How to Make AI UGC Ads in 2026: Sora 2, Claude and the Full Workflow

0:00Loukas:What I will run through is like the workflow I use to create this video because I think people may want to start utilizing this.

0:12Did you see the Sam Altman, Claude backlash on Twitter?

0:19Loukas:Salty Sam. Yeah. Is that what they're calling him now? That's what I call him.

0:26Olly:He's known to bite. he's had a few instances of this where he's like bitten at certain things.

0:33Loukas:He definitely throws his toys out the pram, doesn't he? I think it feels to me like entitlement because, you know, there was a period where OpenAI were just very clearly dominating the consumer AI LLM market, but that has very much changed now. But yeah, give some context on what this is and what happened. well yeah I think

0:57Olly:the Antropic ran a series of of ads about Claude right but taking shots well they've taken a few shots at OpenAI over the last six months but the main like narrative behind it was like that Antropic are never going to run ads on their platform because they're serving a different customer but also they just don't believe in it being like an optimal experience for them.

1:27Loukas:Yeah. And it's very timely because it was very soon after OpenAI introduced ads to their platform. Very quick.

1:35Olly:And the ads are really funny. I like the bodybuilder one where he's on the pull-up bar. Yeah, yeah, yeah. They ran him in the Super Bowl as well, so.

1:41Loukas:Yeah, the joke essentially is like it brings, it almost like recreates what the chat would look like with yourself and the LLM, but in real life. So you might have a therapist speaking to a woman about her troubles and it feels very dry in the sense that it may do when you're speaking directly to AI. And then it just is interrupted with the therapist, as an example, just suddenly trying to shill some form of product or as part of an advertisement.

2:15Olly:Yeah, it was really good, really good execution. There wasn't many good Super Bowl ads this year, I don't think. There's quite a lot of films, but that's like standard, right? yeah I've not actually seen

2:25Loukas:I've not watched many of them I saw was it the Pringles did one with Sabrina yeah yeah

2:31Olly:you got it yeah I don't think but that's the point there was no like big viral moments I think this for me the one that but if like the Anthropic one was probably the the one that felt most viral but it wasn't maybe that's just because we're in this niche bubble

2:49Loukas:yeah it does feel though like there's more we're more consistent consistently breaking through or transcending D2C Twitter and or not necessarily D2C Twitter but like yeah e-commerce Twitter and Super Bowl ads like similarly last year with the Hexclad one yeah yeah I think two other

3:13Olly:things I've seen from these big frontier labs one was the Pentagon thing did you see that? where Anthropic asked the US the US CIA how they were using Anthropic but that in itself was a security risk because they knew that they were using it for certain things so that's started a whole debate I think around the security of using these models and then obviously OpenAI also bought OpenClaw which I haven't actually used that much yet, I've played around with it but haven't used it extensively

3:50Loukas:have you? No No, I haven't. More just because people seem to scare me off with the security risks. But I think I need to surpass that and actually just give it a play or buy a separate device to use it on.

4:04Olly:Yeah, I think I haven't either. Just because I've been so deep in using other things. If you've only got so much time to experiment and test these tools, so I've been more focused on Cloud Code and co-work and just Cloud in general for now. but yeah obviously got a pretty big exit that Peter I can't say his surname Steinberger

4:26Loukas:Steinberger yeah good name yeah he yeah he exited a previous company for like over 100 mil didn't he

4:32Olly:yeah so they bought that which is interesting just feels like a bit of a that's like a big play for them to get back at Anthropic

4:40Loukas:yeah exactly but I mean his story is like super interesting because obviously he sold his business for over 100 mil um so he comes from a very tech code first background um essentially retired for a few years and then obviously as ai um as this kind of ai evolution and revolution revolution started to become more and more um present in in just consumer consumers lives he essentially started looking into it more and more and thinking, okay, what is it? How could I support in this? What can I do to be involved in this journey? And I think there was a period I listened to a podcast with him where he was like, he went six months.

5:27Loukas:He was like, why are people not doing this open source way of actually kind of communicating with agents and getting agents to work for you? Okay, but it's going to come. It's going to come. And then it just never came. And then he actually just went ahead and coded it himself. And I think he, 48 times he failed. and then it was like the 49th that it ended up working. So super interesting story. I know he originally called it Claudebot, which Anthropic came after him for. He had a cease and desist there. And then he re-nate, I know it's open claw now, but it was something in between, wasn't it?

