Make AI UGC Ads with Claude and Seedance 2.5

This ad is all AI. That's not really me eating those tacos. I made it with Claude Code and Higgsfield. I'll show you every step, every prompt, and what went wrong. Hey, I'm Sean. Welcome to AI Builds. Today I'm going to build a real AI video ad starring myself. There's going to be seven steps start to finish. You don't have to have any experience. I'm going to show you step-by-step what I did to create this AI video ad inside Higgsfield. Let's get started. So, this is how AI video works. I started with two pictures and one was a selfie and the other was a photo of my Scorch bottle that I made in that Scorch build. AI video works best when you give it real references. Words alone will work. Like you could do text to video. You get a way better output when you actually have reference images to show AI what to build with. And then, you know, pictures are super cheap. Like images are super cheap to generate. You better You're way better off generating high-quality images to in turn those into videos using those as references to turn them into videos than you would be just using text. Video costs a lot more, so test it with pictures first. I use Higgsfield for that. It has [music] almost every model, if not every model that there is right now. And then, again, if you don't have a Higgsfield account, you can click the link in the description. This channel does get a small commission. Claude Code wrote every prompt and checked every result. We used Claude to do this entire process for us. All right, first up, make my character. Let's go. So, first we're going to build a character sheet. It's basically just going to be me from the front, from the back, and then up close. Without it, my face would change from shot to shot. So, it's super important to create a character sheet whenever you're doing video generation. I ran the same prompt in all four modes. So, Nana Banana Pro, Jat GPT 2.5 Sunburst, which is the newest best out right now, Sea Dream 5.0 Pro, which is also pretty good, and then Nana Banana 2.1. These are all really good models, but basically the GPT 2.5 Sunburst just looked the best. It followed instructions the best and looked the best overall. Yeah, Sea Dream had like a shadowy floor, and like I asked for a flat gray. and you don't want shadows in your character sheets. The name banana 2.1 like it was really close but the face was like super wide didn't look close enough to me. So I I cut that one. Yeah, I definitely I picked GPT image 2.5 sunburst again, it followed instructions the best it looked the best. I thought that that was the best choice for this video. Claude put the face next to my real photo to check it and I just always had like even if I say like if I pick an image and I say hey this is awesome. I'm going to use this. I always double check with Claude say is there anything here that I missed? Is there anything like if I did let's say I did a four image generation of the same prompt. I will choose mine and not tell Claude and then Claude I will tell Claude to choose the best one and I'll see if they match. Often times they do but a lot of times like Claude will catch something like hey this one's got six fingers or you know, whatever this girl's ponytail is not included in the you know, it's supposed to be on the left side it's on the right side. Little stuff like that. I always like I always choose the final output like I double check with Claude to make sure that I don't miss any big detail right before we do image generation. Again, all right right before we do video generation cuz video generation costs a lot of credits. So it's got same eyes same beard and so that that was the one that was the one I chose. All right, next is the place in the product. I chose a kitchen cuz I wanted to I wanted to show off where I'd be eating this hot sauce and first one just had no camera words and like AI picked the angle. So whenever you do image generation with AI like inside Higgsfield, you always get the best output when you decide what the camera angle is. It helps you you know, it helps you with the mood it helps with the shot overall like sometimes you'll get a good output with AI but I highly recommend that you choose whatever camera angle makes sense for that shot. Act as a real director. That's the only way you're going to get a real output. Yeah, AI picked its own angle just added stuff I never asked for. It's the same kitchen but I when I said where the camera stands that's where I got the best the best output. So it listening the camera angle was good but the camera was just still too far back. We're making it for social media and it can't be small right like the the character in the video has got to be large or take up the bulk of the screen you know. My photo became the front and the side sheet. Claude checked every letter on the label and all of them were right. I had Whenever you do any AI video or image generation that has text, you have to triple-check. And again, like there's sometimes where I miss it, so I ask Claude, like, "Hey, can you please make sure that every bit of text on this is 100% accurate?" The red habanero hot sauce the classic has a heat level of six out of 10. When I originally generated the image for this brand on the previous build, AI kept generating five out of 10, so I had to make sure that it said six out of 10. All right, next is the video part. This is the most fun part, and this is where I get the most excited. Let's go. All right, this is where This is where my AI video skill comes in. It runs a Claude Code. You input the skill. Again, the skill is in my free community. The link's in the description. You can just download it and upload it into Claude Code. Claude asks a few questions and then writes the