How to Write AI Video Prompts: A Six-Part Structure That Works

AI video models are good at following directions and bad at guessing them. A short, vague prompt such as "a woman walking in the city" leaves the model to choose the shot size, the camera, the light, the pace and the mood, and it will usually choose the most generic option for each. A prompt that makes those decisions for the model gets you a clip that looks planned.

This guide walks through a structure you can reuse for almost any clip, the mistakes that waste the most generations, and how to turn a rough idea into a prompt in Oraclip.

Think in one shot

A generated clip is a single shot, usually a few seconds to a few tens of seconds long. It is not an edited sequence. If your prompt describes a person waking up, making coffee, driving to work and walking into a meeting, the model has to squeeze four scenes into one shot, and every one of them suffers.

Pick the single strongest moment and describe that. If you need a sequence, generate the shots one at a time and cut them together afterwards. You'll get more control over each shot and you'll only regenerate the one that didn't work.

A six-part structure

Most good video prompts answer six questions, roughly in this order:

1. Subject – who or what is on screen, with the details that must look right (age, clothing, material, color, brand-free product shape).
2. Action – one clear thing that happens. Use a concrete verb: "turns to face the camera and smiles", not "is happy".
3. Setting – where it takes place and what time of day it is.
4. Camera – shot size and one camera move: "medium close-up, slow push-in", "wide shot, static tripod", "low-angle tracking shot following from behind".
5. Light and look – the quality of light and the overall style: "soft window light, warm tones, shallow depth of field", "hard noon sun, high contrast, 35mm film grain".
6. Pace and ending – how fast things move and where the shot lands: "slow and calm, ends on her face in sharp focus".

Here is the vague prompt rewritten with that structure:

A woman in her late twenties wearing a camel trench coat walks toward the camera along a wet city sidewalk at dusk. She glances at a shop window and keeps walking. Medium shot, slow backward tracking move at walking pace. Neon reflections on the pavement, soft rain, cool blue tones with warm shop light on her face. Calm, unhurried pacing; the shot ends with her passing just left of the lens.

It's longer, but every sentence removes a guess.

Write the camera like a camera operator

Camera language has the biggest effect on whether a clip feels cinematic or random. A few rules help:

• One primary move per shot. "Slow push-in" or "orbit left" is clear. "Push in, then pan, then crane up" usually produces a wobbly compromise.
• Name the shot size. Wide, medium, medium close-up and close-up give the model a framing to hold.
• Say what stays still. "Locked-off tripod shot" is a valid and often very useful instruction, especially for products.
• Match speed to the action. A fast whip pan on a calm scene feels like a mistake.

If you're not sure how to phrase a move, the Cinematic Shot Director skill in Oraclip turns a plain description into shot-by-shot camera direction.

Describe motion, not feelings

Models render what can be seen. "She feels nostalgic" gives them nothing to animate. "She runs her fingers over an old photograph and pauses" does. The same applies to products: "the bottle looks premium" is weaker than "a narrow highlight slides slowly across the glass as the bottle rotates fifteen degrees".

Keep the amount of motion believable for the clip length. A five-second shot can hold one gesture or one movement across the frame. It can't hold a full dance routine plus a costume change.

Common mistakes

• Stacking styles. "Pixar, anime, photorealistic, oil painting" cancels itself out. Choose one look.
• Too many subjects. Each extra person or object is something else that can drift. Start with one or two.
• Contradictions. "Static shot" plus "the camera flies over the city" forces the model to pick one, and it may not pick yours.
• Text on screen. Signs, labels and captions are still unreliable in generated video. Add text in your editor afterwards.
• Changing everything between attempts. When a clip is close, change one thing (the camera move, or the light) and regenerate. You'll learn what each part of the prompt is doing.

Using a reference image

If you already have an image of the subject, such as a product photo, a character or a selfie, start from it. The image fixes the subject's appearance, so your prompt can focus on motion, camera and ending instead of describing what's already visible. Our image-to-video tips cover this in detail.

Let the agent draft, then edit

In Oraclip you don't have to write the whole prompt yourself. Describe what you want in plain language, in English or Chinese, and the agent expands it into a full generation prompt covering subject, action, camera, lens, lighting, pacing and mood, without changing your intent. Read what it wrote and tighten anything that isn't what you meant.

To skip the blank page entirely:

• Browse Skills for ready-made templates such as Product Hero Video, Travel B-Roll Clip and Logo Reveal Animation. Each one gives the agent tested instructions for that kind of clip.
• Open Prompts to see prompts other creators have shared and try one in a click, then swap in your own subject.

A checklist before you generate

• Is this one shot?
• Is there one clear action with a concrete verb?
• Did you name the shot size and a single camera move?
• Did you choose one look and describe the light?
• Did you say how the shot ends?

If the answer to all five is yes, your first generation has a much better chance of being the one you keep. When you're ready to make more than a few clips, compare the credit plans on the pricing page.