6:01Loukas:Yeah, all I ever know is is the lobster. Yeah, exactly. The iconic red lobster.

6:07Olly:Which is funny. So yeah, it's interesting that they've obviously had issues with Anthropic and then being bought by OpenAI. Another one, we could go down like what's happening in AI all day, but did you see Manus have started to release quite a bit in Facebook? So Manus are a bit of like a, for those who haven't like experimented with them, Facebook acquired a company called Manus AI for about 2 billion, maybe three months ago now, maybe more, or was announced publicly about three months ago. and what that is is like an agentic layer on top of models um and you've started to sort of see them slowly merge into the uh into the meta platform so um when if you have like real analytics on reels on organic will prompt you to use manas to make more reels or you just leverage the tool to create more content when you're in the ads manager you get a manas pop-up to help with analytics so it's interesting that they're starting to bring that into the platform and merge those two together the weird move though was they brought out like an open claw through Manist replicated the concept which obviously Mark Zuckerberg has always been famous for just stealing people's ideas and doing them better but for some reason they launched it on Telegram and not WhatsApp which made no sense but they own whatsapp why not put it into there maybe because of the encryption side of the platform but yeah i think i'm going to see a lot more from meta manus those two things converging i think

7:49Loukas:would be interesting yeah that's super interesting um cool well let's move on i wanted to actually just touch on some bits that i've been going deep into ai ugc over the last couple of weeks i think like transparently i see this is just a different vehicle like a different way of delivering a message and it also opens up capabilities from the perspective of like being able to cast the net wider with say hook uh variations how can we tap into a broader range of emotional states or test different messaging test different persona call outs um you know for a cheaper price essentially which will mean we get to learnings quicker and as a result we can then drive more impact after the fact.

8:33Loukas:So we've obviously been going a lot deeper into this. It's one that I did, I went through a similar process with last year where started using it a little bit, just it just wasn't where it needed to be. And I just, as a result, put it on pause. I'm like, I'm going to come back to this in like six months where likely we're going to see some improvements there. And I guess like the verdict is like, it's come on leaps and bounds, like kind of hugely significant improvements in the product and how it works. I think the key thing for me is like, how can we do this in a way that, how can we build our own workflow that doesn't require to use one of it at all, like an ArcAds, which is just quite expensive.

9:15Loukas:And that almost kind of defeats the point of the reason you would use something like this. Let's put on the screen, just an example of like one video that we created. We'll just quickly show this now.

9:25Olly:Guys, Monjaro made my appetite disappear. but the problem is I wasn't getting any protein and had zero energy. Then I found a hack that made my coffee a 12 and a half gram protein drink. It's called Nutrition Geeks Collagen. Two scoops every morning in my coffee and just like that 12 and a half grams of protein.

9:43Loukas:When you're losing weight quickly your body can break down collagen faster than it produces it making the actual collagen intake even more important. If you want an easy way to boost your protein and support firmer looking skin click the link below to get yours today so essentially what i will run through is like the workflow i use to create this video because i think people may want to start utilizing this i'm not going to talk about like some use cases or ways that we can apply this to what we're doing now i'm still pretty early in this process so in terms of what you from a use case point of view um more will come and we're going to we're going to talk about this more and more but i think one of the big things that i've i've been able to identify is what um video products or image products are best for different things and i think this is i guess where maybe the bottleneck lies right now is that they're all different it requires a pretty hefty tech stack to be able to get the strongest output because they all do different things uh well and as a result if you want to have a layered piece of creative or layered ad um then you're going to have to use multiple different uh different products it's a lengthy list exactly so the first thing that i want you to do is use either higgs field or kai which essentially are both api marketplaces so they're really great because what it means is you don't actually have to be jumping from Sora 2 to Kling to Nano Banana.