video prompt for us. Claude's going to ask you based on what kind of video you're creating because of the skill what you needed what information you need to provide it to get the best out. Basically, the prompt says what happens second by second and where the camera is. Again, the camera angle is super important, otherwise you can get some random shot that's not going to work. It names my eye color, beard color, and the face stays right. So, that's like the whole thing is Even though you're providing an image reference, you still describe it in the prompt. With AI video generation, so many little things can go wrong. It's They're so powerful and all these models are so good, especially C-Dream 2.5 or C-Dance 2.5, but it could still make errors. Like, you want to make sure it doesn't transform any characters, so you just All you're kind of doing is like reiterating. Just You're just double-checking that the whatever reference you have in the video prompt, you just write in the description again so that Higgsfield or whatever model you're using knows exactly who that character is. It's just super important to re-emphasize all the details of the shot. It spells it out word by word on the label as well. Like, you want to make sure that every little detail is in the prompt as well, even though we're providing image references. So, we reference the four images inside Claude, and they're all in order. It's got me, my face, uh the kitchen, and the bottle. Uh then we do the drafts, right? So, we test it, we check it, and then we fix it. That's That's this whole step here. Making sure that we don't have any details that got over that got overlooked. But, here's the first draft. I forgot to turn the audio on. So annoying. Like, I I always have audio I'd say 90% of the time I have audio off. This time I just forgot. So, I just wasted the credits. It is what it It still happens. It is what it is. You can make sure like I have Claude double-check. But, I tell you when you guys download the skill, it's going to ask about audio before you actually do any video generation. Just by habit I screwed up, but it just super important like triple-check all the settings before you do any video generation. And the pour was like super tiny. Like, it wasn't like a joke. This was a bad script. I'm not going to lie. This is a bad uh video idea. Again, you don't really know until you see it. Like, in my head it looked good, and then when I actually uh the when the video came out, it wasn't It wasn't the best script. And that's That's my fault. I just My imagination wasn't great for this first one. So, yeah, I did the same prompt again, and then it was way too much. Like, it just over-poured. Whatever, it's not a big deal. Yeah, it was There was like sauce everywhere, but the taco was still dry. Like, for as much sauce as you poured on there, like the taco should be like soaking wet, you know? So, with the three drafts side by side, I had Claude check them all second by second, frame by frame. Yeah, it was like the same prompt, but like two completely different videos. Like, it didn't have the same feel whatsoever. And that's That's kind of normal. Whenever you do a regeneration, you kind of don't know what you're going to get, even with the same prompt. Even if it's super detailed, so it That's just kind of part of it. So, we just spelled out the pour, the four shakes, one taco. All right, now we're making it sound like me. This is All right, this is voice draft one. And I made it up, so it's like it's not me. So, what I did is I recorded two lines. You can do this on your phone. Like, you can do it in OBS or whatever you use uh for recording device. What you do is you create what what the lines are, and you try to have the same spacing. If you're a little off, these models are so good, they'll still pick it up. But, try to be as close as you can as far as the timing goes and the delivery goes. And what you do is you record it, but that MP3 you have to turn it into a black video into an MP4. So you can't just have an audio file. So what I did is I use OBS, which is what I use to make all my YouTube videos, and I just made the screen black, right? So like don't record screen, just record audio, and it's still an MP4 output. You use that MP4 output as a reference as a video reference instead of just doing an MP3 audio reference. All these models do a way better job for when when it has an MP4 with a black blank video than it does with an MP3. Don't ask me why, I don't know. That's just it is what it is. And then here's here's draft two of my voice. This is Scorch. Start with one drop. Seriously, one drop. So bad. Hey, that could just be a bad script, but hey, it worked, you know? It took my voice. That's that's what I'm looking for here. It's not like perfect, but it was it's close. All right, now we're finishing it. So I made it sharp that like the draft in 480p like you could see all the pixels like it wasn't super sharp. Um it's just a small size for what we're trying to do here. So inside HeyGen field, if you go to heygen.ai/upscale, I used ByteDance upscale and it went from 480p to 1080p. It took about 4 minutes I think to make that generation happen, but um instead of regenerating this video, I just upscaled it for like 24 credits. So the quality is way better. Um and here it is finished. This is Scorch. Start with one drop. Seriously, one drop. Clod made my face redder uh right on my computer for free. So anytime you do AI video, don't think that because it's not perfect that you can't make tweaks. I'm going to show you later when I do edits in CapCut, but Clod actually made my face redder without it could it was free. Clod code did it all for me for free. Um you're going to see if do some animations and stuff like that, too. But, if