How to Write Video Prompts That Look Cinematic

Most AI video prompts fail the same way: they describe a subject and stop. "A woman walking through Tokyo" gives the model nothing to direct, so you get a documentary shot from a camera that does not exist. Cinematic clips come from prompts written like a director would brief a shot — camera, light, motion, mood. Here is the formula.

1. Camera first, subject second

Name the shot before the content: close-up, wide establishing, over-the-shoulder, drone pull-back, handheld follow. The lens language does more for the look of the frame than the subject does.

• Weak: a chef plating a dish
• Strong: slow macro close-up, shallow depth of field, a chef plating a dish

2. Light like a cinematographer

Light is the difference between a clip and a stock video. Specify source, direction and quality: golden-hour backlight, hard noon sun, single practical lamp, neon bounce at night, soft overcast key.

• Weak: a perfume bottle on a table
• Strong: low-key studio light, one hard key from camera left, a perfume bottle on a black table, rim light on the glass

3. Give the motion a direction

Models animate what moves — so say what moves and how: "steam rises slowly", "the camera pushes in", "her coat flutters in the wind". One deliberate motion reads as cinema; three competing motions read as noise.

4. Set the mood, then stop

End with grade and tone: muted teal-and-orange, faded film grain, high-contrast noir, pastel and airy. Then stop writing. Long prompts dilute; every extra clause is another instruction the model can drop.

A full example, assembled

Slow dolly-in on a lone astronaut standing in a red sand desert, golden-hour side light, dust drifting across the lens, muted warm grade, subtle film grain.

Camera, light, motion, mood — eleven words of scene, all of them directional. That structure, reused across shots, is also how you keep a multi-scene video consistent: lock the light and grade, change only the subject and the camera move.

Iterate in passes

First pass: get the composition right with a fast model. Second pass: swap to a heavier model and keep the winning prompt verbatim. Third pass: change one variable at a time — never rewrite a working prompt wholesale. On PavoAI you can run the same prompt across Veo, Kling and the rest in one window, which is exactly the workflow this formula was tuned for.

Write your next prompt with the four parts and compare it against your old one — the difference is usually immediate.