How do you write a good AI video prompt?

Why video prompts are different from image prompts

An image prompt describes a single moment. A video prompt describes a change over time. The model has to decide what moves, in which direction, how fast, and what the camera does while it happens. If the prompt does not say, the model guesses, and guesses tend to be a slow, vague drift or several actions blended into one.

Text-to-video models generate all frames of a short clip together, so the text has to cover the whole clip, not only the first frame. That is why the most useful words in a video prompt are verbs and timing words: walks, turns, pauses, lifts, for two seconds, then. Adjectives about how things look still matter, but they are the second layer.

Official guides from video model makers say the same thing in different words. OpenAI's Sora prompting guide calls a prompt a creative wish list, not a contract: the same prompt gives different results each time, so you write to raise the odds of the result you want, then iterate.

The seven-part structure

Write the parts in this order. Not every prompt needs all seven, but checking each one stops you from leaving a gap the model will fill at random.

Many tools take duration, resolution and aspect ratio as settings rather than text. OpenAI's guide notes that these technical parameters belong in the API call or interface, not in prose. Put them in text only when the tool's documentation says the model reads them there.

One action and one camera move per shot

The most common reason a video prompt fails is that it asks for too much. 'She walks in, picks up the phone, laughs, then runs outside as the camera circles her' is four actions and a complex move in a clip that may last five seconds. The model usually compresses them, skips some, or morphs one into another.

OpenAI's Sora guide recommends keeping each shot simple: one camera move and one subject action. If the story needs four actions, write four shots and join them in an editor. Shorter clips also follow instructions more reliably, so two short shots often look better than one long one.

• Pick the one action the shot is about. Everything else in frame should be small, ambient motion: hair in the wind, steam, traffic far away.
• Pick one camera move, or say static shot. Do not combine an orbit with a zoom and a tilt.
• Say where the action ends. 'Stops at the window' gives the model an end state; 'walks' alone may never finish.
• Describe speed with plain words: slowly, in one quick motion, over three seconds.

Timeline and beat-style prompts

When one shot has a short sequence inside it, write it as beats. A beat is one small step of the action with a rough time. OpenAI's guide suggests counting steps or beats, for example 'takes four steps to the window, pauses', because concrete counts are easier for the model to follow than 'walks over'.

Some tools respond well to explicit time ranges. Others ignore exact seconds but still follow the order. Either way, a timeline makes the order of events clear.

8 seconds, 16:9, static camera, medium-wide shot.
A small cafe counter at morning, soft window light from the left.

0-2s: A barista in a green apron sets a white cup on the counter.
2-5s: She pours steamed milk in one slow motion; a leaf pattern forms.
5-8s: She slides the cup forward and smiles at the customer off screen.

Audio: espresso machine hiss, quiet cafe chatter, a cup touching a saucer.

Keep beats short and physical. 'She feels happy' is not a beat; 'she smiles and lowers her shoulders' is. Leave the final beat as a held moment, so the clip does not end mid-motion and is easy to cut.

A weak prompt rewritten

Here is a typical first attempt. It names a mood and a subject but leaves the action, camera, light and sound to chance.

A cinematic video of a samurai in the rain, epic, 4K, masterpiece.

Problems: 'cinematic', 'epic' and 'masterpiece' do not describe anything the camera can see. There is no action, so the model may produce a slow pan over a standing figure, or a fight with melting swords. There is no camera instruction and no audio. '4K' is a setting, not a prompt word.

Medium shot, low angle, slow push-in.
A lone samurai in a dark blue kimono stands in a muddy village street at night.
Heavy rain. A single paper lantern behind him gives warm orange rim light;
the rest of the scene is cool blue.

Action: he slowly draws his sword over three seconds, then holds it still
at chest height. Rain drips from the blade.

Style: live-action, shallow depth of field, film grain, desaturated colors.
Audio: steady rain, the metal sound of the blade leaving the scabbard, distant thunder.

Every line now answers a question the model would otherwise guess: who, doing what, for how long, seen from where, lit how, in what style, with what sound. The rewrite is longer, but it is not padded. Each phrase changes the image.

Writing audio and dialogue

Models that generate sound with the picture read audio from the prompt. Google's Veo guide recommends saying clearly that you want audio and describing sound in separate sentences: ambient noise, sound effects, and dialogue.

• Ambient: the room tone or environment, such as wind, crowd, traffic, birds.
• Effects: sounds tied to the action, such as footsteps on gravel or a door closing.
• Dialogue: put the exact line in quotes and say who speaks it. Keep it short; one or two brief lines fit a clip of about four seconds.
• Music: say 'no music' if you plan to add your own in the edit, because generated music is hard to cut.

Lip sync and exact wording are still weak points at the time of writing. If a line must be perfect, generate the shot without dialogue and add a recorded or text-to-speech voice in the edit.

Common mistakes and how to fix them

Physics is the other honest limit. Liquids, collisions, hands holding objects and crowds often break. Keep those actions simple, keep hands partly out of frame when you can, and generate several takes to choose from.