Zero Shot

How to write an AI video prompt: a formula, examples and common mistakes

An AI video prompt is one shot described in order: subject, action, setting, camera, light, sound. A formula, worked examples and the mistakes to avoid.

7 min read

An AI video prompt works when it describes one shot the way a director would brief a camera operator: who or what is in frame, what they do, where, how the camera moves, what the light is like and what you hear. Write those six things in plain sentences, ask for one action and one camera move per clip, and you will get usable footage far more often than with a paragraph of adjectives. The rest of this guide is the formula, worked examples and the mistakes we see most.

What should an AI video prompt include?

Six elements cover almost every shot. You do not need all of them every time, and the order below is the one we find easiest to write and to edit.

Element What to write Example
Subject Who or what the shot is about, with two or three visible details A woman in her thirties in a grey wool coat
Action One thing that happens, with a pace lifts a paper cup and takes a slow sip
Setting Place, time of day, weather at a street market on a cold morning, breath visible
Camera Framing and one movement medium close-up, slow push in, eye level
Light and style Light source, colour, the kind of footage soft overcast light, muted colours, shot on a 50mm lens
Sound Ambience, effects, any spoken line market chatter and a distant bus, no music

Put together: “A woman in her thirties in a grey wool coat lifts a paper cup and takes a slow sip at a street market on a cold morning, breath visible. Medium close-up at eye level, slow push in. Soft overcast light, muted colours, shot on a 50mm lens. Audio: market chatter and a distant bus.”

This matches how the model makers describe it themselves. Google’s prompt guide for its video models lists subject, action, scene, camera angles and movements, lens effects, visual style, timing and audio, and adds that you do not need to use every element in every prompt.

How is an image-to-video prompt different from text-to-video?

In text-to-video the prompt has to describe everything, because the model invents the whole frame from your words. In image-to-video you supply a still image as the first frame, so the look is already decided. The prompt then has a single job: say what changes.

A good image-to-video prompt is short. “She turns towards the window and smiles, slow push in, curtains moving slightly” is complete. Describing the coat, the room and the colour of the walls again is worse than useless, because it invites the model to redraw things you had already approved.

This is why most studio work, ours included, is image-first. The frame is designed and approved as a still, and the motion prompt only has to move it. We cover that whole process in the workflow for making Instagram Reels with AI, and the reference-image side of it in how to keep an AI character consistent across 20 shots.

How do you describe camera movement in a prompt?

Use the words a film crew uses, and use one per shot. The model makers’ own guides are written in these terms, and the models respond to them far better than to “make it dynamic”.

  • Push in or dolly in: the camera moves towards the subject. Good for emphasis.
  • Pull back: the camera moves away and reveals the setting.
  • Pan and tilt: the camera turns left or right, or up or down, from a fixed point.
  • Tracking shot: the camera travels alongside a moving subject.
  • Orbit: the camera circles the subject. Useful for products.
  • Handheld: slight natural shake, for a documentary or phone-footage feel.
  • Static or locked-off: no movement. Say so, or the model may add some.

Add a speed (“slow”, “gentle”, “fast whip”) and a framing (“wide shot”, “close-up”, “over the shoulder”). Two movements in one short clip, such as “orbit, then crane up”, usually produce a muddle. Cut the idea into two shots.

How do you write sound and dialogue into a prompt?

Several current models generate audio with the picture. Google’s documentation, as of November 2026, describes Gemini Omni Flash as producing 3 to 15 second videos with native audio. If your tool does this, the sound belongs in the prompt, or the model will guess.

Google’s guide gives two pieces of advice that are worth copying: say clearly that you want audio, and describe it in separate sentences from the picture. It names three kinds of sound to specify: sound effects, ambient noise and dialogue. So write “Audio: rain on a tin roof, a kettle clicking off” as its own sentence. For speech, name the speaker and give the exact words, and keep the line short enough to be said in the length of the clip. A sentence that takes eight seconds to read aloud will not fit into five.

Even with good native audio, we treat it as a draft. The reasons are in our guide to sound design for AI video.

Do negative prompts work?

