Mixer AI
How to describe a scene for an AI clip, step by step
A good video prompt describes what is seen and heard as the clip plays: who is in the frame, what they do second by second, where the camera is, what we hear and how the scene ends. The block order is almost the same for every model; what differs is how shots change and how sound and references are handled.
One continuous shot, a static wide-angle camera right at ground level. Autumn, fallen leaves and nuts on the ground, blurred houses and trees behind, warm low sunlight. A squirrel sits on the leaves holding a piece of nut in its front paws and looks into the camera. It brings the nut to its mouth and gnaws it, whiskers and ears twitching slightly. The paws and the nut stay in the frame the whole time. At the end the squirrel is still gnawing and does not run away.
Step by step
Pick the model, mode, duration, aspect ratio and sound in the studio settings or in the node on the canvas.
Do not repeat them in the prompt: the text is only for the scene.
If you attach photos, videos or audio, start with their roles: "image 1 is the woman's face, image 2 is the kitchen, audio 1 is her voice".
A reference without a role may be ignored.
Write the action in order or with timecodes: one main action per segment, a reaction and a new state.
Add dialogue with the exact words in quotes, sound and camera.
Finish with the end state of the scene and a line of exclusions: no music, no extra people, no on-screen text.
Check the price before you run it.
Watch the whole clip: the order of actions, dialogue timing, extra objects.
Change one thing at a time so you know what worked.
Ready video prompts
Image 1 is the courier's face and clothes. Handheld camera, documentary style. Evening, the entrance of an apartment building, a dim lamp above the door. 0-2 s: close-up, a finger presses the intercom button. 2-6 s: medium shot, the courier waits holding a box, shifts from foot to foot and checks his watch. 6-8 s: the door opens, the courier holds out the box and says tiredly: "Your order. Please sign here." Sound: intercom beep, wind, the click of the lock. No music. Exactly one box in the frame.
Scene 1, wide shot: a girl in a yellow raincoat walks along an empty pier in light rain, static camera. Scene 2, medium side shot: she stops at the edge of the pier and looks at the sea. Scene 3, close-up: raindrops on her face, she closes her eyes and takes a deep breath. Sound in every scene: surf, rain tapping on the raincoat.
Image 1 is the girl's appearance, audio 1 is her line. Shot at arm's length, slightly from below, no phone in the frame. The girl sits in a parked car, daylight from the side window. She looks into the camera and says the line from audio 1; at the end she raises her eyebrows and laughs briefly. The camera sways slightly with her movements. No music, no on-screen text.
Price in coins
A video prompt formula
Write the blocks in this order and skip the ones you do not need: a clip without dialogue or references needs only shooting style, scene, action, camera and sound.
| Block | What to write | Example |
|---|---|---|
| Reference roles | number and role of each file | image 1 is the hero's face, video 1 is the camera move |
| Shooting style | one line about the look | handheld camera, documentary footage |
| Scene | place, time, light, who stands where | kitchen at night, light from the fridge, the woman at the table |
| Timed action | steps in order or with timecodes | 0-3 s opens the door, 3-5 s freezes |
| Dialogue | who speaks, how, exact words in quotes | whispering, fast: "He's here." |
| Camera | shot size, one move, speed | medium shot, slow dolly in |
| Sound | specific sounds, ambience, music separately | steps on gravel, no music |
| Ending and exclusions | how the scene ends, what must not appear | the door is closed; no extra people |
One or two characters with one or two actions per shot hold best. Prompt length follows the scene: do not squeeze a detailed scene into a couple of lines.
Timed action and shot changes with timecodes
Write action as a chain of visible steps: smiles, steps forward, stops smiling, looks back, freezes. Describe physics in words: "the ball bounces down the aisle, each bounce lower". Count props that must not duplicate: "exactly one cup in her hand".
Several shots in one generation are written as consecutive timecoded segments with no gaps or overlaps: 0-2 s ..., 2-5 s ..., 5-8 s .... Each segment has its own framing, action and camera, and together they add up to the clip length. This works well on Seedance 2.0 and 2.5, Wan 3.0 and MiniMax H3; on Seedance 2.5 and Wan 3.0 the timeline holds for up to 30 seconds.
