AI Video Prompting Basics: Subject, Action, Camera, Time
Learn the four parts of a video prompt: a stable subject, one clear action, a motivated camera and a timeline, with examples that show why each one matters.
Visual AI Club7분 분량
이 가이드는 영어로 작성되었습니다.
An image prompt describes a moment. A video prompt describes what happens between moments, and that is where most first attempts fall apart: the subject melts, three things happen at once, and the camera drifts without purpose. By the end of this guide you will be able to write a short clip prompt in four parts, subject, action, camera and time, and know what each part protects against.
Why video prompts need a different shape
A video model has to generate many frames that agree with each other. Every frame must show the same person, the same room and the same light, while something changes in a believable way. If your prompt describes a beautiful scene but not what moves, the model invents motion, and invented motion tends to be either too little, a near-still image that slowly breathes, or too much, where everything wobbles at once.
So a video prompt keeps everything an image prompt has, and adds rules for change. A practical way to organize it is in four parts, written roughly in this order. Each part answers a question the model would otherwise answer for you.
| Part | Question it answers | What goes wrong without it |
|---|---|---|
| Subject | Who or what is on screen, and what must stay the same? | Faces, clothing and objects morph between frames |
| Action | What changes, how fast and in what order? | Random motion, or several actions blurring together |
| Camera | Where is the viewer, and does that position change? | Aimless drifting or sudden, unmotivated moves |
| Time | How long, what happens when, and how does it end? | Rushed actions, or a clip that stops mid-gesture |
Subject: describe what must not change
In video, the subject description is less about looking good and more about staying the same. Every detail you name becomes an anchor the model checks across frames. Details you leave vague are free to shift: a jacket that is just dark can turn from navy to black to charcoal over five seconds.
- Name fixed features: hair color and length, clothing items with colors, one or two accessories.
- Count things: exactly one cat, two people, a single red umbrella. Unstated counts drift.
- State the setting once, clearly: a narrow cobblestone street in light rain, shop windows lit warm yellow. Then do not contradict it later in the prompt.
- Use a reference image when you can. Starting from a still you already like gives the model a fixed answer for the subject. The image-to-video workflow guide covers this route.
Longer, well-structured prompts in the library often open with a dedicated subject section for exactly this reason. The Wildlife Documentary prompt, for example, fixes the animal's markings and rules out other creatures before describing any movement.
Action: one main action, with a speed
The single most useful habit in video prompting is to give each shot one main action. A woman walks to the window, pauses and looks out is one connected action. A woman walks to the window, answers the phone, makes coffee and waves to a neighbor is four, and in a few seconds of footage the model will either rush them or merge them into something strange.
Write actions with verbs that carry speed and direction. Walks is vague; strolls slowly toward the camera is specific. Turns is vague; turns her head to the left over about two seconds is specific. Physical cause and effect helps too: wind lifts her hair from behind, the cup steams because the coffee is hot. Models produce more believable motion when the reason for the motion is in the prompt.
A man in a kitchen cooking, dynamic and cinematic, lots of movement.
A man in a gray apron stands at a gas stove and slowly stirs a pot of tomato sauce with a wooden spoon, then lifts the spoon and blows gently on it. Steam rises steadily from the pot. Everything else in the kitchen stays still.
Telling the model what stays still is as useful as telling it what moves. Background people standing still, curtains not moving and the table stays empty all reduce random motion in the frame.
Camera: decide where the viewer stands
Camera language is the part of a video prompt that most changes the feel of a clip. The same action filmed on a locked-off tripod, a slow push-in and a shaky handheld phone tells three different stories. Choose one camera behavior per shot and give it a reason.
- Static or locked-off: calm, observational, lets the action carry the shot. The safest choice when a subject's identity matters most.
- Slow push-in: the camera moves gently toward the subject, building focus or emotion.
- Tracking: the camera moves alongside or behind a moving subject at the same speed, keeping them at the same size in frame.
- Handheld: small, natural shakes that suggest a person holding the camera. Good for vlogs and found-footage looks, distracting for product shots.
- Orbit: the camera circles the subject. Dramatic, but demanding, because the model must invent the sides it cannot see.
Lens and shot size still matter, just as in stills. A 24 mm wide shot shows the whole street and exaggerates movement toward the lens; an 85 mm close-up isolates a face and compresses the background. The guide on camera movements for AI video covers dolly, pan, orbit and crane in detail.
Borrow film's natural motion blur
Film and video have traditionally been shot at 24 frames per second with a shutter of about 1/48 of a second, the so-called 180-degree rule. That shutter speed blurs fast movement slightly within each frame, which our eyes read as smooth, natural motion. If AI footage looks jittery or strangely crisp, asking for natural motion blur, as if shot at 24 fps with a 180-degree shutter, often makes movement feel more like a real camera.
Time: give the clip a beginning, middle and end
Short clips feel finished when they have a shape: something starts, something develops, something settles. Without timing, models tend to spread one action thinly over the whole clip or cut it off mid-gesture. A simple timeline fixes this.
0 to 2 seconds: wide shot of an empty rain-wet street at dusk, streetlights flicker on. 2 to 5 seconds: a cyclist in a yellow raincoat rides in from the left, the camera pans slowly to follow. 5 to 8 seconds: she stops under a streetlight and looks up at the rain; the camera holds still.
- Match beats to real durations. A head turn takes about one to two seconds; walking across a room takes several. Cramming a long action into one second makes it look sped up.
- End on a held moment. A pause, a look, a settled pose. Clips that end mid-action feel cut off and are harder to edit.
- Keep light continuous. If sunset begins in second one, it should not suddenly be midday in second six, unless that is the point.
- Add sound when your tool supports it. A line about ambience, such as rain on pavement and distant traffic, no music, shapes the mood. See sound, ambience and dialogue in AI video.
Putting the four parts together
Here is a full prompt written in the four-part order. Read it and notice that each sentence has a job, and that nothing contradicts anything else.
Subject: an elderly man with a white beard, flat cap and brown wool coat, sitting alone on a green park bench in autumn, yellow leaves on the ground. Action: he slowly unfolds a newspaper, then pauses and looks up as a single leaf drifts past his face. Camera: static medium shot at eye level, 50mm look, shallow depth of field. Time: 8 seconds; the leaf passes at around 5 seconds and he holds his gaze until the end. Sound: soft wind, distant birds, paper rustling, no music.
Once this structure is natural, you can write it as flowing prose instead of labeled parts; the order still helps. For multi-shot sequences, write one four-part block per shot, and plan the sequence first with a shot list for a short AI video.
A ten-minute test to see the difference
Run this experiment with any video tool you have access to. It shows how much each part contributes.
- Write a one-line prompt: a dog on a beach.
- Add a subject block: breed, color, collar, the beach, the weather. Generate again.
- Add one action with speed: trots slowly toward the camera, then sits. Generate again.
- Add one camera behavior: static low angle at the dog's eye level. Generate again.
- Add a timeline: trotting from 0 to 4 seconds, sitting and looking at the camera from 4 to 8. Generate again.
- Line the five clips up and note which addition changed the result most. That is the part to pay most attention to in your next prompt.
