A video prompt describes change over time. Static descriptions produce a slideshow.
Video models extend an image prompt with a dimension the image models do not have: time. A good video prompt says what moves, how the camera behaves, and what the shot looks like at the start and the end. A description that only names objects produces something that drifts or sits still, because nothing in the prompt asked for change.
Start with one shot
Write one shot, not a scene. "A cyclist turns a corner on a wet street, camera tracking alongside" is a shot. "A day in the city" is a film, and the model will produce a low-information montage. If you need several shots, generate them separately and cut them together; a single prompt asked to cover a sequence will rush every part of it.
Describe motion explicitly
Name the subject's movement and the camera's movement separately. "The subject walks toward the camera" and "the camera pushes in slowly" are two different instructions and both matter. Keep them compatible: a slow push-in on a running subject reads as a mistake. Motion verbs should be simple and physical — walks, turns, lifts, pours, drifts — because the model grounds them better than abstract ones.
Time is a constraint
Most video models generate a few seconds. Write for that length: one action, one camera move, one beat of change. A prompt that describes a three-act story in five seconds produces something that looks broken. Say the duration when the tool allows it, and design the shot so the end state is visibly different from the start state.
Example: "Five-second shot: a hand pours water into a glass on a wooden table, camera slowly pushes in, soft afternoon light from a window, shallow depth of field, no text. Start with the glass empty; end with it full."
Keep continuity honest
Video exposes inconsistency: a jacket changes colour, a background object moves, a person's face shifts between frames. Reduce the number of things that can drift. One subject, one location, one lighting condition. If you are using an image-to-video mode, the reference image does most of the continuity work, so describe the motion rather than re-describing the scene.
Sound and text are usually separate
Unless the tool generates audio, do not describe it. Dialogue captions and on-screen text are better added in an editor, where you can control timing and legibility. A prompt that asks for a spoken line in a silent model wastes tokens and can distort the visual result.
Review the first second
The opening frame is what a viewer judges. If the first second is a blur or an empty frame, the shot fails regardless of what happens later. Generate two or three variants, keep the one whose first frame is already a composition, and note the prompt version that produced it.
Checklist
- Is this one shot with one action?
- Are subject motion and camera motion both stated?
- Does the end state differ from the start state?
- Have you removed anything that can drift?