your video output isn't perfect, it doesn't mean it's over. Like, Claude can make edits, you can make edits in CapCut. You can cut certain sections and move them around. Like, all this footage is still useful even if it's not perfect. So, don't worry about getting a perfect generation right out the gate. And it's it's so nice that Claude can just make these tiny edits for you, too. It's It's It's honestly incredible. So, again, this Here's the ad finished. There's no music. The silence is what makes it a joke. I just I I just didn't love it, right? I The script was all wrong. I I shouldn't have I shouldn't have put more thought into it. It is what it is. It happens sometimes. So, I tried a new joke using all three bottles, and I think the script's a lot better. [music] So, I had to create two more bottle sheets. And again, Claude checked every letter, every text on this to make sure they were right. But, now I have my mild green jalapeno, my classic red habanero, and then my extreme scorched earth. Again, this is all for my fake scorched brand. But, I created bottle sheets for each one. So, my bottle photos actually had some typos in them. I referenced some wrong sheets when I did this. But, the prompts actually fixed it cuz I put the prompt in to say, "Hey, the classic says six out of 10, right? The green jalapeno is two out of 10." So, the prompt actually fixed my original images. And when I got these bottle sheets, all the text was right. So, again, it's like super important. Just because you're using a reference, it doesn't mean you don't also respecify the exact details inside the prompt. Mild. Nice. Classic. Good. So, here's the first try. All three tacos look the same. And again, I did this in in 480p. That's why the the quality is terrible. And then the the labels were scrambled, too. So, like, not not only did the tacos all look the same, but the labels for the bottles were were all scrambled. It was so much better and so much easier to create a new starting frame. So, again, just think about what we're doing here, right? Like, we're trying to get AI to create generate this entire video scene for us and it's just making it up. It's filling in gaps. So whatever whatever information we give it, if we don't give it 100% of the information it needs, it just fills in those gaps. So I'm like, "Hey, you know what? This would be so much easier if I had a starting frame for the scene, still used all the references, but I had a starting image of myself of exactly how I wanted the scene to look." And this is how I'm going to do it all from here on out. Even though Seedance 2.5 is like really good, the models aren't quite there yet. So you got to still help a little bit. So now it's like my rule from here on out, I'm always going to have a starting frame for every video of what I want the scene to look like first. If I change, I want to put the chicken like for the tacos, I wanted to use chicken because it's super light compared to like carne asada or whatever. And I wanted each sauce, I wanted you to see each color. So that's why I'm like, "Okay, let's do this starting image. I want the tacos sideways and I'll I want it to be chicken so that it's super light and you can see it. There's no guacamole or in the way or anything like that. It's just super clean with chicken so you can see the color of each sauce. The first one Claude kind of screwed me, not going to lie. Chose the wrong model. It is what it is. Like I have to respecify. I'm the director. I'm the one telling Claude what I want. Like I should have respecified instead of allowing Claude to make that call for me. That's on me. I do that sometimes and it's and on the second try we used Dan Banana Pro and uh it came out exactly what I was looking for. So on this next draft, I was staring at the camera the whole time. So I'm eating a taco as I'm sitting like it was just super awkward and unnatural. Um so I had to re-prompt that. Yeah, I did that at 15 seconds and I think it was do I think that the model did that because I didn't give it enough time. I needed to give it a full 20 seconds to like not have that awkwardness in it because the character had to take like three different bites on camera inside 15 seconds, talk about two of the sauces. Like there was a lot going on and it needed a full 20 seconds to get it to get it right. But on the second generation on 20 seconds, yeah, so he didn't go over to the far right taco or the far his far left taco, it's right on our screen to the scorched earth, right? The super spicy one. So, I had to make this edit. So, what I did is the first generation I didn't like had him grabbing the taco. So, I was able to cut that in inside CapCut and I I was able to take that little clip and add it to this generation and I was able to edit it inside CapCut using the first take that I didn't like. I was able just to add that one little part in of him grabbing that third taco. And that way I didn't have to do a full regeneration. I was able to save another, you know, 100 and some credits, right? Again, that's what you do with this. It's like none of this video gets wasted. You can do these edits and piece them all together, right? So, it doesn't have to be perfect on every generation. It just has to be usable, right? You got to have some generations just have to have some parts that are usable and you can edit them inside and make it look good. Yeah, these two like 1080p image generations cost me about 560 credits. But, Claude was able to do these fixes for me absolutely free, right? So, like again, [music] everything can be editable, so nothing's wasted as long as you get something that's usable out of those clips. You don't need a full, you know, 8 10 15 seconds to be good if two or four seconds of those clips