11:13Loukas:All of these image generation and video generation tools can all be centralized in one place, which is great. You just buy tokens directly from Higgs Field or Kai. So that's great. In terms of user experience, Higgs Field is probably the best. You can build storyboards there. You can rerun prompts if, say, it hasn't worked as you would have expected versus actually having to just recopy and paste it kind of actually gives you a bit of an insight into maybe why something didn't work you can also upscale as well through through Higgs field which you can't do on Kai AI but essentially you would use a tool like that now the first thing you would want to do is use Claude which I found is the best for actually creating prompts now click if you if you comment down below prompt will give you a really strong prompting template which will give you a good baseline that you can build on but we've essentially created a Claude skill that really deeply understands how to best prompt for image and text to video essentially so Claude is best I've found for that specifically so essentially using our Claude prompt we are I give it some just pretty basic information I we want to have a let's say woman that is running that is walking through her kitchen make sure there's camera shake um she needs to say this she's from england uh with a british accent and it needs to be shot on an iphone 14 if you put that in very kind of basic level information as well as your dialogue that you would like her to say um and then just use the prompt that i shared below it will actually put that information in the exact prompt um so do that first and then you essentially will have a prompt ready to then put into Sora 2 now Sora 2 is the best product to use for visuals with uh audio hooks so I would say like you can obviously you know you can extract video that's up to 15 seconds on Sora 2 and that works really really well from the perspective of like being able to to get hyper-realistic UGC-style content where you have good camera shake, the audio is really good from the creator, it makes it sound like it's actually shot on iPhone.

13:32Loukas:I think sometimes these text-tip video, you can kind of tell there's almost a slight distortion to it.

13:38Olly:Yeah, I think you can. When you're close to it. It's a question of, could the consumer tell? But yeah, I definitely think you can.

13:45Loukas:Yeah, exactly. So Sora 2 is definitely best for that. um so then you put your prompts into Sora 2 it obviously outlines what the subject looks like etc etc the dialogue it should do a pretty good job then of outputting um a you know a creative up to 15 seconds that could be used either for your hook or the first 15 seconds of the video the example we shared was it was the subject talking about a story it felt very raw comment reply on screen and then from that point we go into voiceover and I feel it felt like that was the most efficient way versus trying to replicate voice with video um after the after the fact so the next thing you would do once you've got that sora 2 output is put that directly into 11 labs and 11 labs can literally just recreate that voice just from a 10 second clip which i think is quite scary actually um from like a um yeah reproducing the exact same voice so then you can put the rest of your script into that voice as a voiceover from that point essentially what all we're doing is is creating b-roll footage of that subject now i layered in some actual real footage of b-roll because i think if we want to move quicker you should have lots of kind of low prod ugc style b-roll of like people um you know opening the product unboxing the product um you know um putting it into a drink whatever it might be so you should have footage there which is a good way to just speed up the process and not have to rely 100 on ai output to to be able to finish um the the actual ad itself but one thing that you will want to likely do is take image take an image of an existing ugc subject who might be sat down let's say talking to camera or might be filming themselves self in selfie mode drinking a coffee and literally screenshot that screenshot the subject from our opening Sora 2 prompt and then you essentially will want to take that put it into chat gpt and just say can we put our avatar our avatar that we created from our Sora 2 prompt and put her into this scene to replace the other woman and what that does is it creates a really great json prompt output that means that we can actually create a really good reference point to turn that into video so we've got our json we then put that into nano banana which will then essentially um bring that to life so it will then take that the two images take your prompt and it will recreate the image that we asked it for with our exact same subject that we had within the hook and i'm going to be showing some of this on screen as we go through it now um then essentially what once we've got that nano banana output cling which is uh is another text to video tool or image to video tool that is is best for b-roll creation all we need to do there is literally take the image the the reference point and say bring this image to life so if it's like the subject's sipping a cup of coffee but it's a static from nano banana it will literally bring it to life and turn it into like an actual video so then i would layer those in as part of it so yeah that's a rough rundown as i say i we will be shown this through on screen as we've gone through and we're going to be going a lot deeper into this and other use cases and how we can utilize this but um that's how i would recommend getting started and as i say comment down below uh prompt and we will send you that resource so that's really complicated there's a lot of steps

17:27Olly:where do you think it's like how do you think that's going to simplify say let's say wait what's your expectation for say end of this year what do you where do you think it'll go do you see it being in an arc ads do you see it just being able to be done in one to two tools what's that hypothesis

17:43Loukas:i think the hypothesis is like if we're in a position right now where they can all do different things well based on the law of accelerated returns and how ai is you know continuing to transcend expectations of of development there is no reason why within the next six months every single one of these tools should be able to do this individually on its own versus having to go into chat gbt because it does the best json output etc etc so i i would say i'd like to say we will be there within the next six months where you would be able to do this in just one of these LLMs or video generation tools.

