couple of words, how I see things. I think we are right now in a transition phase, as as OpenAI and Anthropic are preparing with these tools like Codex work and and Cloud Co-Work, the step into the broader usage of their products, you know. So everyone in the future that the the end game the end game will be that everyone working with computers will be working with Codex or Claud. And they're not working on on their systems anymore because Codex has all the MCPs, all the connections to the systems. So just talking to an assistant, your your to your agent, doing stuff for you. And for most of the people in the real world all this code based stuff, they they they don't want to learn that. They cannot learn it or don't want to. And in the in the future I think they don't have to because the AI is getting so powerful of maintaining that all by them themselves. And if something breaks in the in the in the back end where the customers are not seeing that, they will fix it themselves. I think that's the end game. And right now we are the early adapters but we're just in the transition phase to this final stage
Ben Sufiani: It feels so early, doesn't it? Like it's so clanky still and we still, we already have superpowers. So one of the superpowers we now get as marketers, we get huge support in creative generation. I'm not talking about stupid AI slop, but actually images, videos, copy that converts. What is your content creation process look now?
Marcus: So maybe okay, Paul, it's our topic, you can proceed.
Paul from DreamyTales: So it so parts of it have changed with AI and parts of it have been the way they have always been. So in the beginning, you always have to get an idea or inspiration. What does the creative actually is is going to look like that I want to run, right? So you can get inspiration via AI, obviously. You can just ask ChatGPT or Claude. I want to sell this product, you know. What kind of creatives could I run? What's the angles? What's the hooks I could test? Blah blah blah. this can be one approach. The better approach that we do most of the time is take a look at your competitors and especially the meta ed library and try to scout creatives that have been running for a long time, you know, from brands that try to sell the same stuff that you are trying to sell. And then you try to get like three to four concepts out of this process before actually try to build the actual creative, right? So say you found this one ad from your one competitor in the meta-ad library and you know it has been running for three years or whatever, and you're like, man, this is printing. I have I have to build an ad that looks exactly like this. there's basically two cases. Like one case A it's a static, case B it's a video. If it's a static, you are very lucky because you can just pump it into GPT image two and tell it to, you know, rebrand it to your brand guide and put your product on it, whatever. Because this image model is just, you know, so good that this will work. from our experience, static images don't work as good as video ads. So What we are trying to do is to gr create video ads with AI.
Ben Sufiani: Is it your experience that videos are converting better than statics? It's
Paul from DreamyTales: Yes.
Ben Sufiani: not always mine, it's mixed project by project basis. What is it for you right now?
Marcus: Yeah, I think I think it really depends on the product. so for us, for our products, I think it's very very much important to explain a lot of things to the customers, how is it working? to get their own like personalized gift. and also it's a lot about emotions. It's a lot about connections between people and this is the feeling and the atmosphere you want to deliver to the customer. And this for us at least works better by by videos. I think we we never had a real success with images at least on a scale base.
Ben Sufiani: Interesting.
Marcus: Just a little bit, but yeah, video works best.
Paul from DreamyTales: Yeah, I think so for DreamyTales this is true. with Vibe TV we are seeing mixed results, but we're still in early testing phase there, so not sure where this will go. But yeah, for Dreamy Tails, we usually then try to scout like some kind of ad. And then we try to first of all try to digest this app. So we as humans take a look at this app because we At this ad because we think that AI is it's pretty good at describing the you know details of a video. It can tell you, alright, there's a lamp in the background, the girl is around this and that, age old, and you know, has blonde hair, whatever. But it's really bad at describing the aesthetics of it. So is why is this probably performing great, right? Is it because it
Ben Sufiani: Mm-hmm.
Paul from DreamyTales: looks Weird in a certain way? Is it because it's very long or very short or what is it? So we try to look at the ads and try to come up with hypotheses why this ad might work well, right? So it could be a certain hook or an angle or a lighting or the person or the music or the voiceover or whatever. And then we try to break this ad down into these different hypotheses. And from there we first of all try to generate prompts that would generate videos. So we have this engine build that we for example put in an idea for an ad and it turns our idea into a very detailed long prompt for a video scene, right? So that's the first step.
Ben Sufiani: Mm-hmm.
Paul from DreamyTales: And then you from that prompt you try to generate The actual video with the model of your choice. For us, you know, currently we are using Sora 2, but we are also switching these models. like underwear basically always depends on the use case and what's what's the state of the art model right now. And what you wanna do is you wanna you don't wanna create just one video, but you wanna create like on bulk, right? So one from one prompt you would then create, I don't know, 10 videos or whatever and choose the one that you think fits best. You know, because still it can be there can be weird stuff happening in the background or the lighting is too perfect or too imperfect or whatever. So then again you take a look at it and basically choose the one that you like best.
Ben Sufiani: Now, is it one video or is it sequence steps that you still need to put together?
