AI Car Commercial · Claude, Seedance 2.0 and Higgsfield

Quick Summary

• Credit: Higgsfield AI published the embedded tutorial, presented by Adil. The concept, examples, finished commercial and original workflow belong to him and the Higgsfield team.
• The process has three steps: building reusable assets, setting up references and rules in Claude, and generating each scene in chunks small enough to maintain control.
• Consistency doesn't just come from the prompt. It comes from character sheets, a car isolated on a gray background, clean locations, approved frames saved as Elements and specific references for face, steering wheel, interior and position.
• Adil uses Soul Cinema for presence and acting, Seedream Pro 5 or Nano Banana Pro for identity, clothing and editing, and Seedance 2.0 for video. The principle is to specialize models instead of asking just one for everything.
• The five scenes are montages of various outputs. When a generation only gets the first and last shots right, those fragments survive and the rest are discarded.
• The most instructive errors include warped traffic, characters that teleport, steering wheels that change, unstable logos, wrong doors and impossible positions for the car to exit. Each error generates a new reference, a tighter framing or the division of the action into two generations.
• The official Higgsfield guide publishes the exact project prompts. The pack below is an original and organized adaptation for study and reuse, not a complete copy of this material.
• The video does not provide total cost, total number of generations, latency, project duration or success rate. It also does not replace the verification of rights to brands, faces, music and commercial outputs.

Credit, sources and limits of analysis

The video was published by Higgsfield AI and is presented by Adil. All credit for the ad concept, characters, experiences, failures shown and final solution belongs to the creator and the Higgsfield team.

This analysis cross-references the full transcript provided with the video, the official Cinematic Car Commercial · Full Breakdown page and public information from Cinema Studio. The search was updated on August 3, 2026.

The official page contains the complete original prompts and the skill file used in the tutorial. To respect authorship and make the material more reusable, the pack in this article rewrites the method as its own templates. Anyone who wants to compare with the exact texts should open the official guide to the sources.

The Short Answer: How This Ad Was Actually Made

It wasn't a prompt that spawned an entire film. Adil built a small production system. First he created the product, characters, clothes and locations. Then you saved the approved assets as named references. Finally, he asked Claude to transform scene descriptions into technical prompts for Seedance 2.0.

Each scene was treated as a separate problem. The generation could only provide one good movement, one reaction shot or two seconds of driving. The montage brought these fragments together and hid the flaws.

That's the tutorial's main takeaway: Generative AI makes production accessible, but direction, continuity, and editing continue to decide whether the result looks like an ad or just a model demo.

The story in five scenes

The ad starts at a dealership. A pilot receives a car and appears disguised to avoid recognition. Then, a woman stops the car with her hands and embarks on a journey that passes through a residential area, urban center, highway and tunnel.

At a checkpoint, the couple tries to find change while two controllers recognize the pilot. The story ends at the airport with a skid to park, a goodbye and leaving the car.

Narrative is useful because it tests almost every difficult AI video problem: dialogue, acting, two characters in the same interior, traffic, physical contact with an object, clothing continuity, identity, geography, vehicle movement, and visual gag.

The three-step production architecture

Step 1, assets. The project creates the car, pilot, disguise, consultant, female protagonist, controllers and locations. These materials are treated as a virtual art department.

Step 2, setup. Each approved image receives a name and becomes an Element in Cinema Studio. The same name is explained to Claude. Thus, the prompt text can ask for @car, @hero_clean or @airport and maintain the relationship between instruction and file.

Step 3, generations. Adil describes the intention in simple language, receives a structured prompt from the skill and generates it in Seedance 2.0. If the result fails, correct the plate, add a reference, reduce the framing, simplify the choreography or separate the sequence.

Folders, versions and traceability

Adil creates a main folder for each scene, an assets subfolder, and a new subfolder for each iteration. This discipline seems banal, but it prevents a corrected frame, an old plate and the approved take from mixing.

A practical structure is 01_assets, 02_lookdev, 03_scene_01, and so on. Within each scene, use 00_refs, 01_generation, 02_selects, 03_fixes, and 04_final. Also save the prompt, template, duration, proportion and rejection reason.

Without this record, the team does not know which reference produced the good steering wheel, how much the scene cost or how to repeat the aesthetics in a new product.

Why does the car start on a gray background

The car is generated as a clean asset, with multiple angles and a neutral gray background. The goal is not to create a pretty image. It is to give the model an unequivocal definition of shape, paintwork, rims, headlights and proportion.

A complex background can contaminate future compositions. The model starts to associate the product with that location, invents incompatible reflections or transports objects to the new scene.

The same logic applies to sneakers, perfume, packaging or electronic devices. First build a controlled product reference. Only then place the object in the final art direction.

Character sheets: face, front and back

The pilot sheet brings together a facial close-up, the complete body from the front and the clothing seen from behind. This gives the model information that an isolated photograph does not provide when the camera changes angle.

