How to Write AI Video Prompts That Actually Direct the Shot

Why most AI video prompts underperform

AI video models are trained to turn descriptions into moving images. When a prompt only names a subject — “a perfume bottle” — the model fills every other decision with its own defaults: a generic background, a random camera move, flat light. The result is technically a video, but rarely the shot you imagined.

Good prompts read like a shot list. They make the decisions a director and cinematographer would make, in plain visual language, and leave the model less to guess.

The anatomy of a strong video prompt

• Subject — what or who is on screen, with the one or two details that matter most.
• Environment — where it happens and what surrounds it (surfaces, weather, time of day).
• Composition — framing for the format: vertical 9:16 keeps the subject in the upper two-thirds; 16:9 benefits from foreground and background depth.
• Camera angle and lens — eye level vs. low hero angle; 24mm wide energy vs. 100mm macro detail.
• Camera movement — one clear move per shot: slow dolly-in, orbit, crane up, locked-off.
• Lighting — motivated key light, rim light, soft daylight, low-key studio; describe where the light comes from.
• Materials — glass, brushed metal, linen, wet stone. Materials tell the model how light should behave.
• Motion — what moves in the frame and how fast. Ask for physically plausible motion and no morphing.
• Timing — for 6–8 seconds: establish, main action, reveal. Beats help the model pace the shot.
• Audio — if the engine supports sound: ambience, music feel, foley.

Before and after

Before: “Create a luxury commercial for this perfume.”

After: “A luxury commercial for this perfume. Setting: minimal premium set of polished stone, dark glass and soft drapery. Composition: vertical 9:16, subject centered in the upper two-thirds. Camera: low three-quarter hero angle; 100mm macro for details; slow orbit and macro push-in. Lighting: controlled low-key studio light with specular highlights sculpting the bottle. Product fidelity: match the reference exactly — same shape, colors, packaging and label text.”

The idea didn’t change. The direction did.

Rules that prevent common failures

• One subject, one action, one camera move per clip. Stack complexity across scenes, not inside one shot.
• Describe motion, not a slideshow. “The camera slowly pushes in as steam rises” beats “beautiful coffee”.
• Avoid text in the frame unless the engine is known to render typography well; add captions in the edit instead.
• For product shots, always attach reference images and state fidelity rules explicitly.
• Keep negative instructions short and concrete (“no warping of the bottle”), not a long list of fears.
• Don’t ask for real celebrities, trademarks of other brands or copyrighted characters.

Image-to-video prompts are different

When you start from an image, the picture already defines the subject, colors and composition. Re-describing it can make the model redraw instead of animate. Focus the prompt on movement: which way the camera travels, what moves in the scene, and how fast.

Let Sardav write the first draft

Sardav’s Prompt Enhancer applies this anatomy automatically, using the format, duration and style you pick. It keeps your original wording, shows every line it adds, and lets you edit before anything renders — so the structure is automatic, and the taste stays yours.