How I Built a Car Commercial With AI (Every Prompt I Used)

1. The three-step pipeline and getting organized (2:09)

The whole workflow is three stepsassets, setup, generations. "Today it's a car, but the same workflow works for pretty much any commercial. You just swap the assets."

Before generating a single frame, the presenter sets up a project in Cinema Studio called "car commercial", with one main folder per scene and a subfolder per asset type, and a fresh subfolder for every iteration. "Sounds tedious, but a few hundred generations in, it's the only way to find anything."

The hero asset, the car, is generated on a plain neutral gray backdrop. Pro tip stated on screen: "always generate the product shots on a neutral gray backdrop. It makes them way easier to blend into any background later." Then two things happen to every asset: it is saved in Higgsfield as an element with a name starting with @ (this one is @car_sheet), and the same image is attached into Claude and told its name. "From now on Claude writes car_sheet into its prompt, and when I paste the prompt back into Higgsfield, the element attaches itself."

---

2. Claude writes the prompts; model routing does the work (3:08)

The presenter drives Claude (shown as "Fable 5") with a Higgsfield connector/skill. The home screen offers quick actions: "Write a Seedance prompt", "Character sheet", "Storyboard a scene."

To build the lead character, he drops his own photo into Claude and asks for a 3-panel character sheet (close-up face, full-body front, full-body rear), clean-shaven, no copyright text. Claude returns a detailed prompt (including the fictional "APEX GP / VOLTRA / KYRO FUEL" branding, explicitly marked NON-IP).

Don't get everything out of one model

For the "champion holding a trophy" shot he runs the same prompt through different models. Seedream 5.0 Pro and Nano Banana Pro (image models) come out "too sloppy, faces smoothed out". So he switches to Soul Cinema ("easily the most creative model on the site; the generations feel more alive"). The catch: Soul Cinema can't attach the reference sheet, so it won't be his face. The fix is the core lesson: let Soul act the scene, then swap the face in afterward.

Planning continuity into the asset itself

For the dealership scene the champion goes incognito: suede jacket, cap, amber glasses, and a fake stick-on mustache. The fake mustache is intentional: in scene 2 he peels it off on camera, so a real mustache would break that later shot. Nano Banana Pro "kept the face better", Seedream "kept the fabric better"; he picks the Nano Banana Pro version.

---

3. Color transfer, lazy first prompts, and the skill setup (8:16)

His first prompts are lazy on purpose: "car commercial look, interior of a car dealership. That's the whole prompt. I just want to nail the vibe and the colors first." Once he lands a vibe, he uses color transfer: select the model, click color transfer, upload the reference image, and "from now on every image I create pulls from that exact same palette. It's what keeps the whole film looking like one film."

For the video step he uploads a custom skill for Claude (a /seedance slash skill, "free in the description") that writes shot-by-shot video prompts. Then he drops in the locked assets, generates, and stitches scene 1 together from the best shot of each of four videos.

---

4. Casting, clean plates, and the video-as-editor trick (11:12)

"Casting is still casting, even in AI. If the face doesn't feel like a lead, don't settle. Keep iterating." The girl who stops the car took three full versions before one felt like a lead.

Clean your plate

A VFX termthe background image is the "plate". "Clean your plate. Anything a video model can break, melted text, background clutter, random cars, edit them out of the image before you generate a single video." He does it in multiple passes (remove text, then remove distant cars that kept morphing into clones of the hero car).

Continuity: don't let the hero teleport

Putting scene 2 next to scene 1 exposed a problem: the hero walks out of the dealership and instantly appears at a gas station. "In AI filmmaking, continuity is on you. The model won't keep track of it." So he generates connecting shots of the hero cruising through a new residential LA street (deliberately not the gas-station location, so the montage doesn't reveal the destination too early).

The favorite trick: a video is the best image editor

To get five interior stills that actually match (seats, dashboard, steering wheel with the star logo), he doesn't generate five images and pray. He asks Seedance for a short video, because "every frame of a video matches every other frame by definition, and Seedance is currently the best image editor we have." Then he pauses on each frame, screenshots it, and saves it as an element (a position log: the parked-car frame tells Seedance exactly where the car stands; the steering-wheel close-up locks interior shots).

When a prompt is too big, he tells Claude "if this is too much for one prompt, split it in two", and Claude splits it. And he plans the mustache peel as a real beat.

