# I Sold a $1,500 AI Commercial in 60 Minutes With Claude Opus 5.5

Channel: Sanji Nai-Chien
Video: https://www.youtube.com/watch?v=MuiUtTvGD8I
Duration: 26 min
Language: English
Words: 4355
Transcript page: https://viewrankai.com/tools/youtube-transcript/MuiUtTvGD8I

---

[0:00] I made this entire commercial from scratch with the new Claude Opus 5.5 and a real-life client paid me $1,500 for it. Hey! [music] What kind of work is this? I demand [music]

[0:23] [music] results. Are you listening? [music]

[0:38] [music]

[0:45] [music]

[0:50] [music]

[0:56] [music]

[1:04] [laughter] Now, if you're somebody who makes AI videos, then you know that generating them is the easy part. But getting somebody to actually pay for them is where it gets hard. So, I gave myself 60 minutes to do the whole job, to find a client, make the commercial and its variations, and get paid. And absolutely everything here counted, waiting for replies, generating the footage, and editing the final package. Now, I'm going to walk you through all of it, from the very first message to the final edit, so that you could actually go and do the same thing for your own clients,

[1:36] even if you don't have a portfolio. Every prompt and the Claude skill are in the blog post link below. Now, at the end, I'll show you exactly how much of that $1,500 I kept after my costs and how to keep your generation costs as low as possible. Now, the clock starts right now. And step one is to find a client. I find all of my clients by searching Upwork for AI video ads sorted by newest. Now, generally fresh posts have the fewest proposals, and that's exactly where you want to be. But, since we only have an hour, we can't just take any old job. The brief has to give you three

[2:15] things: a specific [music] product, a clear place where the ads are going to run, and product photos and branding that are ready to use. Now, if any of these three is missing, then you need to skip it, or you're going to burn half an hour just figuring out what it is they want. Okay, this one looks promising. They're a small smart TV brand launching a new model, and they need an ad for YouTube, Instagram, and TikTok. Now, what I don't know yet is the deadline or whether their assets are ready. And

[2:43] that's going to go into my proposal. But, here's the big problem. They've never worked with me, and I'm about to ask them for a thousand dollars. If they ignore me, I'm back to searching with even less time on the clock. So, what makes them take this seriously? Now, in my [music] case, it's actually a finished commercial. This one is from my previous video, and it's already in my portfolio. Before they even reply, they can see the product shots, the consistency, and the quality of the output. But, if you don't have a portfolio, in just a minute I'll show you how to land your first job without

[3:16] one. So, here's my proposal. Feel free to pause the video and read it. Now, you always want to end your proposal with questions that they can answer in one line. Their answers give you the scope, and the deadline gives you the price. While I wait for a reply, here's what to send if you don't have a portfolio yet. Pick a specific business and make a short commercial concept for their product. If you're not sure how, then check out my previous videos where I actually walk you through making your first AI commercial step-by-step. You want to label it as a concept, watermark the preview if you don't want it to be

[3:52] stolen and then send it with a personal message. Here's an example on screen. Now, they can see their own product in an ad before they pay for anything. That's what makes business owners choose you. You're showing them exactly what they'll be paying for instead of just asking them to trust you and that's what

[4:10] gets you paid right away. 13 minutes in and they reply. And honestly, I got a little bit lucky that I got them to reply this soon because they need the job done within 24 hours. And also, their product photos and logo look pretty much ready to be used. Now, I need to price the package. The vertical version has to work on a phone and the three openings let them test

[4:34] different ways to introduce the product. Every single version means more work, so I need to take all of them from the same scenes. I simply recut and reframe for every new version. Now, for the package, I'm quoting a thousand dollars. [music] Why would a brand pay that? Because they're not comparing me to other AI creators. They're comparing me to a traditional production. Actors, locations, a crew, and potentially weeks of work. And instead of one ad, they're getting a whole set to test. And since they need it by tomorrow, I'm adding a rush fee. If a client needs it fast, then you should always be charging for

