Free-text prompts drift. For campaign work you want the same shot grammar every time. Structured, JSON-style prompting is how video models are told exactly what to do.
The shape of a structured prompt
Break the shot into named fields: subject, action, camera, lighting, style, duration. The model treats each field as a constraint instead of a suggestion.
{
"subject": "matte-black wireless earbuds on a concrete plinth",
"action": "slow 90-degree turntable rotation",
"camera": "macro, slow push-in, shallow depth of field",
"lighting": "single softbox key from top-left, cool rim light",
"style": "35mm product film, high contrast, no text",
"duration": "5s"
}Why this matters for ads
Variants stay on-brand: change one field, keep nine locked.
A creative brief becomes a template your whole team can fill in.
Test cells differ by exactly one variable: real experiments, not vibes.
Field-by-field tips
camera: name the move ("slow push-in", "handheld orbit"). One move per shot.
lighting: describe the source, not the mood ("single window, late afternoon").
style: anchor to a medium ("35mm", "smartphone vertical") rather than an artist.
A prompt template is an asset. Version it like code, because it is code.