The best AI video prompts are written like directions to a camera crew, not like requests to a chatbot. Name one shot size and one camera move, describe one clear action in the present tense, then the place, the light and the sound. Keep it to two to five sentences, describe motion instead of a still picture, and when you start from an image, write only what moves. Those habits work on every current video model. The ten practices below show each one with a before-and-after example.
A lot of prompting advice online was written for chatbots and then pasted onto video. We think that is why so many first clips disappoint: “give the AI a role”, “explain your target audience”, “ask it to think step by step”. Video models do not read prompts that way. They turn your words into pictures and motion, so the only words that help are words a camera could film.
1. Lead with the action, not the adjectives
A video model’s first job is to decide what happens. If your prompt opens with “a beautiful, cinematic, moody scene”, it has learned nothing about the shot yet. Open with who is doing what.
- Weak: A cinematic, atmospheric, stunning shot of a lighthouse keeper, moody lighting, 8K.
- Better: An old lighthouse keeper climbs the last steps of a spiral staircase and pushes open the door to the lamp room. Wind tugs at his coat.
Every current model guide says some version of this: the subject and the action come first, everything else supports them.
2. One shot size, one camera move
“Close-up” plus “slow push in” is a direction. “Low aerial handheld orbit push-in with a crash zoom” is a wish list, and the model usually picks one item at random or blends them into mush.
- Weak: Dynamic camera movement, sweeping and orbiting around her, then zooming in dramatically.
- Better: Medium close-up. The camera slowly pushes in as she reads the letter.
If a shot really needs two moves, it is usually two shots. Cut between them.
3. Describe motion over time, not a still picture
A prompt that would make a good photograph often makes a dull video. “A woman standing by a window at sunset” gives the model nothing to animate, so it adds random movement: hair blowing, a slow drift, a blink. Tell it what changes from the first second to the last.
- Weak: A woman by a window at sunset, looking thoughtful.
- Better: A woman stands at the window as the sun drops behind the rooftops. She lifts her phone, reads the message, and lowers it without replying.
4. Show emotion through the body
Models cannot film “sad” or “nervous”. They can film what sad and nervous people do. Swap feeling words for physical ones.
- “He is angry” becomes “his jaw tightens and he keeps sanding the table, harder”.
- “She is nervous” becomes “she turns the ring on her finger and glances at the door”.
- “They are relieved” becomes “their shoulders drop and one of them laughs once, quietly”.
Several model guides now say this outright. It is also simply better screenwriting.
5. Give the light one source
“Dramatic lighting” means nothing in particular. “Warm light from a single bulb overhead, the corners of the room in shadow” gives the model a light it can place. Pick one light logic per shot and name where it comes from: a window, a bulb, neon signs, a campfire, the screen of a phone. Mixed and unexplained light is one of the quickest ways to make a shot look generated.
6. Write what you hear
Several of the strongest current models generate sound with the picture, and they follow sound directions closely. Name the actual sources rather than a mood:
- Weak: Atmospheric sound.
- Better: Rain on a tin roof, a dog barking somewhere down the street, the hum of the fridge.
For dialogue, put the exact words in quotation marks, say who speaks them, and keep stage directions out of the quote. “(pause)” inside the quotes can end up being read out loud. A line of ten words needs about four seconds of clip, so check the length before you blame the model for cutting someone off. If you do not want the words to appear as subtitles, say so: “no on-screen text, no subtitles”.
7. With an image, write only what moves
When you start from a frame (image-to-video), the frame already answers every question about how things look. The prompt should answer the one question the frame cannot: what happens next.
- Weak: A young woman with short dark hair in a navy raincoat stands in a doorway of a wood workshop with a workbench, slow push in.
- Better: Slow push in. She steps inside and sets the key on the workbench, then hesitates before letting go of it.
Re-describing the picture is the most common image-to-video mistake. It tempts the model to change what you wanted to keep. If you are not sure which mode to use, our guide to text-to-video vs image-to-video walks through it.
8. Keep the character’s description word for word
If the same character appears in several shots, write their description once and paste it, unchanged, into every prompt they appear in. “Theo, man in his 40s, short brown beard, red and black flannel shirt with rolled sleeves” every single time. Rewording it as “bearded man in a checked shirt” in shot four is an invitation to cast someone new.
If your model accepts reference images, give each one a single job in plain words, and say what to ignore: “Image 1 is Theo. Use only his face, hair and clothes. Do not use the background.” Without the last sentence, models tend to drag the reference photo’s background into the shot.
9. Short, simple shots beat long, busy ones
The longer a clip runs, the more it drifts: faces soften, hands grow extra fingers, objects appear and vanish. The busier the frame, the faster it happens.
- Aim for one action in four to eight seconds.
- Keep crowds, signage and on-screen text out of the shot unless you need them. Text in particular is still unreliable on most models.
- Choose plausible physics. A person sitting down works. A person backflipping off a moving car usually does not, yet.
A scene made of eight clean five-second shots looks more professional than one forty-second take that falls apart halfway, and it is far cheaper to fix.
10. Change one thing at a time
Your first prompt is a draft. When a clip comes out wrong, change one part of the prompt and run it again, ideally on a cheaper draft setting first. Change the camera, or the action, or the light, but not all three, or you will not know which change fixed it.
Keep the versions that work. After a few projects you will have your own small library of prompts for the shots you make most often: the walk and talk, the insert of a hand on an object, the reaction close-up.
A prompt formula that works on every model
Put together, the ten practices give you a simple order to write in. Drop any part a shot does not need.
- Shot: size and angle. “Close-up, eye level.”
- Subject and action: who, doing what, in the present tense.
- Place: where, and the time of day.
- Camera: one move, or static.
- Light: one source.
- Sound: real sources, and any dialogue in quotes with the speaker named.
Close-up, eye level. An old lighthouse keeper pushes open the door to the lamp room and squints into the wind. Top of the tower at dusk. The camera holds still. Cold blue light from the sky, the lamp still dark. Wind whistling through the railings, the creak of the door.
Each model then has its own preferences on top of this. Some like labelled fields, some want camera moves described with a size and speed (“pushes in with small amplitude at slow speed”), some want a reference image labelled a certain way. The free prompt translator takes one description of a shot and writes it in the format each model prefers.
FAQ
How long should an AI video prompt be?
For one shot, two to five sentences is the sweet spot on most models. Short prompts let the model fill in framing and mood, which is useful when exploring. Longer prompts give you more control, but every sentence has to describe something visible or audible, or it only adds noise.
Should I write “cinematic” or “8K” in my prompts?
Those words do very little on their own. A specific camera move, a named light source and a real action do far more to make a shot look cinematic than any quality keyword.
Do the same prompts work on every model?
The principles do. The exact format does not: models differ in how they want dialogue, references and camera moves written. Start from the formula above and adjust for the model you use.
Why does my video ignore half of my prompt?
Usually because the prompt asks for too much at once: several moves, several actions, several moods. Cut it back to one move and one action, check the result, then add the next detail.
How do I keep a character the same across clips?
Use one fixed written description in every prompt, start each clip from a frame or reference image of that character, and keep clips short. Our guide to consistent style across AI video clips goes deeper.