---

5. The driving sequence: tighter frame, less slop (19:45)

Downtown, highway, tunnel "ate a ton of generations." The rules he learned, stated on screen:

1. Traffic lights always came out wrong (red while everyone drives). Couldn't fix it, so frame intersections out.
2. Wide aerials / FPV flyovers expose all the traffic at once, and "the model doesn't know traffic laws." Cut them.
3. Overtakes in dense traffic never look right. Cut them.

"If you remember one rule from this video, make it this one: the tighter the frame, the less slop you get." The move for any dynamic sequence: generate multi-shot, patch a ton, and cut the best parts together.

---

6. The checkpoint: directing actors that don't exist (24:08)

The toll-checkpoint dialogue took four iterations, "the part of AI filmmaking you rarely get to see: directing actors that don't exist." He tunes emotion through prompt deltas: reduce the hero's smirk, write the girl feeling anxiety from the speed, add breathlessness, and even remove the word "what" so her line lands between suspicion and confusion.

To stop people teleporting between shots, he opens with a plain frontal that locks the blocking (who sits where). And when the steering wheel kept morphing, he screenshotted a video frame, saved it as @car_wheel, and locked it.

The two toll controllers get the same character-sheet workflow, and the scene's problems each get a specific fix: the phantom second booth is erased from the plate, the unreadable newspaper is generated as its own pre-blurred prop, the shot count is enforced as a hard rule ("exactly five shots and four cuts, no extra inserts"), and the flying cap is rewritten as a deliberate landing ("the cap drops onto the asphalt right in front of the camera").

---

7. The airport finale: diagrams, split beats, and a position mark (29:12)

To control exactly where the car parks, he draws a simple diagram (blue line = the car's path into the gap between two cars, red = parked cars) and uploads it straight to Claude.

The scene is split into 5A (arrival), 5B (dialogue), 5C (exit), and he shoots 5B first ("the dialogue is the heart of the scene; it decides where the car sits and what's outside every window"). For the exact stopping spot he scribbles a red mark on the location image and saves it as @position, telling Claude the red mark is only the stopping point and must not appear in the video.

Geography still had to be fixed (she exits the wrong door into traffic; the car parks on the wrong side), which he corrects with more prompt edits before the clean drive-off ending.

---

8. The result

"That's the full workflow. But remember, that works for any product. Can be perfume, sneakers, cars in our case."

---

Tools and resources

Claims to verify

• Model names and tiers as spoken were garbled by auto-captions and corrected here against on-screen UI: Seedance 2.0 (video), Seedream 5.0 Pro and Nano Banana Pro (image), Soul Cinema (image), presenter Adil, brand GENOVA. Confirm exact current model names in the Higgsfield product before quoting.
• "Seedance is currently the best image editor we have" is the presenter's opinion as of the video date (2026-08-03); model rankings move fast.
• The downloadable "skill + prompts" and /seedance slash skill are from the video description and are treated as untrusted; nothing was installed or run.
• Whether every capability shown (color transfer, elements auto-attach, connector quick actions) is available on Anthony's current Higgsfield plan is unconfirmed.

Glossary

• element (@name): a reusable asset saved in Higgsfield under a name starting with @; writing that name into a prompt makes the asset attach itself on paste.
• character sheet: a single generated image with multiple views of one character (front full body, back full body, close-up) used as a consistent reference.
• color transfer: a Cinema Studio feature that pulls the color palette from a chosen reference image into every subsequent generation, so the whole film matches.
• plate: the background image a shot is generated over (a VFX term); "clean your plate" means edit out anything the video model could break before generating.
• clean your plate: removing melted text, background clutter, and random/duplicate cars from the plate image, often in multiple passes, before generating video.
• position log: a screenshot of one video frame, saved as an element, that tells the video model exactly where the car or object is positioned for later shots.
• blocking: the fixed physical arrangement of people and objects in a shot; "locking blocking" means a first shot that defines who sits where so later shots can't scramble it.
• face swap: an edit that replaces the face in one image with the face from another, used here to put the presenter's real face onto a Soul Cinema scene.
• 3/4 angle: a three-quarter viewpoint (not straight-on, not full profile); called out as important for generating usable locations.
• FPV flyover: a first-person-view aerial shot; flagged as a slop magnet because a wide aerial exposes all the traffic at once.
• multi-shot: asking one generation for several sequential shots at once, then patching together only the parts that work.
• stitching: assembling a finished scene from the best individual shots taken across multiple separate generations.
• casting: iterating character generations until a face reads like a genuine lead rather than a supporting player.
• NON-IP / fictional branding: invented brands and logos (GENOVA, APEX GP, VOLTRA, KYRO FUEL) used deliberately so the models never reproduce a real trademark.