Cross-platformPrompting basics

How to Write AI Video Prompts That Work on Any Platform

By Marco Cavazzana · Updated · 5 min read
ShareEmail
Short answer

A portable AI video prompt describes six things in plain sentences: the subject, what it does, the camera, the light, the sound and the style. Runway's guide boils it down to two essentials, visual description and motion description, and Google's Veo guide lists subject, context, action, camera motion, composition and ambiance. Write one clear shot with one main action and one camera move. Then adjust for each platform: drop the motion of the scene into an image-to-video prompt, and leave out sound for silent models.

Every AI video platform publishes its own prompting guide, and each one only covers its own model. In practice you'll use several, so it pays to write prompts that travel. This guide gives you one base structure that works across the major tools, then explains the handful of things that really differ by platform.

The portable structure

PartWhat to writeExample
SubjectWho or what, with the details that must stay fixedA woman in her 30s with short grey hair and a yellow raincoat
SettingWhere, and whenOn a wet city street at dusk
ActionOne main action, at a real-world speedWalks briskly towards camera, then stops and looks up
CameraOne move and the framingMedium shot, slow handheld push in
LightSource, direction, colourNeon shop signs from the left, cool blue ambient light
SoundAmbience, effects, any short line of dialogueRain on pavement, distant traffic, no music
StyleMedium and finishShot on 35mm film, natural grain, realistic colour

This matches what the vendors say. Runway's text-to-video guide says effective prompts contain visual descriptions and motion descriptions, and that natural language "usually gives you more control" than keyword lists. Google's Veo prompt guide names subject, context, action, camera motion, composition and ambiance as the building blocks. Neither insists on a strict order. Runway says structure and order "are far less important than clearly conveying an idea".

Rules that hold everywhere

  • One shot per prompt. If you need a cut, make two clips. Multi-shot prompts are a speciality of some models, not a safe default.
  • One main action. "Picks up the cup and drinks" works. "Picks up the cup, walks to the window, waves to a friend and sits down" asks for four shots in one.
  • Say the speed. AI video drifts towards slow motion. Words like "briskly", "quickly" or "at normal speed" help. Midjourney says its default Low Motion setting is "more likely to result in" still scenes and slow motion.
  • One camera move, named plainly: dolly in, pan left, orbit, static tripod, handheld. Our camera control guide covers the tools that go further.
  • Describe what you want where you can. "An empty street" gives the model something to draw; a list of things to leave out gives it less to work with.
  • Keep spoken lines short and put them in quotes, with who says them.

What differs by platform

Image to video: describe the motion, not the picture

When you start from a still, the image already sets the subject, colours, composition and style. Runway's Gen-4 guide says the text prompt "should be almost entirely focused on describing the desired motion". Cut the Subject, Setting, Light and Style rows down to a few words and spend the prompt on Action, Camera and Sound. This is the biggest single adjustment, and the most common reason a prompt that worked as text-to-video fails from an image. See image to video: which platform for what.

Sound: only where the model makes it

Kling 4.0, Seedance 2.5, Hailuo H3 and Google Flow's Omni and Veo models generate audio with the picture, so the Sound row matters there. Luma Ray3.2 and Midjourney video are silent, so leave sound out rather than confuse the model. Flow also has a project setting to "Return silent videos" if you plan to add sound later. More in AI video with sound.

Parameters and controls outside the prompt

Some things belong in settings, not in words. Midjourney video uses parameters: --motion low or --motion high, --raw for less added flair, --loop and --end. It strips your original image parameters when you animate. Luma and Kling take keyframes, Flow and others take start and end frames, and Higgsfield has named camera presets. Where a control exists, use it instead of describing the same thing in the prompt. See start and end frames.

References for consistency

If a character or product must match across shots, words alone won't hold it. Use the platform's reference feature (Kling Omni Reference, Flow Ingredients, Seedance or Hailuo references) and keep the prompt's description consistent with the references. Our consistent characters guide compares them.

Examples you can adapt

Text to video, with sound:

A woman in her 30s with short grey hair and a yellow raincoat walks briskly along a wet city street at dusk, then stops and looks up at the sky. Medium shot, slow handheld push in. Neon shop signs light her from the left, cool blue ambient light. Sound of rain on the pavement and distant traffic, no music. Shot on 35mm film, natural grain, realistic colour.

The same shot as image to video, starting from an approved still of her:

She walks briskly towards camera, then stops and looks up at the sky. Slow handheld push in. Rain falls steadily and the neon reflections ripple in the puddles. Sound of rain and distant traffic.

Product shot for a silent model such as Luma Ray3.2:

A glass perfume bottle on a black marble plinth, slow 90-degree orbit to the right ending on the front label, soft overhead light with a thin rim light, the bottle stays perfectly still, clean studio look.

Try it

Write one prompt with the structure above and run it unchanged in two platforms. Then write the image-to-video version and run it from the same still. Check whether each model followed the action, the speed and the camera move. Whichever words a model ignored are the words to rework for that model, and the phrasing that worked is worth keeping.

Keeping what worked

After a few projects you will have a set of phrasings each model obeys, and they are easy to lose across several sites. Dreamdrive, a free Chrome extension, saves each clip as you make it with its prompt, model and settings, in one searchable library across your AI tools. You can search for a camera move or a style phrase and see exactly which wording worked in which model.

Related: how to waste fewer credits.

Frequently asked questions

What is the best structure for an AI video prompt?
Describe the subject, setting, one main action, one camera move, the light, the sound (if the model generates audio) and the style, in plain sentences. Runway's guide boils this down to visual description plus motion description; order matters less than clarity.
Why does my AI video ignore my prompt?
Common causes are too many actions in one clip, several camera moves at once, describing the picture instead of the motion in an image-to-video prompt, and asking a silent model for sound. Simplify to one action and one camera move, and use the platform's controls (keyframes, presets, references) where they exist.
Should I use keywords or full sentences for AI video?
Full sentences usually work better. Runway says natural language gives you more control, while keywords act more like suggestions that the model interprets loosely.
How do I stop AI video looking slow motion?
Say the speed in the prompt (briskly, quickly, at normal speed), give the clip a clear action that fills its length, and use higher motion settings where they exist, such as Midjourney's --motion high. Midjourney notes its default Low Motion is more likely to produce slow motion.

Nothing you makeis lost.

Free · No card · Read-only · 2-minute setup

DreamdriveFree plan · Chrome
Add to Chrome — it’s free