[5:11] it. I added 50% which is another 500 dollars before production has even started. That brings the total to 1500 dollars. You might be wondering, why would they pay 500 dollars more on top of what I'm already asking for? And the answer is simple, because they need a high-end commercial in just 24 hours. And when a business needs an ad that urgently, speed is what they're really paying for. And most people out there

[5:36] will gladly pay extra to get it on time. [music] Now, let's see if they accept. I'm sending the full breakdown. The main ad, a vertical version, and three alternative openings, all delivered within 24 hours with two rounds of revisions included. And yes, [music] of course they know it's made with AI because it says so right in the breakdown. They said yes, $1,500 for the

[6:00] full package delivered within 24 hours. That's 19 minutes gone. We only have 41 minutes left for the assets, the scenes, the edit, and getting paid. [music] So here's the idea. To make it look expensive, we're not making a product video about the TV's features. Instead, we're actually telling the story of a family. Each person gets the remote and uses it to take control of their own moment. And then we see all of them on

[6:28] the same couch watching the same TV. Now, we're going with a story like this because every TV lists has the same features. 4K, smart apps, a thin frame. Ultimately, specs don't make anybody feel anything and nobody really remembers them once the ad ends. On the other hand, a story lets viewers see themselves on that couch. That's why Apple, Coca-Cola, and IKEA rarely lead

[6:55] with technical features or specs. Instead, they sell the feeling. People buy on emotion and they justify it with the specs later. Now, each scene works out on its own, so any of them can open the ad. That gives me three openings: [music] Grandma in the van, the office, or Grandpa at the carousel. The product is the TV, so the TV obviously gets the

[7:18] final scene. The next step is to create assets. [music] We need them to make sure that the faces, locations, and items stay the same from scene to scene and don't turn into slop. The production comes down to three simple steps: assets, prompts, and generation. I used Hexel's image generator for the references, then I plug those into its video generator for the scenes. Now, all in all, this commercial needs 16 reference images, six character sheets, six locations, two props, and two more faces. The woman in the subway and a child version of the grandpa. Almost all of them are made in Soul Cinema. If grandma's face changes

[8:01] between the crosswalk and the couch, then the client's immediately going to notice. So, everyone needs to have a character sheet. I use a Claude skill that I built for this workflow, which actually runs on Opus 5.5. I describe how I see the person, and then I let Claude write the whole prompt for me. Every character sheet has three views, front, back, and face up close. The back matters because grandma's going to walk

[8:26] away from us as she crosses the street. So, I want to make sure that the AI model has that information to keep everything consistent. It's also super important to keep the clothes and proportions the same to make sure that nothing morphs or turns into slop. Now, notice the lighting on this sheet. There's no studio lights, just window light and an overhead lamp that feels like real skin texture. And that's on purpose. The model doesn't only take her face from this reference. It also takes her skin, her clothes, and even the light that's on her. So, if I give it a glossy studio portrait, then she's going

[9:02] to look like a studio photo pasted into the New York streets. But, if the reference already looks like a normal photo, then she's actually going to look like she belongs in the setting that we placed her. I pick the most consistent sheet, and then I repeat that for the grandpa, the office guy, his boss, and [music] the two school boys. Now, one more thing about these characters, their reactions have to stay small. For example, when she misses her train on the subway, the grandma doesn't throw up her arms or open her mouth in shock. She simply blinks slowly and her shoulders drop just a little bit. And there's a

[9:37] reason for that. The remote is always doing something impossible. It stops time and rewinds trains. So, if people overreact on top of that, then the whole ad starts to feel like a cartoon. But, if they react the way that real people would, then the magic actually feels believable. So, once we have our assets ready, we now need to prepare the locations. For six scenes, we obviously need six locations. We use the same workflow for that. We describe the set to Claude, and then we take that prompt into Soul Cinema. Now, here's a super important detail. I generate every location empty [music] with no people and nothing that moves in it. That way,

