A prompt covers one shot, not a film
A shot is one continuous piece of video from one camera position. OpenAI's Sora 2 prompting guide asks for one clear camera move and one clear subject action in each shot. When a prompt covers several shots, the guide keeps each block to one camera setup, one subject action and one lighting recipe.
Clips are short, which is another reason to write one shot per prompt. The chart gives the longest clip a single call makes, according to each vendor's own documents.
Some tools add cuts on their own. Google's Gemini Omni notebook covers Omni, a Google video model you can talk to for edits, and says it may add multi-shot cuts by default. For one continuous camera it suggests writing "In a single unbroken scene..." or "Continuous handheld shot, no scene cuts..." into the prompt.
Sources: OpenAI Cookbook: Sora 2 Prompting Guide (read November 6, 2026), GitHub: google-gemini/cookbook, Gemini Omni notebook (read November 6, 2026), GitHub: zai-org/CogVideo, README (read November 6, 2026), GitHub: google-gemini/cookbook, Get started with Veo notebook (read November 6, 2026)
The parts every vendor puts in a prompt
The guides name the same ingredients in different words. Tencent's HunyuanVideo-1.5 handbook boils a basic prompt down to Subject + Motion + Scene and treats shot type, camera movement, lighting, style and atmosphere as optional extras.
Lightricks' LTX-2 README gives seven steps, from the main action to sudden events, written as one flowing paragraph of up to 200 words. OpenAI's guide says to sketch a shot as if you were drawing it onto a storyboard.
| Part | What to write | Where the advice comes from |
|---|---|---|
| Subject | Appearance, clothes, species: a few distinctive details | Tencent lists appearance, attire and species. OpenAI says a few distinctive details keep a subject recognizable |
| Action | One action, written as steps | OpenAI: one clear subject action, in beats. Tencent: "First... then... next... meanwhile... finally...". Lightricks: start with the main action in one sentence |
| Setting | Place, background, weather | Lightricks: background and environment details. OpenAI: scenery and weather |
| Camera | Framing, angle and one move | Google: "single shot", "eye level", "dolly shot", "tracking shot". Tencent: shot type and camera movement. Lightricks: camera angles and movements |
| Light and style | Light, color, lens, overall look | Google: "shallow focus", "macro lens" and styles such as "sci-fi" or "animation". Tencent: lighting, style and atmosphere. Lightricks: lighting and colors |
| Sound | Sounds and speech, as text | Google: Veo 3 and Omni make audio from the prompt. OpenAI: speech goes in a Dialogue block |
Write a prompt, step by step
Keep the first draft short. Tencent's handbook marks the items in brackets as optional, so write the three required parts, then add the others one at a time until the clip matches the picture in your head.
The example below is Google's own timing prompt from the Omni notebook. It is one continuous handheld shot in two beats of three seconds each, with a subject, a place and a weather detail.
[0-3s] A continuous handheld shot of a woman in a red coat standing still on a snowy city sidewalk looking forward. [3-6s] She slowly opens a bright yellow umbrella as soft snowflakes fall around her.- Say the shot in one sentence: who is in it, what they do and where.
- Add what the shot needs from the camera: a shot type, one move, a light and a style.
- Write the action as steps with connecting words. Tencent suggests "First... then... next... meanwhile... finally...".
- Swap feelings for things the camera can see. Tencent's formula is Subject + Action + Small Detail, as in "The boy smiles, eyes crinkling slightly."
- Say who is where and where they move, in simple words such as left, right, foreground and background, as Tencent advises.
- Add sound or speech as text, then run the prompt more than once before you change it.
Timing, sound and speech
Omni takes timing in the prompt itself. Its notebook shows plain phrasing, such as "after 3 seconds, a woman enters the scene", and a bracketed form that splits the clip into timed beats, like the example above.
Sound is text too. Veo 3 and Omni make audio from the prompt, so a sound goes in the sentence like any other detail, and the Veo lesson shows the call that sends it. OpenAI's guide moved speech into a Dialogue block under the scene, which the Sora lesson walks through.
| You want | Write it like this | Source |
|---|---|---|
| One continuous camera | "In a single unbroken scene..." | Gemini Omni notebook |
| An action at a set time | "[3-6s] They stop and turn around." | Gemini Omni notebook |
| Spoken words | Each line in quotes (Veo), or a Dialogue block under the scene (Sora) | Veo notebook, Sora guide |
| Sound without speech | "The hum of espresso machines and the murmur of voices form the background." | Sora guide |
| Nothing spoken, no text on screen | "No dialogue or voiceover. No text overlay on screen." | Gemini Omni notebook |
Keep a character the same from shot to shot
Each prompt starts from nothing, so a character stays the same only when you hand the tool the same things each time. OpenAI's guide warns that small changes in phrasing can change identity, pose or the focus of a scene. Its advice is to keep descriptions consistent, reuse phrasing and avoid mixing traits that compete.