will work for you and you can just mash them together in the editing. So, again, once he took a bite, the tacos kept showing up and [music] what I wanted to do is I wanted to act like I just cut like after the first bite and he set it down, I wanted to do like a a sharp cut to him saying his line and we just skipped the part [music] where he was actually eating the whole taco, right? So, we did that with the green sauce one, we did that with the red sauce one and Claude was able to just make that edit for me. I didn't have to like I don't know about masking. I don't know how to do that inside CapCut. I can learn, but again, that's just time I don't have right now. Claude was able to make that edit for me and I was able to make the changes inside CapCut with Claude's edited file. So, it was it's pretty awesome. Anytime you're struggling with any type of video issues, again, Claude can now make those edits for you absolutely free, so it's awesome. So, basically you just take all these edits like all these parts of these videos and we can just make the edits inside CapCut and it's going to look seamless. So, what I wanted to do is I wanted to add the sound effects like I wanted I wanted that last part to have some type of sizzling or some type of like you like you know this is super spicy, so I just added that little um sound effect in. And then also, I wanted to add a filter on the whole thing. So, sometimes AI videos, especially like C Dream C Dream 2.5, it it looks too perfect, and that's how you know it's fake like cuz it's too perfect. It's not real life, especially when it's a social media ad or a social media video like this. So, I put a filter on. I forgot what this one was called, but I put a filter on the whole thing so that it looked a little less perfect. And you can do that inside CapCut. It's super easy. Like there's a bunch of ones that you can use. But whatever whenever a video doesn't look perfect, you can just put this filter on, and it'll make it look so much so much more realistic. And then one finishing touch I wanted to do is I wanted to like you know get this is a video ad. Even though it's a fake brand and fake video ad like I wanted to make it look like like it was from from a real brand. So, what I did is I asked Claude I said, "Hey, this is what I'm looking for. I want the logo, and I want this being said in this type of voice, right?" And it's crazy how Claude was able to generate this for me absolutely free using Claude design. And we're able to make this little like video card this little logo card at the end like the signature sign-off uh part of the ad using Claude design. Like I said, "Here's my logo deck or here's my brand deck for uh Scorch, and then this is this is the line I want you to say, and then this is the type of sound I want I wanted it to be like a monster truck announcer's voice, right?" And Claude just generated this for me like in minutes. Claude just generated for this uh this for me in Claude design. It was absolutely incredible. So, I'm now able to attach this clip at the end of my video ad, and now it sounds like a like an actual video like it's awesome. This is a real video ad now with all the elements you need. Claude design was able to do this for me in minutes absolutely free. So, I added the fire, the logo, and the if you dare at the end, which is the Scorch uh signature. Yeah, that announcer voice like it runs free on my computer like no credits whatsoever. It's freaking awesome. All right, here's the first ad. This is Scorch. Start with one drop. Seriously, one drop. The script was bad. Like, I just wasn't creative enough. It is what it is. It's not terrible, but it's not great. It works, whatever. All right, here's the second ad. Mild. Nice. Classic. Good. Scorch. If you dare. It didn't take that long at all. It's just pretty awesome to be able to do this and editing was super simple. Claude helped us with all of it. I did this whole project inside Claude Code and it's It's just awesome to be able to do this and not have to know how to do everything, but be able to get a good output. You be able to generate your own AI video ad. It's pretty awesome. So, guys, that was the build. Uh we did the pictures first. Again, the pictures are cheap. If you generate good assets, you're going to get a good video out of it. Again, don't forget the starting frame image. That is, for me, the most important part of video generation is even though I I generate all the video all the image assets, that starting frame image sets the tone. It just makes the video so much easier to generate. And then we do all the video testing small like in 480p. Once you know the prompt works, once you like the style and the camera angle and everything, then you can bump it up to 1080p. And again, sometimes you're not going to love it. Like, sometimes for 24 credits or 20 whatever it is for 480p, like, it's not a big deal if you don't love it. You can just have a different idea, adjust the script, adjust things you don't like. It's not the end of the world if you don't like what comes out of when you're doing 480p. And again, we're able to use our own voice now in all these video ads, which is freaking awesome. And again, every prompt I used for this build, guys, it's in the description. It's in my free Skool community. You can download whatever you want, make whatever changes to it. You can do this build along with me in there. All the prompts, all the skills, everything is in there. And if you want to see how I built the Scorch website, the video will be up top here. All right, guys, that was today's build. I appreciate you following along. Please like and subscribe. And again, everything I do on YouTube is in my free Skool community. Check it out. I'll see you guys on the next one.