[10:17] the carousel has no kids riding it and the metro station has no train. And this is the reason why. In the metro scene, the train has to pull away and then rewind back into the station. But, if the train is already in the reference image, then the model tries to keep it there. And the reference ends up

[10:34] fighting the action that I'm asking for. So, the image only gives the model the place itself. The platform, the tracks, and the light. Everything else which is moving, like the train, the riders, and people, come from the video prompt itself. Now, there's one more thing that makes all of these locations look like they belong in the same commercial. Four of them use the exact same style block, the same camera, a full frame anamorphic lens at 40 mm, and that overall background filter. I write that block once and I paste it into every single location prompt. And here's why that matters. If you describe the camera differently each time, then you're going

[11:16] to get six images that look great on their own, but feel like six different ads. So, you lose that consistency. The style block makes sure that everything looks consistent. [music] Then, I add in this to the scene prompt. Use the reference image for geography, materials, and lighting, not for framing. Without it, the model copies the reference's camera angle, and every shot ends up looking like the [music] same photo. The remote shows up in all six scenes and in a different hand every single time. So, for that reason, it gets its own reference, just like the actors. For this clean product shot on white, I switch to C Dream 5.0 Pro. And

[11:56] Claude writes out every button in the prompt, so the model knows exactly what to draw. And the layout of the remote really matters here, because every button has a job in the story. One freezes time, another rewinds it, and plus, for example, multiplies whatever you point at it. In a minute, I'm going to show you what happened when one of my scene prompts asked for a button which this remote doesn't even have. Now, this is the client's remote. They sell the TV, and I designed this one just for the story. So, before I build six scenes around it, I'm going to send it to them

[12:31] for approval. And now, it's approved. Getting that approval was super important, because if they'd said no after all six scenes were done, then I'd have to rebuild every single one. The last prop is the schoolboy's lollipop. And for that, I go back to Soul Cinema. In his scene, he hits plus, and the one lollipop turns into dozens, so every

[12:52] copy has to look exactly the same. That's why I designed the prop and its trick together instead of separately. Now, we move on to generating the scenes themselves. Let's start with Grandma. She stops a speeding van with a remote. This prompt worked on the first try, so let's open it up and see why.

[13:12] Hey! [music] In Claude, I attach her character sheet and the street with the remote, and then I tell it what each one is. The skilled in tags them in every single shot, so the video model knows exactly who and what goes where. Each scene is a single generation, but with cuts inside of it.

[13:35] So, it ends up being like a mini edit. [music] Grandma's going to get five shots in 10 seconds, the office gets nine and 15. So, really what I'm writing here is a shooting script. It's going to start the scene in one paragraph. Then, come down to the references and then a map of the street, where grandma stands, which way the van comes from, and where the people around her actually are. [music] They have to stay in the same spots every time we cut. Otherwise, the audience is busy figuring out the street

[14:06] when they would be watching the trick. Then, every shot gets its own camera position and lens. The first one is a 600 mm lens from 70 m away. That squashes the distance, so the van feels like it's right on top of her. Then, I write the action down to the second. At 3 seconds, the man needs to yell, "Hey." [music] At 5.2, she presses okay, and on the very next frame, the van freezes. All of that detail pays off when the video comes back because I can simply check it against the script, moment by moment, instead of going back and forth re-watching every single prompt. The

[14:43] locks cover whatever could break the scene. [music] Now, for grandma, freezing means movement stops. The street doesn't turn gray, people don't become statues, and the lighting doesn't stay the same. Freeze time leaves a lot open to interpretation. I always want a normal-looking street where something impossible happens, so I'm going to describe exactly what stays normal. The metro, on the other hand, needs a different rule. She moves normally while the world rewinds, stopping when the train doors open. The carousel needs its geography locked down, where grandpa stands at the south of the carousel, always facing north. That exact spot is going to matter in just a minute. I also