Sometimes, and it depends on the tool. Some generators have a separate negative prompt box, some read exclusions from the main prompt and some ignore them.

Where they are supported, the wording matters. Google’s guide says to describe what you do not want to see as a plain list, and not to use instructions such as “no” or “don’t”. Its example is writing “wall, frame” to keep those items out. “No walls” tends to backfire, because you have just written the word “walls” into the prompt.

The more reliable habit is to say what you do want. “An empty beach at dawn” beats “a beach with no people”. “Clean white background” beats “no clutter”.

A prompt template you can copy

  1. Subject: who or what, plus two or three visible details.
  2. Action: one action, with its pace.
  3. Setting: place, time of day, weather or season.
  4. Camera: framing, angle, one movement and its speed.
  5. Light and style: the light source, the palette and the kind of footage (film, phone, animation, product studio).
  6. Sound: ambience, effects and any line of speech, in separate sentences.
  7. Exclusions: only if the tool supports them, as a list of things.

Here is a product version. “A matte black water bottle stands on a wet slate surface. Water droplets run slowly down its side. Dark studio background. Close-up, slow orbit from left to right. A single soft light from the upper left, cool tones, sharp focus on the bottle. Audio: a low hum and a single drop of water.”

One warning about products: a prompt alone will not reproduce your real bottle, label or logo. Accurate products come from your own photos used as references, a point we explain in what AI video still cannot do well.

What are the most common prompt mistakes?

  • Too many actions. “She walks in, sits down, opens the laptop and laughs” is four shots. Ask for one.
  • Adjectives instead of things. “Cinematic, stunning, epic” tells the model nothing it can draw. “Low sun behind her, long shadows, shallow depth of field” does.
  • Contradictions. “Handheld” and “perfectly smooth”, or “bright midday” and “moody shadows”.
  • Describing the image again in image-to-video. Describe the movement only.
  • Asking for readable text. Signs, labels and captions still come out wrong often enough that we add text in the edit.
  • Changing everything between attempts. Change one element, generate again and compare. Otherwise you learn nothing from the take.
  • Carrying one prompt between tools unchanged. Each model has its own habits. Our comparison of the best AI video generators in 2026 shows how different they are.

Why do the same words give different results each time?

Generation starts from random noise, so the same prompt gives a different clip on every run unless you fix the seed, and many tools do not let you. That is normal, not a fault in your writing. Professionals plan for it: several takes per shot, one chosen, the rest thrown away.

It also explains why prompting is a small part of the job. The prompt gets you a usable take. The idea, the shot list, the consistency between shots, the edit and the sound decide whether the takes add up to a video anyone watches. A beautiful clip of the wrong thing is still the wrong thing.

At Zero Shot, an AI video production studio in London, prompts are written last: after the script, the shot list and the approved still frames. If you would rather hand over the whole process, from idea to finished reels and short-form video, send us a short brief and we will reply with a few questions and a fixed price.

Questions people ask

How long should a prompt for AI video be?

Long enough to cover the subject, the action, the setting, the camera and the light, which is usually two to four sentences for text-to-video. When you animate an existing image, one sentence about the movement is often enough.

Do the same prompts work in every AI video generator?

The structure carries over, but the details do not. Each model reads camera terms, sound instructions and negative prompts in its own way, so check the tool's own prompt guide and expect to adjust wording when you switch.

Can ChatGPT or another chatbot write video prompts for me?

Yes, and it is a reasonable way to expand a rough idea into full sentences. The chatbot cannot see the result, though, so you still have to watch each take and decide what to change.

Why does my AI video ignore part of the prompt?

Usually because the prompt asks for too much in one clip. Several actions, two camera moves or a long list of adjectives compete with each other, and the model keeps some and drops the rest. Split the idea into separate shots.

Keep reading

Contact

Tell us what you need

Tap a few options. We will open a chat with your brief already written.

What do you need?
When?
Budget
Your messageHi Zero Shot! I would like to talk about a project.
Send it via

Prefer to talk? Send a voice message instead. English or Russian, whichever is easier.

WhatsApp Start a project