Castaly · Blog · Guides
How to Write AI Video Prompts: Seven Parts, Shown on Real Clips
The seven-part prompt structure behind our sample clips: shot type, place, subject, event, camera, light and finish, each shown with the exact prompt and its result.
A video prompt has more to do than an image prompt. Besides what is in the frame, it has to say what happens, in what order, and where the camera is while it happens. Prompts that leave those out let the model decide, and you pay for every guess.
This guide breaks down the structure we use for the sample clips on this site. Every prompt quoted here is the exact text of a recorded generation, and the clip next to it is what came back.
The seven parts of a video prompt
Here is one complete prompt, from a 5-second Seedance 2.0 clip:
One continuous take, no cuts. An empty indoor swimming pool at night, tiled walls and underwater lights. A swimmer in a chrome-finish suit stands on the edge and lifts both arms; the water answers by leaving the pool, rising as a single rotating column of turquoise that twists up past the roof beams and holds there, the emptied tiles glowing underneath. The camera starts low at the pool edge, cranes up alongside the column and tilts to look down into the drained basin. One hard cyan key from the underwater lights, warm sodium spill from the high windows. Real weight and inertia, spray and refraction, no text.
It has seven parts, in this order:
| Part | In the prompt above |
|---|---|
| 1. Shot type | One continuous take, no cuts. |
| 2. Place and time | An empty indoor swimming pool at night, tiled walls and underwater lights. |
| 3. Subject | A swimmer in a chrome-finish suit stands on the edge |
| 4. Event | lifts both arms; the water answers by leaving the pool, rising as a single rotating column… |
| 5. Camera | The camera starts low at the pool edge, cranes up alongside the column and tilts to look down… |
| 6. Light | One hard cyan key from the underwater lights, warm sodium spill from the high windows. |
| 7. Physics and finish | Real weight and inertia, spray and refraction, no text. |
The sections below take them one at a time.
1. Say what kind of shot it is
Most of our recent prompts open with One continuous take, no cuts. A video model is free to cut between angles, and a cut in the middle of a 5-second clip often wastes it. Stating the shot type first asks for one unbroken camera move.
For vertical clips we add the format to the same sentence, One continuous take, vertical. For a stylised piece, the style goes first instead: Stop-motion claymation style: a small orange clay fox sneaks through a miniature felt forest…
2. Set the place in one clause
An empty indoor swimming pool at night, tiled walls and underwater lights. One clause, three concrete details, no adjectives about mood. The model gets a location, a time of day and the light sources it will need later.
3. Give the subject one handle
Name one subject and give it a single strong visual handle: a swimmer in a chrome-finish suit, a figure in an acid-green vinyl coat, a woman in a cobalt-blue structured trench coat. A distinctive colour or material makes the subject easy to follow through the motion, and easy to describe again in the next shot.
For a recurring character, repeat the description word for word. These two Kling 3.0 clips open with the same subject description:
A woman in her late twenties with a short dark bob, a small scar above her left eyebrow and a worn olive field jacket …
The bob and the jacket carried over between the clips. The scar did not stay put: in the market it sits on her cheek. A repeated description keeps the broad look; for exact details, use a reference image of the character instead. The image-to-video guide shows how.
4. Write the event as cause and effect
The event is the part people under-write. Describe a trigger and what it sets off: lifts both arms; the water answers by leaving the pool. A semicolon between the action and the reaction keeps the order clear.
Size the event to the length. A 5-second clip has room for one action and its consequence. A 10-second Seedance 2.5 take can carry a whole sequence:
One continuous take, no cuts. Rush-hour subway car, commuters staring at their phones. A paper cup tips over; the spill does not stop but swells within seconds into a full waterfall bursting out of the car, blowing the doors open and flooding the platform with seats, newspapers and phones while the commuters keep scrolling, unbothered. …
5. Direct the camera from start to end
Give the camera a starting position, a move and an end point: The camera starts low at the pool edge, cranes up alongside the column and tilts to look down into the drained basin. Ending on a specific view gives the last second of the clip something to land on.
Use ordinary film words. Dolly, crane, tilt, push in, handheld, tracking shot and Steadicam following behind all appear in our recorded prompts. Lens choices such as wide lens or wide-angle lens help when the framing matters.
6. Name the light sources
One hard cyan key from the underwater lights, warm sodium spill from the high windows. Naming where each light comes from, and its colour, gives consistent lighting that matches the place you described in part 2. Cool grey-blue palette broken by one warm sodium lamp works the same way.
7. Close with physics and finish
The last sentence carries the material qualities: Real weight and inertia, spray and refraction. Then the exclusions. Almost every prompt we record ends with no text: night streets, shops and screens invite a model to invent signs and captions.
Prompting from a first frame
When the clip starts from an image, drop parts 2 and 3: the model can see the place and the subject. Spend the prompt on the event and the camera, as in this Wan 2.7 I2V prompt:
One continuous take, no cuts. The ribbon of film keeps unwinding and speeds up until every frame is lit, tightening into a neon helix that lifts off the deck and climbs past the parapet into the dusk sky…
Writing in Chinese
Prompts do not have to be in English. This Seedance 2.5 request was written in Chinese and follows the same structure, subject, camera, event and an ending condition:
探索潜航器缓慢驶向海底环形遗迹，摄影机从侧后方平稳跟随。灯束扫过发光珊瑚，探险者保持与潜航器的相对位置；在拱门内新空间显露时结束。保持设备外形、光线方向和水中运动连贯。
That is: an exploration submersible moves slowly toward a ring-shaped ruin on the sea floor while the camera follows steadily from behind and to the side; the beams sweep across glowing coral; the shot ends when a new space opens up inside the arch; keep the vehicle's shape, the light direction and the motion through water consistent.
Draft cheaply, then render
A long prompt is cheap to test. The same text runs on Seedance 2.0 Mini at 480p for about $0.09 per 5 seconds, and you can open it there with one click:
Try the swimming-pool prompt as a 480p draft →
Once the action and camera read the way you want, render the final on the model and resolution you chose. Treat the draft as a check of the idea; a different model will render the details its own way.
FAQ
How long should a video prompt be?
Long enough to cover the seven parts. Our 5-second prompts run to about 100 words, and the 10-second ones a little longer. Length costs nothing: the price depends on the model, resolution and duration, never on the prompt.
Should I write "no text" in every prompt?
If you do not want lettering in the frame, yes. Scenes with streets, shops or screens invite invented signs and captions.
How do I keep a character the same across clips?
Repeat the same description word for word, and for exact features use a reference image of the character. A repeated description alone kept the hair and clothing in our clips but moved a small scar.
Does the prompt language change the price?
No. A Chinese prompt and an English prompt for the same settings cost the same.