Shots do not have to be equal: a detail or close-up takes less time, a wide shot or an action takes more. The framing and what happens in the shot decide its length.
When to split a scene into separate generations
Keep the shots of one scene in one generation when the place, time and characters are the same and everything fits the model's duration. A prop or place described only in text stays the same only within one generation.
Split into separate generations when a shot needs its own retake, the scene changes or the total length exceeds the model's limit. Then attach the same character, prop and location references to every generation and describe the current state again: generations do not remember each other.
A frame makes a clean seam: the last frame of the finished clip becomes the start frame of the next one. End each part on a settled state, a held pose or a finished action, and the cut will be invisible.
Dialogue and sound in the prompt
Each line: who speaks, tone, pace, emotion, then the exact words in quotes. The spoken language is set by the text of the line; video models have no language switch. Write numbers in speech as words.
Allow about three seconds for a short line. If the speech rushes, add time or spread the text across shots.
Name each sound together with its action: a glass hits concrete, steps on gravel. Write ambience and music separately; if you want no music, say "no music".
Voice from a file: Seedance 2.0 and 2.5 and MiniMax H3 accept audio together with a photo or video, Wan 3.0 also without them. Name the file in the prompt: "speaks with the voice from audio 1". Kling 4.0 can only keep the sound of its reference video; Gemini Omni accepts no audio.
How prompts differ between models
The prompt structure is shared, but models take shots, sound and references differently.
| Model | How to change shots | Sound and voice | References |
|---|---|---|---|
| Seedance 2.5 | timecodes in one prompt, clips up to 30 s | sound switch; audio together with a photo or video | up to 30 photos, 10 videos, 10 audio clips; give each a role by number |
| Seedance 2.0 | timecodes, clips up to 15 s | sound switch; audio together with a photo or video | up to 9 photos, 3 videos, 3 audio clips, no more than 12 files in total |
| Kling 3.0 | Multi-shot: up to 5 scenes, each with its own prompt up to 500 characters, 3-15 s in total | sound is always on in Multi-shot | elements: up to 3, each with 2-4 photos; name a character the same way in every scene |
| Wan 3.0 | timecodes, clips up to 30 s | the line can come from audio, even without a photo | up to 10 photos, 5 videos, 5 audio clips |
| MiniMax H3 | timecodes, clips up to 15 s | state the tone directly: "she speaks in a whisper" | up to 9 photos, 3 videos, 3 audio clips; give a frame or character photo for realism |
| Veo 3.1 | one short shot: 4, 6 or 8 s | accent and timbre next to the line | Lite and Fast take up to 3 photos, Quality takes text or a frame only |
Keep a Kling 3.0 prompt short and concrete: one or two characters, a simple action, one camera move. On models not in the table, test timecodes with one try first.
The camera in a prompt
Give the angle and shot size first: low angle, overhead, close-up, wide shot. Then one move with a speed: slow dolly in, smooth pan left. Several moves without an order break the shot; if you need a chain, place it on the timeline.
If the camera must not move, write "static camera, no zoom". The model may put a named device into the frame: instead of "drone shot" describe the path, high above the field, slowly forward.
When words cannot set the camera path, attach a video with the move you need as a camera reference on a model that accepts video: Seedance 2.5, Wan 3.0, MiniMax H3 or Kling 4.0.
Which model to use
FAQ
How do I write a prompt for an AI video?
Describe in order the reference roles, shooting style, scene, timed action, dialogue, camera, sound and end state, then the exclusions. The model, duration and aspect ratio are set in the settings.
What language should a video prompt be in?
You can write in your own language; the spoken language in the clip is set by the language of the quoted line. If the model misreads a phrase, rewrite it simpler and shorter.
How do I get several shots in one video?
Write them as consecutive timecodes, 0-2 s ..., 2-5 s ..., with framing and action for each. This works on Seedance 2.0 and 2.5, Wan 3.0 and MiniMax H3; in Kling 3.0 shots are set as Multi-shot scenes.
Why does the AI ignore part of my prompt?
Usually the prompt has competing actions or directions, such as a static camera and an orbit at once. Keep one or two actions per shot, give every reference a role and remove contradictions.
Where can I get a ready video prompt?
Take an example from this page, swap in your own character, place and line, and keep the structure: shooting style, timed action, sound and exclusions.