Paul from DreamyTales: So it's sequences that you stitch together in the end. And
Ben Sufiani: Hmm
Paul from DreamyTales: you would it kind of depends on the angle and on the ad. There's also ads that are very short, you can do them in just one video. But most of the ads that we have are multiple scenes stitched together.
Ben Sufiani: And what do you use for editing?
Paul from DreamyTales: So we to put it all together we use a tool that's called Creatomate. ever heard of it?
Ben Sufiani: No. What is it?
Paul from DreamyTales: CreatorMate is a tool that's basically a video editor that allows you to edit a video just as you would in CapCut or whatever. But CreatorMate then lets you export this template of a video as JSON. And then you can replace every little piece of that video with another piece.
Ben Sufiani: Awesome.
Paul from DreamyTales: So say you have a template for a video and the voiceover starts exactly at you know second. Three, Creator Mate lets you switch out that voiceover just by you know feeding that JSON into into your AI agent of your choice.
Ben Sufiani: Again, with codecs at the center, managing all of these tool connections.
Paul from DreamyTales: So for this, we actually built a simple HTML file that speaks to different N8N workflows. So you have one you Yeah. So you have
Ben Sufiani: Yep. okay. Good idea.
Paul from DreamyTales: one input field in this HTML file that speaks to an N8N workflow that generates the prompt for the video. Then you have another input field that speaks to the N8N workflow that generates three videos from that prompt. Then you have another input field, you know, that makes headlines, voiceover, puts it all together. Just to save on contacts, to save on tokens and just because it's so repeatable, basically. But then you can also just hand this to Codex and tell and tell it, hey, you know, look at this.
Marcus: So this is one way of doing it. It works very, very great if you're having always the same template which which works for your ads. But if you wanna experiment on on cutting times on voiceovers and stuff like that, we've also made very, very good experience with yeah, with with codex skills. So they're different kind of skills. There's a lot of tutorials how to do that. So that codex handles all for you. They have the tool base, they are basically rendering the videos on your on your device. putting putting voiceovers to it, putting like simple captions and stuff like that. so you can basically with one prompt create a complete ad which is cut it and stitched together.
Ben Sufiani: I've been messing around with the re-motion and the hyperplex tools and accompanying skills. Is that what you're talking about Markus?
Marcus: Yeah,
Ben Sufiani: Which one of the two?
Marcus: To be honest, I have no idea. I I I told Codex took
Ben Sufiani: Yeah, it's okay. It's okay. It doesn't matter.
Marcus: Codex to make a research what is the best one, what is the state of the art, and he he he he proposed something and I said, Okay, go for it. don't let me hear about it, just do it.
Ben Sufiani: one of the reasons we are doing this open exchange is because we have nothing to sell. Like no Paul, Markus, you have nothing to sell here except for your wonderful B2C products, which here is not the primary audience, but because we're looking for others to share with us as well. Their tools, what they're using. Like I'm sure I'm going to be busy exploring Creatomate afterwards. for example. And maybe there's this or that idea that you explore and we need more people like us who have been AI-pilled but who have this performance-marketing mindset. Yeah, there are not so many, I feel. And I would just love to in touch, get in the comments, connect with Markus Paul, with me and talk about this stuff. And just to share a little bit about my process, I have... this all in Cloud Code in skills and references file. I also do this analysis of what makes it work, what makes an ad work. But I have everything documented by Cloud in docs files essentially from our design language. to the tone of voice that we're using, different formats of ads from single images to carousels to shorts to infographics and it's always a consistent feedback process. Every batch of creatives that you are creating is an opportunity for you to nag on what is not good enough yet. And if you look at pirateskills.com on the insides, yeah, where all of the stuff we are blogging, we are creating custom images for every blog post and animate them. And that is such a smooth process now. As Marcus said, when you have a repeatable template, it is very, very, very nice. And for exploration, for Paul's idea, something I was thinking about is... This is new, no not new, but a popular Go-Dev product called YouTube Download, YTDL. And I
Paul from DreamyTales: Mm.
Ben Sufiani: was like, we could let our agent research the best YouTube channels and just have a daily or weekly... routine that pulls all of the highest virality videos from these channels that we are observing or the niches that we are observing, then slicing it and seeing what was the hook, what was the general content, but as marketers we know how important the hook is and the format, the idea of the creative and then just pitch us that based on validated data. Do think that could work?
Paul from DreamyTales: I mean, in the end, it's all testing, right? So I think this might work, might as well not work. Surely depends on the product you're trying to sell with it. So one hook and one creative might work specifically for that one use case in that market for that audience. Might not work for another product, for another market, for another audience. But I think that it surely makes sense to build a system that tests this, right? So I think the real superpower isn't in lying the i isn't in finding the the secret ingredient, but is in building a system that can test so many things that you know you will inevitably find a winner just because you tested so much.
