Prompt Engineering for AI Video: How to Write Prompts That Actually Move

AI video prompts fail differently than image prompts. An image prompt can be vague and still produce something pretty. A video prompt that's vague produces something that melts — faces morph, hands multiply, the camera drifts into soup. Video models need choreography, not just description. Here's how to write it.

The Golden Rule: One Shot, One Prompt

The single biggest mistake beginners make is prompting a whole scene: "a car chase through the city at night with explosions." A video model generates a few seconds per prompt — so write like a director breaking a scene into shots. Each prompt describes one continuous camera setup: one subject, one action, one camera move. If the camera would cut or the action would change, that's a second prompt.

The 5-Block Shot Formula

Every video prompt needs these five blocks, in order:

1. Subject. Who or what is on screen, with specifics (age, clothing, material, color).
2. Action. What the subject DOES during the shot — one clear verb phrase. "Walks slowly toward camera," not "is walking around doing stuff."
3. Camera move. How the camera behaves: static tripod, slow push-in, lateral dolly, crane up, handheld shake, 360 orbit, FPV dive. This is the block most people skip — and it's what makes footage feel cinematic instead of slideshow-like.
4. Lighting. Time, direction, quality — same as still images, but remember light can change during a shot ("clouds parting, sunlight breaking through").
5. Style lock. Two to four style words repeated verbatim in every prompt of a project: "cinematic, 35mm film, shallow depth of field." These are your continuity glue.

Add a video-specific negative prompt to every generation:

morphing, warping, extra limbs, flickering, jitter, watermark, text overlay, logo, cartoonish, low quality

"Morphing" and "flickering" are the video killers — name them explicitly.

Sample Prompt 1: Product Commercial

Matte-black wireless headphones rotating slowly on a dark reflective pedestal, dramatic rim lighting tracing the earcup edges, camera orbiting 180 degrees around the product at constant speed, shallow depth of field, floating dust particles catching the light, cinematic product commercial, photorealistic, 35mm film

Why it works: one action (rotating), one camera move (180-degree orbit), lighting that sculpts the product. The dust particles give the motion something to play against — pure black backgrounds often expose AI flicker.

Sample Prompt 2: Cinematic Landscape

Golden-hour desert dunes with long wind-carved ridges, heat shimmer rising, a lone traveler with a backpack walking along a ridge line, drone shot gliding forward and slowly descending, warm low sunlight raking across the sand texture, cinematic, 35mm film, epic scale

Why it works: the traveler gives the vast landscape a scale reference and a focal point — empty AI landscapes tend to drift. "Gliding forward and slowly descending" is one smooth compound move, not two conflicting ones.

Sample Prompt 3: Character Scene

A detective in a soaked trench coat in his 40s standing under a flickering streetlamp on a rainy night street, rain streaking down, he looks up slowly into the light, handheld camera with subtle shake pushing in from medium to close-up, neon signs blurred in the background, reflections on wet asphalt, cinematic, 35mm film, moody noir

Why it works: the action is minimal and face-forward — "looks up slowly" keeps the face stable, which is where morphing strikes hardest. Rain and reflections add production value while the camera move stays simple.

Sample Prompt 4: Food Macro

Thick golden honey pouring in a slow continuous ribbon from a wooden dipper onto a stack of pancakes, extreme macro shot, camera static on tripod, warm morning kitchen light from the side, steam rising softly, shallow depth of field, hyper-detailed texture, appetizing commercial food video

Why it works: static camera plus one fluid action is the safest possible video prompt — and for food, it looks expensive. Macro hides AI's weakness with complex scenes entirely.

Consistency Tricks Across Shots

• Image-to-video first. Generate a still image of your key frame first, then animate it. You control composition; the model only adds motion. This single habit fixes most continuity problems.
• Repeat style-lock words verbatim in every prompt of a project. Even small rewording ("cinematic film look" vs "cinematic, 35mm film") drifts the aesthetic.
• Describe recurring characters identically each time — same age, clothing, hair, in the same words. Copy-paste the description block between prompts.
• Chain with overlaps: end shot 1's prompt describing the frame that shot 2's prompt starts from ("...ending on the detective's face" → "close-up of the detective's rain-streaked face...").
• Keep clips short. Generate 3–5 second shots and edit them together rather than asking for one long take — quality collapses as duration grows.

Common Failures (and the Fix)

• Melting faces/hands → simplify the action, keep faces toward camera, shorten the clip.
• Camera drift → name ONE camera move; "slow push-in" beats "camera moves around dynamically."
• Style jumping between shots → verbatim style-lock words in every prompt.
• Garbled on-screen text → don't prompt text at all; add titles in your editor afterward.

Key Takeaways

• One shot, one prompt — break scenes into single camera setups.
• Use the 5-block formula: subject, action, camera move, lighting, style lock.
• Animate stills (image-to-video) for maximum control, and repeat style words verbatim for consistency.
• Edit short clips together — the edit is where AI video becomes real video. And disclose AI-generated content where platforms require it.

Related reading

• Prompt Engineering for AI Images: A Framework That Works Every Time — the 6-ingredient prompt stack for stills
• How to Create AI Videos: A Practical Workflow — put your prompts into a real production workflow