One subject, one action, one camera move.
If you have written prompts for image generators, you know the basics: say what is in the frame, add a style, mention the lighting. Video prompts look similar but fail differently. A picture only has to be good once. A clip has to stay good for several seconds while things move, and the model must decide how. Leave those decisions open and you get drifting faces, wandering cameras and actions that never finish.
Why video prompts are not image prompts
Three things change when you add time.
Time. Your prompt is closer to a one-line shot description than a caption. "A lighthouse at dusk" says nothing about what happens. "A lighthouse at dusk, the beam slowly sweeping across the water" gives the model something to animate.
Motion. Every moving thing needs a speed, a direction and an amount. Supply them, or accept the model's default, often a slow, vague drift.
One action per clip. Short clips have room for one clear beat. Ask for a person to walk in, pick up a cup, laugh, turn and leave, and most models blur it into mush. Treat each clip as one shot in an edit and join shots for longer scenes.
A simple structure to start from
A reliable template has five parts, in roughly this order:
Subject: who or what, with two or three concrete details.
Action: the one main thing that happens.
Camera: how the camera behaves and how close it is.
Light: time of day, direction, quality.
Style and pace: the look (documentary, soft film grain, clean commercial) and the tempo (slow, steady, brisk).
Example: A ceramic teapot on a wooden table, steam rising in a gentle curl, slow dolly in from a medium shot, soft window light from the left, calm documentary feel.
When a result disappoints, check which of the five you left out.
Camera vocabulary that models respond to
Ordinary cinematography words tend to work well. Keep to one camera instruction per clip.
Static / locked-off: the camera does not move. Use it when you want the subject's motion to carry the shot.
Dolly in / dolly out (or push in / pull back): the camera moves toward or away from the subject. Add "slow" unless you want urgency.
Pan left or right: the camera turns horizontally on a fixed point. Tilt up or down: the same, vertically.
Tracking shot: the camera travels alongside a moving subject. Say "tracking from the side" or "tracking from behind" to fix the angle.
Orbit / arc: the camera circles around the subject.
Handheld: slight natural shake, good for a documentary or candid feel.
Crane / rising shot: the camera lifts up and away, often used to reveal a landscape.
Shot size matters as much as movement: "wide shot" shows the setting, "medium shot" frames a subject from the waist up, "close-up" fills the frame, "extreme close-up" isolates a detail. Lens words such as "shallow depth of field" or "wide-angle" shift the look without adding motion.
Describing motion without confusing the model
Choose one main action. "A dancer spins" beats "spins, jumps, bows and exits."
Use speed words. "Slowly", "steadily" and "quickly" are cheap and effective.
Avoid contradictions. "A static camera with a fast orbit" or "the boat is perfectly still as it races across the bay" gives the model two opposing instructions, and it will often compromise on neither.
Describe the end state. Telling the model where things should end up gives motion a direction: "the door swings open and stays open", "the fog lifts to reveal the hillside." This works better than "the door moves".
Prefer concrete verbs. "Moves" is vague; "pours", "glides", "drifts" and "sweeps" each carry a picture.
Light and time of day
Light sets mood and also helps a clip look like one place in one moment.
Name the time of day (golden hour, blue hour, overcast noon, night), the direction (backlit, side-lit, window light from the left) and the quality (soft, hard, diffused, hazy). Practical sources help too: "warm lamplight", "neon reflections on wet pavement", "candle glow". If the light should change during the clip, say so as the single motion event: "the sky slowly darkens from orange to deep blue."
Image-to-video: describe only what changes
When you start from an image, the model already knows what is in the frame. Your prompt should describe the change, not the picture.
Do not re-describe the whole image. Repeating every detail invites reinterpretation, and the subject can drift from your start frame.
Name the movement and the camera. "The woman in the red coat turns her head toward the window; slow push in" is enough.
Stay consistent with the start frame. If the image shows overcast daylight, do not ask for golden hour. If the subject is facing away, do not ask for a close-up of their face. Contradicting the frame is one of the most common reasons image-to-video clips warp or cut.
Ask for plausible motion. Small movements (wind in hair, steam, a slow walk) hold together better than dramatic transformations.
Patterns that cause trouble
Too many actions in a few seconds.
Conflicting camera moves, such as "zoom in while pulling back".
Vague words with no visual meaning ("epic", "amazing").
Style words piled up from several genres, which flattens the look.
Readable text on signs or screens, which many models still garble.
Iterating cheaply
Treat the first generations as sketches, not finals.
Draft short and small. Use the shortest duration and lowest resolution the model offers while you test the idea. If the composition and motion work at 480p for a couple of seconds, scaling up is a safer bet.
Change one variable at a time, or you cannot tell which change helped.
Keep a log. A prompt plus a one-line note shows which words reliably do something.
Compare models on the same prompt. Different models read the same words differently. Services such as SpicyAPI let you try several video models side by side and show the price before each run, which makes this kind of comparison practical.
Three before-and-after rewrites
1. Cyclist on a seafront
Before: A cyclist riding along the sea, cinematic and beautiful, camera moving around, fast, sunset.
After: A cyclist in a yellow jacket rides steadily along an empty seafront promenade, tracking shot from the side at medium distance, low golden-hour sunlight from behind the camera, calm waves in the background, smooth and unhurried pace.
What changed: one action, one camera move, a named light direction, no vague adjectives.
2. Coffee being poured in a cafe
Before: Coffee in a cafe, pouring, nice atmosphere, lots of movement.
After: Close-up of a white ceramic cup on a wooden counter. Dark coffee pours in a thin, steady stream and fills the cup to the brim, steam curling upward. Static camera, shallow depth of field, soft warm window light from the left.
What changed: the action has an end state (fills to the brim), and the locked camera lets the pour carry the shot.
3. Sailboat at sunset
Before: A sailboat at sunset, epic, the camera flies around it and zooms in while the boat sails away.
After: Wide shot of a small white sailboat gliding slowly across calm water toward the horizon, slow crane shot rising gently above it, low orange sun reflecting on the surface, quiet and steady mood.
What changed: the contradictory camera moves (flying around, zooming in) became a single rise, and the boat's motion matches the camera's pace.
Conclusion
Good video prompts read like short shot notes: one subject, one action, one camera behavior, a clear light and a pace. When a clip goes wrong, look for contradictions and for decisions you left to chance. Draft small, change one thing at a time, and let the clips you keep build your vocabulary.