The costume gets its own set: brown suede jacket, light shirt, jeans, cap, amber glasses and a deliberately fake fake mustache. This last detail was designed for the future action in which the mustache comes off.

The consultant and controllers also receive separate sheets. When a full-body view shows a worse face, Adil replaces that area in an editor rather than spending generations hoping to fix everything at once.

Soul Cinema, Seedream Pro 5 and Nano Banana Pro

Soul Cinema is used for images that need presence, agency, human imperfection and a sense of a living scene. It is useful for casting, trophy photography and atmosphere.

Seedream Pro 5 and Nano Banana Pro come in when the priority is identity, clothing, face swap or localized editing. The most natural result may have the wrong face; the solution is to preserve the performance and replace only the identity.

The lesson is operational: assign each model the task at which it is most predictable. Requesting casting, styling, composition, identity, lighting and movement in a single step increases the error surface.

Look development and color transfer

The first showroom prompt is intentionally open-ended. It serves to look for contrast, atmosphere, glass, materials and palette, not to deliver a final scene.

When he finds an image with the right color, Adil uses color transfer to bring the same visual intention to the remaining locations. This transforms a collection of plates into a coherent universe.

It is best to separate look from content. The plate defines geography and objects; the look reference defines color, contrast, light and texture. Mixing the two functions makes it more difficult to understand what to fix.

Elements as visual memory of the project

Each approved asset is saved as an Element. The name needs to indicate function, not just appearance: @hero_clean, @hero_disguise, @lead, @car_exterior, @car_interior_front, @wheel_lock and @airport_stop_mark.

When a logo or steering wheel changes in an output, the best frame is corrected, captured and promoted to the new reference. The problem no longer depends only on language and now has a visual restriction.

Element works as external memory. On long projects, this memory is more reliable than repeating descriptions like the same car or the same woman across dozens of prompts.

What Claude's skill adds

The skill does not generate video. It converts a creative description into a consistent specification for Seedance. References are declared, the technical block fixes duration and format, the shot list defines temporal progression, the audio describes dialogue and effects, and constraints protect important elements.

This layer reduces writing variations between scenes and allows you to reuse a production grammar. The person continues to decide the intention, but does not need to remember all the technical instructions in each prompt.

The official file can be downloaded from the tutorial page. Before running it, read the content, confirm what actions you can take, and remove integrations or instructions that don't make sense for your environment.

Anatomy of a Robust Seedance Prompt

Start with REFERENCES: say what each image controls. One reference can fix identity, another the car, another the location and another the exact position of the steering wheel.

Continue with TECHNICAL: duration, aspect ratio, intended frame rate, camera language and exact number of shots. In SHOT LIST, describe actions in order and use explicit cuts.

Close with AUDIO and LOCKS. Write dialogue without adjectives within lines, describe environment and effects separately, and block out identity, clothing, logos, direction of movement, number of characters and geography.

Scene 1: The dealership is a montage

The opening needs to feature a showroom, car, consultant and undercover pilot. A single generation would have too many goals. The tutorial tries out several versions and retains the useful shots.

The end result appears continuous because the plate and Elements maintain materials, color and characters. In practice, each fragment solves a function: establishing the space, revealing the car, handing over the keys or showing the reaction.

This scene introduces a rule that is repeated throughout the film: generate editorial coverage. A long shot, a medium shot, a detail and a reaction give more freedom than a long take that needs to be perfect from start to finish.

Scene 2: stop the car and clean the plate

The female protagonist goes through several versions until the casting matches the energy of the story. The location of the gas pump is prepared at a three-quarter angle to allow approach and contact with the car.

Adil removes text, clutter and expendable vehicles before animation. Ambiguous elements are points where the model can deform the image, create unsolicited movement or compete with the main automobile.

The moment when the woman stops the car with her hands was built on the best generations. Contact between body and vehicle is a high-risk event, so editing cuts before errors accumulate.

Scene 2: interior, statics and dialogue

To prevent each angle from creating a different cabin, Adil generates a static video with several views of the same interior. Then capture the most coherent frames and save them as Elements.

One image fixes the position of the characters and another fixes the steering wheel. The dialogue sequence is divided into four shots, which reduces the burden of synchronization, acting, continuity and camera movement in each generation.

When the description becomes too dense, the instruction to Claude is simple: if it's too much for a prompt, divide it into two. Splitting is a steering technique, not a workflow failure.

Scene 2: Remove the disguise and activate Sport Mode

The mustache and glasses are removed at separate times and, when necessary, slowed down or prolonged during assembly. From then on, prompts will stop using @hero_disguise and start using @hero_clean.

Starting in Sport Mode combines details of the button, box, pedal and passenger reaction coming from different outputs. Asking for four perfect inserts in a single generation would be less manageable.

