The same AI video model can produce a forgettable blur or a striking cinematic shot, and the difference is often the prompt. Video prompting is a learnable skill. It has more in common with briefing a camera crew than with writing a story. This guide gives you a formula, examples and a method for improving results quickly.
Think in shots, not stories
A video model generates one continuous shot of a few seconds. It cannot follow a plot, jump between locations or show cause and effect over minutes. So the first rule is to describe a single shot: what is in frame, what moves and how the camera behaves. If your idea needs a scene change, it needs two prompts.
The shot-description formula
Most strong video prompts contain six elements, roughly in this order:
- Shot type: wide shot, medium shot, close-up, extreme close-up, aerial, over-the-shoulder.
- Subject: who or what, with key visual details. “An elderly fisherman in a yellow oilskin coat.”
- Action: one clear movement. “Hauls a rope onto the deck.”
- Setting: where and when. “On a small wooden boat in a grey, choppy sea, early morning.”
- Camera: movement and lens. “Handheld, slight sway, 35mm lens.”
- Look: light, colour and style. “Overcast diffused light, muted blue-green palette, documentary film look.”
Put together: “Medium shot of an elderly fisherman in a yellow oilskin coat hauling a rope onto the deck of a small wooden boat in a grey, choppy sea, early morning. Handheld camera with slight sway, 35mm lens. Overcast diffused light, muted blue-green palette, documentary film look.”
Word order matters
Many models pay most attention to the beginning of a prompt. Lead with the most important element, usually the subject and action, or the shot type if framing is critical. Put secondary details such as film grain or colour grading at the end.
Describe motion explicitly
Motion is what separates video from images, and models need guidance. Use clear verbs and speed words: “walks slowly”, “turns her head towards the camera”, “leaves drift gently”, “waves crash rhythmically”. Words such as “slow”, “gentle”, “subtle” and “steady” help avoid chaotic movement. If you want stillness, say so: “static camera, the subject remains still, only the candle flame flickers”.
Use camera language
Models have learned a lot from film and photography, so cinematic terms work well: dolly in, pan left, tilt up, orbit, crane shot, tracking shot, rack focus, shallow depth of field, wide-angle lens, telephoto. Use one camera move per shot. Combining several usually produces a confused result. Our guide to camera movement prompts covers each move in detail.
Style and look keywords
Style words set the overall feel:
- Lighting: golden hour, blue hour, neon, candlelight, overcast, hard noon sun, rim light.
- Film and format: 35mm film, 16mm grain, VHS, anamorphic, documentary, music video.
- Art styles: watercolour, claymation, cel-shaded anime, oil painting, paper cut-out, low-poly 3D.
- Colour: teal and orange, desaturated, pastel, high contrast black and white.
Two or three style terms are enough. A long list of conflicting styles muddles the result.
Length: say enough, not everything
Very short prompts give the model too much freedom. Very long prompts bury the important parts. Somewhere between 30 and 80 words works well for most models. Some recent models understand long, detailed paragraphs better than older ones; check the model’s documentation or example prompts and match their style.
What to avoid
- Abstract ideas: “a feeling of loss” means little visually. Show it: “an empty chair by a window, rain outside, dim light.”
- Several actions: “she opens the door, walks in, sits down and starts crying” is four shots.
- On-screen text: most models write gibberish. Add text in your editor.
- Negations in the main prompt: “no people” can sometimes add people. Use a negative prompt instead, see Negative Prompts.
- Real people’s names: legally risky, ethically wrong without consent, and usually inaccurate anyway.
An iteration method
- Write the prompt using the formula.
- Generate three or four variations with different seeds.
- Pick the best and list what is wrong.
- Fix the seed and change one element of the prompt.
- Repeat until the shot works, then render at final quality.
Keep a prompt journal with what you changed and what happened. Patterns emerge fast, and your first drafts get better each week.
Before and after
| Weak prompt | Strong prompt |
|---|---|
| “A cat in a city.” | “Low-angle tracking shot of a ginger cat walking along a wet brick wall at night, neon signs reflecting in puddles below, shallow depth of field, cinematic, moody blue and pink light.” |
| “Sad soldier.” | “Close-up of a young soldier in a 1940s uniform sitting in a muddy trench, staring into the distance, smoke drifting behind him, overcast light, desaturated colours, slow push-in.” |
| “Fantasy forest, magic.” | “Wide shot of an ancient forest with glowing blue mushrooms on mossy roots, fireflies drifting slowly, morning mist, soft volumetric light beams, static camera, painterly fantasy style.” |
Prompting for image-to-video
When you start from an image, the picture already defines the subject, setting and style. Your prompt should focus on motion and camera: “the woman slowly turns her head and smiles, hair moving gently in the wind, subtle camera push-in”. Repeating every visual detail of the image can confuse the model. See Image-to-Video Prompting.
Build a personal prompt library
Every time a prompt produces a shot you love, save it with the model name, seed and settings in a simple document or spreadsheet. Tag entries by type: landscapes, portraits, products, action, stylised. Within a month you will have a library of proven starting points, and new projects begin from something that already works instead of a blank box.
Frequently asked questions
Do prompts work the same across tools?
The formula transfers well, but each model has quirks. Check its example prompts.
Should I use commas or full sentences?
Newer models usually prefer natural sentences; older ones were trained on comma-separated tags.
Why does my subject move too fast?
Add speed words such as “slow” and “gentle”, and reduce motion strength if the tool has that setting.
