How do you write better AI video prompts using filmmaking and cinematography knowledge?

Write AI video prompts as a fixed cinematography stack: camera spec, lens and aspect ratio, lighting source, palette, composition, atmosphere, mood register, film/DP attribution, and a negative prompt — phrased in precise on-set language, not generic adjectives. One documented short film held that exact 9-element assembly order across every frame to keep its visual style consistent.

Assemble every prompt in the same order, every time. A documented production used a fixed 9-element sequence — camera spec, lens and aspect ratio, lighting source, palette, composition, atmosphere, mood register, film/DP attribution, negative prompt — and held it across every frame of the film. A fixed order works because nothing gets dropped under deadline: each shot prompt answers the same nine questions a cinematographer would answer on set. A stricter variant of the same idea outputs 12 parameters per shot, adding shot length, emotional register, blocking, and a revision prompt.

Within that stack, name the light source, not the mood. "Warm lighting" is a guess; "warm yellow from the lamps only, like all the refs" is a lighting plan — and source-specific corrections like that are documented to produce more accurate results than generic descriptors. The same precision applies to ratios and lenses: one production encoded James Wan's lighting grammar as an 85:15 dark-to-light ratio, and corrected a lens error mid-project because spherical glass means circular bokeh and no horizontal flares — Wan shoots 35mm spherical at 2.40:1 hard matte, widescreen by extraction, not optics. Wrong lens vocabulary propagates wrong flares, wrong bokeh, and wrong distortion into every generation, so verify technical claims before you lock them into your prompt language.

Define palette as data, not vibes. Encoding a filmmaker's color philosophy as named tonal modes — "Mode A — Split-toned amber and emerald" — with exact hex values gives you reproducible palette control across shots, where "moody teal-orange" drifts generation to generation.

End every prompt with a negative prompt that states what the shot must never be. One animated production's style block read: "This MUST look and feel like Arcane animation — not live action, not photorealistic. Every surface has hand-painted brushstroke texture." Explicit prohibitions are what prevent style drift; telling the model what to leave out matters as much as what to include.

For motion and coverage, write directorial intent instead of parameter soup. A direction like "I want to stay on the feral guy when we run this scene. No back and forth cutting. We hold on him right up till he lunges" specifies a held shot, the cutting pattern, and the action trigger in one sentence — exactly how a director briefs a DOP. If you have on-set experience, it transfers directly: shot-size, blocking, and coverage vocabulary is the prompt language these models respond to.

Once your cinematography stack works, codify it once instead of re-typing it per shot — the invideo agent can hold a full visual-language document in persistent context (one production codified a director's complete system into 14 sections, from camera and lighting to prompt templates and negative prompts) so element nine of shot 40 matches element nine of shot 1.

Watch some of these to see what works for you:

Pretty much exactly like how I would talk to my DOP on set or how I would talk to my DA on set.

— invideo's creative team, on directing AI video with on-set language