The most reliable prompt structure for AI video generation is a fixed 9-element assembly order: camera spec, lens and aspect ratio, lighting source, palette, composition, atmosphere, mood register, film/DP attribution, and negative prompt — physical mechanics first, exclusions last. One documented production held this exact order across every frame of a 21+-scene project.
Describe your subject and action in one plain sentence first, then assemble the nine stylistic elements around it in this order:
- Camera spec — the camera body or format behavior you want (e.g. 35mm digital, locked-off or slow push-in).
- Lens and aspect ratio — lens type carries real visual consequences: spherical lenses give circular bokeh and no horizontal flares, anamorphic gives the opposite. State the ratio in your film's delivery format.
- Lighting source — name the actual source, not an adjective. "Warm yellow from the lamps only, like all the refs" produces more accurate results than generic "warm lighting." Ratios work too: one production encoded an 85:15 dark-to-light ratio as reusable prompt language.
- Palette — name the color mode separately from lighting. Encoding tonal modes like "split-toned amber and emerald" with exact hex values makes the palette reproducible shot to shot.
- Composition — framing logic: what's centered, what's withheld, foreground obstructions, negative space.
- Atmosphere — the physical air of the shot: haze, rain, smoke, neon reflections.
- Mood register — the emotional temperature, kept distinct from atmosphere: dread, longing, restraint.
- Film or DP attribution — anchor everything above to a stylistic reference ("in the visual language of Wong Kar-wai"). It sits late in the sequence so the model interprets your specifics through that lens rather than letting the reference override them.
- Negative prompt — a mandatory closing element, not an optional add-on. State what the shot must never be. One animated production locked this as "This MUST look and feel like Arcane animation — not live action, not photorealistic" and prefixed that block to every one of its 164 generated clips, which is what held the style across the whole episode.
The order has a logic you can lean on when adapting it: physical camera mechanics come first because they constrain everything else, color sits in the middle, emotional register follows color, stylistic attribution interprets the whole stack, and exclusions close the prompt to fence off drift.
A full worked example, all nine elements applied to one shot: "A woman waits alone in a hotel corridor. 35mm camera, slow push-in / spherical lens, in your film's aspect ratio / lit only by warm practical wall lamps / split-toned amber and emerald palette / framed through the doorway, subject off-center, foreground wall obscuring half the frame / humid night haze, wet reflective floor / restrained longing / in the visual language of Wong Kar-wai / negative: no daylight, no handheld shake, no modern signage."
The structure is model-agnostic — it works across Veo, Kling, and Seedance 2.0, all of which are available inside invideo, where the invideo agent routes each shot to the right model. If you're writing many shots, you can also load these rules into the invideo agent once as a context document so the same assembly order holds on every prompt without retyping it.
Watch some of these to see what works for you:
This MUST look and feel like Arcane animation — not live action, not photorealistic. Every surface has hand-painted brushstroke texture. Every element in frame must feel painterly and handcrafted like a moving Arcane frame.
— invideo's creative team, from a documented animated production's locked style block