[15:25] need to specify how the hands actually hold the remote. In a product ad, that grip is a key part of the product shot. I'm also going to add two rules about the references themselves. Now, they sound obvious, but the model doesn't know them unless you spell them out. First, character sheets show three [music] views: front, back, and close-up. So, I need to tell the model these three views are one person. And second, the remote reference is shot on a white background. So, I tell the model to take out only the remote and not the white background around it. The office

[15:59] and subway were used two style blocks. While one says no people, the other says no cuts. Those conflict with our scenes. So, I explicitly switch off those restrictions and say which blocks take priority. The original style text stays intact, but its job is clearly defined. I want its look without it accidentally importing instructions from a different scene. We're almost done with the prompt, so at the very bottom, I repeat only the things that would ruin the scene if they changed. Grandma moves and the van doesn't. The remote is in her

[16:30] right hand and the cane is in her left. And there's one van and one grandma. Now, I don't need every rule in every scene. I simply need the ones that protect the particular idea. And with that, the prompt is done. Let's watch the result and see what actually held up. 27 minutes left. Now, those references and prompts have to turn into something that can actually deliver. And here's a pro tip. C Dance 2.5 just got a draft mode. You can generate everything that you need in 480p first at a quarter of the price, and then bring the exact takes that you're going to keep up to

[17:05] 1080p. That way, you're not burning credits on takes that you're never going to use. And right now, it's one of the cheapest ways to generate any kind of high-quality video. So, the loop is super simple. [music] The Claude skill writes the prompts, C Dance 2.5 generates, I review, and then we fix whatever breaks. I attach the references in the same order every single time because the prompt is going to call them by number. All 17 video generations are C Dance 2.5. I fit them

[17:34] into 15 minutes by overlapping the work. While one scene renders, Claude is writing the next prompt, and I'm reviewing whatever just finished. Now, half the scenes never needed a prompt change. Grandma, the metro, and the school just needed me to pick the best take. The office, grandpa, and the ending needed actual fixes. And that distinction matters. Another run fixes bad [music] luck, but it can't fix a bad

[18:00] prompt. Okay, the office scene. [music] The boss is yelling at his employee. The employee freezes him with the remote, draws a mustache on his face, sits back down, and starts the time again. The boss continues yelling, but he now has a mustache. But remember that remote that the client approved? It doesn't have a pause and play button. My first prompt used both. So, I had to go back to Claude and kept everything, and only fixed the remote. The new version gets the actual reference image and the exact button layout. The employee presses okay now to stop time, and he clicks it again to start it. The scene still does the

[18:41] same thing, but this time with the prop the client actually needs. So, if you've already approved the product's appearance, then make sure that the next prompt doesn't quietly change everything. Next, the grandpa. Now, this is my favorite fix in the whole commercial. He stops by a carousel full of kids, feels a little bit sad, and then presses rewind. He then becomes a 4-year-old laughing as he rides a carousel. Now, it looks nice, but who's that kid? You couldn't instantly understand that it's the grandpa. I take one shot and the boy turns precisely

[19:16] where the grandfather had been standing. The same fence, same direction, and the same remote in his small fist. He slips under the rail and then gets on the horse. At the same location, using the same remote, it's now your brain that makes the connection between the grandpa and the boy as the same person. When a character changes, the change has to happen to something while other things remain the same. I also made the kid bigger in the frame. You can't care about what just happened until you recognize him. And then I find a contradiction. One line puts the camera south of him looking north. Another puts

[19:53] it north looking south. The model can't follow both, so it picks just one. You always want to make sure you're reading your prompts for contradictions. [music] The model won't tell you, it'll actually just take a guess. The transformation itself, you actually never see it. You see a click, a reverse swoosh, a cut, and then he's a kid. There's no morph, no glow, because you want to hide the impossible part in the cut, and the viewer's brain should just be doing the rest. And lastly, the family on the couch. This is where all of those separate scenes have to feel like one commercial coming together, and where