18:25Loukas:I think from a short-term point of view, we are going to look at how we can automate some of this process through the use of agents that do specific parts of the process. So for example, one that is really good at picking an image and turning it into a reference nano banana one, as an example, one that's really great at B-roll. And then we can look at how these work together, essentially. But I would say, as a starting point, the easiest thing you can do that is super quick and low barrier to entry is use Sora 2 for hook variations. I'd say, just start there. The example I showed you is an end-to-end creative.

19:10Loukas:You don't really need to go that far at this point. You can create 15 second videos from Sora to have a subject just talking to camera, walking with the phone. We know that in itself could work as an ad, like as a standalone creative without any editing associated with it. We know that that, you know, could be, we could test different messaging, like tapping into different emotional states, you know, different personas. as we talk a lot about these micro personas in the last episode with meta and how do we how do we tap into different desires different fears different hook frameworks we could as a result and very low additional cost if 10 if it's 10 cent an output for the for for for one dollar we could do 10 a different 10 additional hook variations of a concept that already exists

20:00Olly:yeah and i think it's using it strategically isn't it like where does it fit into the process based on what it's able to provide in terms of value right now. Just see it enabling us to validate so many more different messages that we maybe wouldn't have prioritized if we were to actually do the shoot with a creator or with our production team because we're able to do it for such a low cost. And so I had to fill some of those gaps with those sub-personas on micro-personas that are less obvious when we first start on an account. yeah 100 % I guess the other area that we've been going pretty deep on as an agency and I think arguably the most valuable part of Claude right now maybe even the whole like AI stack is skills I know we've referenced these on a previous episode and I know you referenced a couple in that workflow I think I sent you this article didn't I yes you did this recursive self-improvement loop I've been experimenting a lot with skills so building them for different parts of our workflow whether it's research, analytics certain parts of the creative process skills enable just vastly better output and better efficiency through Claude use just by being able to partly just because they're able to inject themselves and Claude's able to understand when to call a skill throughout a workflow, even if it's like a multi-step process.

21:35Olly:How have you been using this recursive self-improvement loop? I can give an example of my side after you've led.

21:41Loukas:Yeah, for sure. So essentially what this is, is like, how can we build context or build a skill within Claude that allows us to quantify what success looks like at different parts of the process and almost like a grading system. So it's like working with your manager, right? Like you create a script, you send it to your manager to review, your manager sends a load of feedback back. This could be better. Maybe you could have done this. You great hook though. And what essentially this creates is the reviewer. It creates that instant feedback loop with the individual that may be creating a script or creating, you know, asking for feedback on a hook or whatever the part of the process might be.

22:27Loukas:and it will be able to give, it will be able to directly grade each part of the script based on its knowledge and context and how I guess we've set up that recursive self-improvement loop. So I've built one specifically for script writing. So essentially what that means is if you call the script writing self-improvement loop as a skill, it will then be able to take the skill that you send it and it will it will grade each part of that script so it will grade for things like hook quality storytelling does it align with the persona that we are speaking to um are we bringing in curiosity loops are we tapping into core human desires all of these elements it will grade and essentially it will like it will need a minimum of an eight out of ten on all of them to be able to pass so therefore you will need to make changes yourself and then send it back and it will regrade it and give you context on that grading and i think it's a great it's a great thing to do and we'll put the shout out to vibe marketer who who put this um article together it's a great thing to do if you've got new people coming into the business and maybe one of your constraints is time to to actually give feedback to these individuals it's it's a good way to actually almost download the way you think and what you grade success to look like and whatever part of the process it is and apply that to to yeah whatever that thing is you need grading so we do it for script writing hooks and then also static ads as well how have you been using it

23:59Olly:i've started to build one for landing page reviews so again persona personas top ads and then rating the congruency between ad and landing page and whether we need to break whether and a lot of that is like many of our clients will be making a thousand ads but we'll be sending 400 ads for the same product through one pdp even though it's speaking to like five different personas so just putting that uh put uploading those top ads maybe it's 12 15 pieces of creative into gemini either pairing that with the creative and getting it to like just prompting it to understand how those ads are split across those across those creatives then uploading a pdf a website to pdf of the pdp and just um getting it to rate uh alignment between headline copy offer um objection handling and then suggesting like five key levers rated from highest impact to lowest impact, impact in time or impact in effort for each page and that's leading often to like such easy quick wins of like splitting out a page across three new personas and then every time we launch a new ad for a new persona we're launching it to that lander.