How to Write AI Video Prompts That Actually Work (With Examples)

The gap between a mushy AI clip and a scroll-stopping one is rarely the model — it's the prompt. Video models are literal collaborators: they render what you specify and improvise everything you don't. This guide gives you the structure working creators use, real before/after examples, and a shortcut for when you'd rather answer questions than write prose.

The 6-part prompt structure

Strong video prompts answer six questions, roughly in this order:

1. Subject — who or what, with concrete visual detail. "A woman" is a coin flip; "a woman in her 30s with a red windbreaker and wind-blown hair" is a casting call.
2. Action — what happens over time. This is video, not a photo: give the model motion. "She turns toward the camera and laughs mid-stride."
3. Setting — where and when. "A rain-soaked Tokyo street at night, neon reflections in puddles."
4. Camera — shot type and movement. "Handheld medium shot, slow push-in." Camera language is the most under-used lever in amateur prompts.
5. Lighting & mood — "moody tungsten with cyan rim light," "overcast soft daylight, muted colors."
6. Style — the overall grade. "Shot-on-film look, subtle grain, cinematic color grade."

You won't always need all six — but when a generation disappoints, the fix is almost always in a slot you left empty.

Before and after

Weak: a chef cooking in a kitchen

Strong: A middle-aged chef with rolled-up sleeves flips vegetables in a flaming wok, sparks of oil catching the light. Busy restaurant kitchen at night, stainless steel surfaces, warm tungsten key light with cool fill. Handheld close-up, slight slow motion. Documentary kitchen-film style.

The strong version specifies motion (flipping, flames, sparks), a camera (handheld close-up, slow-mo), and light. Every phrase removes a decision the model would otherwise make randomly.

Format changes what you should ask for

The same structure flexes by format:

• UGC / creator style — specify the talent and energy: "smiling creator in her 20s talks straight to a selfie camera, natural daylight, casual bedroom backdrop." Then script the hook. (Full walkthrough: UGC ads with AI.)
• Product videos — the product is the star: name the hero angle, the surface it sits on, and the reveal ("slow orbit around a matte-black earbud case on brushed concrete, studio softbox lighting").
• Cinematic shots — lean on real cinematography vocabulary: lens feel, blocking, era references. Models respond remarkably well to DP-speak.

Image-to-video prompts are different

When you animate a start frame, the image already defines subject, setting, and style — so your prompt should spend its words on motion: what moves, how fast, and where the camera goes. Describing what's already visible wastes tokens and can even fight the image. More in image-to-video vs text-to-video.

Five mistakes that ruin generations

1. Describing a photo, not a video. No verbs, no motion — you'll get a slow zoom on a static scene.
2. Contradicting your reference image. If the frame shows daylight, don't prompt "night."
3. Overstuffing. Two subjects, three actions and a plot twist in 5 seconds produces soup. One beat per clip.
4. Vague style words. "Epic, stunning, 8K" does less than one concrete reference ("hand-held vérité, drab office fluorescents").
5. Ignoring duration. A 5-second clip fits one action. Write sequences as multiple clips instead — a storyboard helps.

Or: let the AI interview you

Here's the honest truth — most people don't want to learn prompt engineering, and you shouldn't have to. SpeedReel's built-in prompt assistant flips the process: pick a format (UGC, cinematic, product…), answer 3–4 short multiple-choice questions about subject, setting and mood, and it writes the full director-grade prompt into your clip. It reads what you've already typed and builds on it rather than starting over, and if you've attached reference images, it looks at them and asks only about what it can't see — the motion.

The assistant produces prompts using exactly the structure in this guide. Think of this article as the manual transmission, and the assistant as the automatic — same engine either way.

FAQ

How long should an AI video prompt be? Usually 40–100 words. Long enough to fill the six slots, short enough that every phrase earns its place.

Do prompts transfer between models? The structure transfers; the dialect doesn't always. Kling rewards motion detail, Seedance handles multi-shot direction well, Veo responds to audio cues. Multi-model editors like SpeedReel make it cheap to A/B the same prompt across models.

Should I use negative prompts? Most modern video models don't take them. State what you want instead of what you don't.