[20:29] the client's TV gets to have its moment. In version one, the camera pushes towards the TV. Now, it works, but all we really get is a screen that gets bigger. I want the room and the family to open up around it. So, version two starts close to the TV and then pulls back. [music] But another line in the prompt still says the camera can drift or push. I have to fix that, too, and then repeat

[20:55] the new direction in positive locks. Away from the TV, never towards it. One leftover word can undo the change that you just made. When you change a rule, check that the whole prompt follows the new one and doesn't keep anything from the old [music] one. The camera feels better now. The family feels like five people waiting for a photo. So, in

[21:15] version three, they actually interact. Grandpa says something, the young woman laughs and nudges him, the boy turns around, and the popcorn gets passed. There's no scripted lines and just indistinct conversation, laughter, and popcorn crunching. [music]

[21:45] The first shots all move a little and the cuts land in the middle of those movements. Then everything settles into one still wide shot. The family, the room, and the TV. [music] All the life comes from the people. That's the shot for your screen, your [music] rules. The subway uses the same ideas grandpa. A

[22:03] clear problem, one button, and a payoff. She misses the train, presses rewind, and the station runs backwards until the doors open right in front of her. In the school, one boy has a lollipop and his friend is too shy to ask. So, he taps plus and lollipops start pouring out all over the floor. The whole hallway stops and stares. Different tricks, same visual style. The same references, the same remote, and small, believable reactions all are placed in a world with people in it. The pigeons at the crosswalk and the janitor in the hallway helps those places feel alive. The

[22:42] generated sound is just what's in it. What kind of work is this? Voices, clicks, and the carousel's organ. Things like music, titles, and the client's logo go in during the edit. [music] We're 48 minutes in. Every scene is done and we have 12 minutes to turn them into the package that I sold. The cuts inside each scene are already in the generation, so I'm assembling the complete sequence. The crosswalk, the office, the carousel, the metro, the school, and lastly, the couch. And the TV in the lodge plays the same scene that you just watched. The whole commercial folds back into one living room. Now, we have three openings. Each

[23:23] scene already has a problem, a click, and a payoff. [music] So, really I can lead with grandma in the van, the office, or grandpa at the carousel. Everything afterwards always stays the same. That's why I always plan the variations before generating. The client's going to get three openings to test, and I don't have to build three more scenes at minute 48. So, for the vertical version, I'll use reframe in Hicksfield. You always want to check the results on a phone-shaped frame. Faces, hands, the remote, and the TV all need to stay readable. So, we've got 5

[23:57] minutes left, and it's time to deliver. Main ad, vertical version, three openings. [music] Everything that we agreed upon delivered in one single go. $1,500 with a clock stopping at minute 58. [music] Now, the number that I promised you at the start. What did it cost to make? 16 images and 17 video generations with a total generation cost of $250. [music] That leaves 1,250 after generation cost, but before any platform fees or other

[24:28] expenses. [music] Now, keep in mind, 500 of those dollars was the rush fee. I quoted it before starting, they agreed, and ultimately, I delivered inside of the deadline. That's why the scope and the price had to be super clear in those first few messages. And after this job, I still have every asset. The prompts, the locations, and the remote. [music] So, if they need a holiday ad next month, then I already have the cast and the world. [music] Now, there's obviously still a few new scenes to make, but I don't have to rebuild everything that you just watched. So, when you deliver, tell them what else

[25:07] this cast can do. Give them a reason to come back for the next campaign. And that's the whole job. It ultimately comes down to five simple steps. Go out there and find the client, agree with them on the package, build your scenes, fix what's breaking, and deliver what you promised. And by the way, every prompt, every reference, the cloud skill, and every step that we took are in the blog post linked below. So, if you want to see the commercial that I used in that first proposal, it's right here as well. I'll see you guys in the next

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt