Getting Started with AI Video Prompts: How to Write Camera Moves, Actions and Duration

An image prompt describes a single moment; a video prompt describes a span of time. To generate an image, you only need to say what is in the frame. To generate a video, you also need to say how the frame changes: what the subject does, how the camera moves, and how light and atmosphere carry through. Video prompts are therefore usually longer, and read more like a director's script written for the camera.

The most common beginner mistake is feeding an image prompt straight into a video model: piles of style tags and scene elements, but no action and no camera language. The result is often a subject frozen in place or a wildly drifting camera. Keep one simple idea in mind: when writing a video prompt, think of yourself as a director giving instructions to the camera, not a painter describing a still painting.

A complete video prompt is usually organized in this order: subject, action, scene, camera movement, atmosphere. The subject is the main character of the frame — be specific, for example “a young woman in a beige trench coat” rather than “a woman”. The action is what the subject does on camera, for example “turning and walking down the stone steps”. The scene sets the place and time, for example “a stone-paved lane in an ancient river town at dusk”.

Camera movement says how the camera moves, for example “slow push-in”. Atmosphere covers light, color and mood, for example “warm sunset backlight, fine drizzle in the air, cinematic”. Write one short clause for each of the five elements, join them with commas, and you have a solid starter prompt. Before submitting, run a five-question check in order: Is there a subject? What is it doing? Where is it? How does the camera move? What are the light and mood? When all five are covered, your success rate improves noticeably.

Camera-move terms are the jargon of video prompts, and mainstream models understand the common English terms quite reliably. For moving closer: slow push-in and dolly-in bring the camera gradually toward the subject, great for amplifying emotion. For pulling away: zoom out and pull back reveal the wider environment. Pan left and pan right swing the view sideways; tilt up and tilt down move it vertically.

To circle around the subject, write orbit, a favorite for close-ups. To follow a moving subject, write tracking shot or follow shot. For the slight shake of documentary-style footage, write handheld. For a ground-hugging push, write low-angle tracking; for an overhead view, aerial drone shot. Use only one or two camera terms per clip: the more movement instructions you stack, the more confused the model gets.

Mainstream text-to-video tools commonly generate five to ten seconds per run — only enough for one “action sentence”. So the first rule of writing actions is: describe one coherent action at a time. “She picks up a teacup, blows on it gently, and takes a sip” works well; “She drinks tea, puts down the cup, walks to the window and opens it” is too crowded — the model will likely swallow the later actions or render ghostly figures.

The second rule: give the action a start and an end, and describe the process as concretely as possible. “She slowly raises her head and gazes toward the distance on the right of the frame” is far more controllable than “she is looking at the view”. The third rule: attach speed and emotion words to the action — “slowly”, “lazily”, “suddenly” — which clearly shape the pacing of the result. Favor everyday, continuous, moderate movements for people; complex fights, multi-person interaction and rapid direction changes remain a universal weak point across models, so avoid them when you can.

Duration and aspect ratio are generally not written into the prompt itself; they live in the tool's generation settings, usually as duration and aspect ratio options near the prompt input. Common single-run duration tiers sit around five to ten seconds, and some tools also offer automatic timing or multi-clip continuation; common ratios include 16:9 landscape, 9:16 portrait and 1:1 square.

Portrait suits short video for phones; landscape suits cinematic scene-setting. You may echo the settings with a framing line in the prompt such as “vertical composition, subject centered”, but the actual duration and ratio should follow the generation parameters. Also, the same model usually offers several quality tiers, and the higher-quality tier is often slower and costs more — check the current tier before generating so you do not waste attempts.

In image-to-video, a still image becomes the first frame and the model animates it. The division of labor changes: the image carries the visual content, while the prompt only answers “how does it move”. Focus on action and camera work — for example “she slowly turns toward the camera, hair strands lifted by a breeze, camera slowly pushing in” — with no need to repeat the clothing or background details already in the picture.

Before writing, look closely at the original image: the pose, the direction of the gaze, the fall of light. Write motion along those clues so action and picture do not fight each other — if the original shows a figure from behind, do not write that she turns with a smile toward the camera. The quality of the first frame also sets the ceiling of the video: a picture with a clear subject, clean composition and natural light is far less likely to fall apart in motion.

Finally, three complete examples, all written for text-to-video; copy them and swap in your own subject and scene. Example one, an emotional close-up: a young woman in a beige trench coat stands on a wet stone-paved lane of an ancient river town. She slowly turns her head toward the camera, hair strands lifted by a gentle breeze. Warm sunset backlight, fine drizzle drifting in the air. The camera slowly pushes in to a close-up of her face. Cinematic, shallow depth of field.

Example two, a soothing still-life: a white ceramic teacup on a wooden table, thin steam rising from the cup and drifting slowly upward. Morning light slants in from the window, dust motes floating in the beam. The camera slowly pans from the side of the cup to the front. Quiet and soothing, photorealistic. Example three, a night city aerial: an elevated highway under nightfall, headlights streaming into ribbons of light. The camera rises slowly from a low angle and pushes forward, gradually revealing the whole skyline. Neon reflections on the wet asphalt, blue-purple night tones, commercial-film polish.

When copying these, start by replacing only the subject and the scene, keeping the camera and atmosphere parts, and run a few rounds to feel how the model responds before switching to your own combinations. There is no single correct video prompt — the same instruction yields different results on different models and tiers. What matters is building the habit of “subject plus action plus scene plus camera plus atmosphere”; the rest comes from repeated experimentation.