Tencent adds a rule for frames with multiple characters: tell them apart by an attribute or a position. Its example has "The black cat hands the bomb to the gray cat" and then sends the gray cat running toward the right side of the frame.
| Method | What you do | Where it comes from |
|---|---|---|
| Same words | Paste the same description of the character into every prompt | OpenAI Sora guide |
| A start picture | Pass the picture as the first frame, so the look is set before the video begins | Veo notebook, Sora guide |
| Reference images | Up to 3 in Veo 3.1 (not Fast). Up to 6 in Omni, across up to 2 characters plus props or wardrobe | Veo and Omni notebooks |
| Told apart in the prompt | Name each character by an attribute or a position, such as the black cat and the gray cat | Tencent handbook |
| A follow-up edit | Change one thing in a new turn and add "Keep everything else the same" | Omni notebook |
| A stored character | Sora kept a character from a short reference clip. That ended with the Sora API | Sora guide |
When a clip is close but wrong
OpenAI's guide says to make controlled changes one at a time and to say what you are changing, for example "same shot, switch to 85 mm" or "same lighting, new palette: teal, sand, rust." When a result is close, it suggests pinning that clip as the reference and describing only the tweak.
Google's Omni notebook gives the same advice for follow-up edits: keep them short and surgical. Its good example is "Change the car color to metallic blue. Keep everything else the same." Its bad one re-describes the whole video before it asks for the change.
When a shot keeps going wrong, the Sora guide says to strip it back. Freeze the camera, simplify the action and clear the background, then add complexity one step at a time once it works.
Editing is for nudging, not gambling.
Common mistakes
- Packing several scenes into one prompt. A clip lasts seconds, and the Sora guide wants one camera setup and one action per block.
- Leaving cuts to chance. If the tool adds them, write "single unbroken scene" into the prompt.
- Describing feelings instead of things. Tencent says to turn abstract emotion words into action details.
- Rewriting the whole prompt to fix one detail. Both OpenAI and Google say to change one thing and keep the rest.
- Rewording a character between prompts, which can change who appears in the clip.
Try it yourself
- Write one shot of your own as Subject + Motion + Scene, then add a single camera move and a light.
- Take the timed prompt above into a video tool that accepts timecodes, then change only the second beat.
- Write a character description once, paste it into two prompts and compare faces and clothes.
- Fix a near miss with one edit that ends "Keep everything else the same."
- Add a line that names what must stay out, such as "No dialogue or voiceover."
Check each clip against your own list: the subject looks as described, the camera does the one move you asked for and nothing you ruled out has appeared.
Next: run an open video model yourself
Everything above also applies to open models, which run on your own computer. Open video models covers Wan, HunyuanVideo, LTX and CogVideoX, with the sizes and licences each project states.
Check yourself
A tool keeps adding camera cuts to your clip. What does Google's Omni notebook suggest?
Show answer
C. Write "In a single unbroken scene..." into the prompt The notebook says Omni may add multi-shot cuts by default and gives that wording to force one continuous camera.
Tencent's handbook gives Subject + Motion + Scene as the basic prompt. Which parts does it call optional?
Show answer
A. Shot type, camera movement, lighting, style and atmosphere The items in square brackets in its formula, from shot type to atmosphere, are the optional ones.
You want a character to look the same in a second prompt. Which step does the Sora guide give?
Show answer
B. Reuse the same phrasing for the character The guide warns that small changes in phrasing can change identity, so it says to keep descriptions consistent and reuse phrasing.
A clip is close, but the car should be blue. What does Google's notebook tell you to write?
Show answer
D. Change the car color to metallic blue. Keep everything else the same. The notebook calls for short, surgical edits and sets this sentence against a long re-description of the video.
Frequently asked questions
How do I write a good AI video prompt?
Describe one shot in plain words: who is in it, what they do and where. Camera, light and style come next, only where the shot needs them, and any sound is written as text.
How long should an AI video prompt be?
The vendors differ. Lightricks asks for one paragraph under 200 words, Tencent for simple, direct words, and OpenAI says short prompts leave more to the model while long ones restrict it.
Can I put several scenes in one video prompt?
In some tools. OpenAI calls multi-shot prompts valid when each shot block stays distinct, and Omni takes timecodes such as [0-3s]. A clip lasts seconds, so each scene gets little time.
What is a negative prompt in AI video?
A short list of what you do not want to see. Veo takes it as the negative_prompt setting, while Omni's notebook writes a plain sentence such as "No text overlay on screen."
Why does my character change between clips?
Each clip starts from your words alone. Reuse one description, or give the tool a picture or reference images: Veo 3.1 takes up to three and Omni up to six.
Sources
- OpenAI Cookbook: Sora 2 Prompting Guide https://developers.openai.com/cookbook/examples/sora/sora2_prompting_guide
- GitHub: Tencent-Hunyuan/HunyuanVideo-1.5, Prompt Handbook https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5/blob/main/assets/HunyuanVideo_1_5_Prompt_Handbook_EN.md
- GitHub: Lightricks/LTX-2, README https://github.com/Lightricks/LTX-2
- GitHub: google-gemini/cookbook, Get started with Veo notebook https://github.com/google-gemini/cookbook/blob/main/quickstarts/Get_started_Veo.ipynb
- GitHub: google-gemini/cookbook, Gemini Omni notebook https://github.com/google-gemini/cookbook/blob/main/quickstarts/Get_started_Omni.ipynb
- GitHub: zai-org/CogVideo, README https://github.com/zai-org/CogVideo