Continuity is treated as state. Before the gesture, the pilot is disguised. After the gesture, it's clean. This state must also change in the reference library and not just in the text.

Scene 3: city, traffic and the tight framing rule

The urban center plate has four lanes, glass towers, vegetation and clean space. The gray sky is replaced and the film palette is transferred before inserting the car.

Very open aerial shots, FPV movements and dense overtaking expose the limits of the model. Traffic lights, vehicles and lanes begin to transform. The practical answer is to reduce the field of view and the amount of traffic.

Adil's rule can be summarized as follows: the tighter the frame, the smaller the margin for slop. A short lateral tracking can sell speed with much less risk than a traffic choreography seen from above.

Scene 3: cabin, highway and tunnel

The first shot of the cabin is frontal and simple. It serves to establish where each person is sitting before switching between other angles. Without this anchor, characters may switch places or appear to teleport.

On the highway, the FPV attempt in traffic fails and a match cut between steering wheel and wheel is not understood. A detail with sudden entry, wind and slight camera recoil works best, so the sequence is reconstructed around that moment.

The tunnel is a beauty pass. Light lines run through the bodywork and sell finish, speed and production value without requiring complex geography.

Scene 4: checkpoint, dialogue and a failure turned into a gag

The dialogue inside the car took four iterations. The initial acting was weak, so emotion and intention were rewritten line by line. The logo and steering wheel also changed until they were fixed by their own Elements.

In the control booth, teleporters, a second ghost booth and an unreadable newspaper appeared. The solution combines a clean location, separately created newspaper, five shots and four exact cuts, and the first shot attaching the controller to the window.

A cap that fell the wrong way became a controlled joke: the prompt now asks him to land on the asphalt in front of the camera. A recurring flaw can be absorbed by the staging when it serves the story.

Scene 5: airport, diagram and production order

The location is presented in three rooms. A diagram drawn over the top indicates the car's path and parked vehicles, but the prompt warns that lines and colors are just instructions and cannot appear in the video.

The first version tries arrival, skidding, dialogue and exit at the same time. The result loses positions. Adil divides the sequence into 5A arrival, 5B dialogue and 5C departure, and produces the dialogue first because it fixes the position of the car and the geography of the window.

A red mark on the plate defines the exact stopping point. When he realizes that cars in front and behind are preventing a clean exit, he corrects the sign. The end result uses a 180-degree rotation parallel to the ride and the tailgate on the right side.

Batching and editing: the real engine of results

Only one of four generations of residential driving works, and even that contributes mainly to the foreground and backplane. The tutorial does not hide that selection is a central part of the process.

Make batches with one variable at a time. Maintain assets, duration and composition; change emotion, lens or movement. Then mark each output as approved, partially usable, recoverable problem, or rejected.

The editing must cut before the visual drift, use reaction shots to hide transitions and use sound to connect separately produced images. The final video is an editorial decision, not proof that a model generated everything in one take.

How to transfer the system to perfume, tennis or technology

Replace @car with a product-neutral asset and keep the architecture. For perfume, the scenes can be laboratory, human gesture, liquid macro, city travel and packshot. For sneakers, they can be unboxing, putting on, running, sole detail and arrival.

The useful question is not which prompt creates an ad. It is what references fix the product, person, space, state and movement in each beat of the narrative.

The system becomes reusable when the names, templates, continuity rules and approval criteria survive the product change.

Costs, attempts and time that the video does not report

The tutorial does not provide total cost, number of images, number of videos, queue length, editing time or percentage of rejected outputs. Without these values, it is not possible to estimate cost per final minute from the video.

Record credits spent per asset and per scene, plus hours of preparation, prompting, selection and editing. Divide the total cost only by the approved seconds that went into assembly.

Compare this value to an equivalent traditional production, but include legal proofreading, corrections, audio, color grading, subtitles and deliverables. AI changes the cost structure, it does not eliminate all phases.

Brands, faces and commercial use

Use fictitious brands during prototyping and obtain permission before representing a real manufacturer, logo, pilot, location, or person. The ability to reproduce an identity does not create a right of publicity.

Confirm each model and vendor's current terms for inputs, outputs, training, retention, compensation, and business use. Keep the version of the terms consulted on the project date.

Document the provenance of images, fonts, music, voices and sound effects. For paid campaigns, request legal review appropriate to country, platform, and license duration.

Verdict: it's not one-click, and that's an advantage

The tutorial is valuable because it shows the work between the idea and the output. AI advertising becomes compelling when references, direction, and editing reduce model freedom at critical points.

Claude helps formalize prompts, Seedance produces movement, and Higgsfield organizes generation and Elements. None of these pieces replace the decision to clean a plate, change the casting, fix a steering wheel or cut two frames before the error.

For those who want to repeat the method, the best first step is to create a short scene with a product, a person and a location. Measure attempts, save references and only then expand to five scenes.