Ben Sufiani: Yes, and I still notice how I am limited in this idea. have this campaign idea for some product. I create the creators, I set up the campaign, I monetize and optimize them, blah, blah, end of line. Instead of thinking of it as a closed loop. Yeah, like have you ever come across this auto research repo from Kapathi? The idea is that you let an agentic system in a loop. Do the creatives first batch with you, set up the campaign, see the results, but then continue the process. Yeah. And create more creatives and go left and right. See what is working and as long as early stage indicators like higher click-through rate and in e-commerce, you have a very short cycle time to conversions ROAS. Yeah. Is positive, go more in that direction. Yeah. Explore to a certain degree and it just goes round and round and round. And this idea with a YouTube downloader or Instagram downloader or whatever is how can I, what Marcus said, pull myself a little bit out of that interactive loop and thinking about it as an assembly line that starts from A to B, but build this. Have you guys already tried that or is that something we should think about together after the pot? What do you think?
Paul from DreamyTales: Yeah, one hundred I think this is absolutely the right way to think about this process, right? As e-commerce guys, we are heavily dependent on paid ads and to automate as much as it as possible. and optimize is absolutely the right way to do. What we do is we do it in a different way. So we are not trying to optimize the creative, but we try to optimize basically our Our broad campaigns, right? So we run campaigns in I don't know five to ten countries depending on season. And the system that we build allows us to test angles in one market, and if they work, we can replicate exactly the same ad in a different language for another market, right? So on the second. I mean I'm not talking about like roughly the same, but I'm talking the exact same ad, just you know optimized for that other market. So in that kind case, I think this is some some variation of that loop that you are talking about, but a primitive one. I think there there's a way to go and especially like with testing the creative and optimizing the creative, that is definitely a a road to to explore.
Ben Sufiani: So maybe to wrap things up slowly, Markus, from you I got the impression that your views at the moment, lots of people can now build stuff. Very few can also distribute and market that. Can you expand on that a little bit?
Marcus: yeah, I think I think that pretty much comes from the fact that right now you can build an eighty percent product pretty, pretty quickly. So if you're having a an idea, you can with the help of an AI j agent, doing doing eighty percent of the work and you're having something solid, you're having a prototype or something. But the the the rest of the one hundred percent, the twenty percent, this is the tough nut crack, you know. and this is where where also marketing comes in. into place. I think y y you're still rely relying on on this those this human input, this human feeling, the atmosphere, the the knowledge, the the experience you have as a marketer. and I hope this stays like this. I mean I'm not sure but it can be that in twelve or twenty-four months this twenty percent has been shrinked to like five percent or even lower
Ben Sufiani: Mm-hmm.
Marcus: and then Everyone can do what we we are doing, yeah. but yeah, I think nobody knows what's what's going to happen that in that terms.
Ben Sufiani: But it's worth staying at the edge of what's possible.
Marcus: Absolutely. If you're not if you're not staying at the edge somebody else's and then yeah, you're losing.
Ben Sufiani: such a new opportunity space and pre-AI, probably like since I'm 16, I have this image of my mind of like the Neanderthals. Like, I was always fascinated how slow progress was. Why the fuck did it take 30,000 years to go from a round stone to a sharp stone? And then 15,000 years to put it on a stick like for the first spear. And for me fundamentally, the image in my mind was... This is one monkey sits down on his ass and is thinking about round stone. What can I do with round stone? Yeah. And it's up to the monkey that sits down on his ass and thinks, how can we improve our quality of life? How can we now build a little piece of that Iron Man suit that we're able to build for ourselves? And we got a whole new toy put in our lab and it's down to us and it's not just the developers anymore that get to play but everyone. Yeah that is very fascinating for me. I hope we have inspired a couple of people to join us, think about... maybe regularly getting in touch with Marcus, Paul and myself to build a little mastermind, nothing paid, where we really want the practitioners of performance marketing combined with AI in what is left and right for that. Analytics, creative, the whole landing page, customer conversion experience. would be very very welcome to hear. Are there any last words or asks that you have for the audience before we sign off?
Paul from DreamyTales: thanks again just for having us, Ben. It's been a lot of fun and you know, our eyes widen when we talk about this stuff because we play with it every day, just like you said. So I think it's a great initiative to try to get more people together to come together and play and see what we can build.
Ben Sufiani: Not cuz you.
Marcus: Yeah, also from my side. Thank you, thank you so much for having us. it was very, very fun. and hope hope to be one of your guests for for another time.
Ben Sufiani: For sure, for sure. And we see each other on September 2nd at the Vibe Coding Cologne event where you actually open a laptop, show stuff, as well. So I'm very much looking forward to that. And I would say have a good day. Bye.
Paul from DreamyTales: Have good day.
Marcus: Thanks.
Paul from DreamyTales: Bye Ben.
