AI filmmaking
The Same AI Video Prompt, Weak and Cinematic
"Cinematic" is the most common word in AI video prompts and one of the least useful. Here is what people mean by it, written out as the specific decisions it stands in for.
The starting point
A man walking down a street at night.
Nothing is wrong with it. It will produce something. But every decision that makes an image look considered has been left to chance: where the camera is, how far away, what the light is doing, what the lens does to the background, whether anything moves.
Pass one: say where the camera is
A low-angle medium shot of a man walking down a street at night.
Shot size and angle are the first two things a director decides and the first two most prompts omit. A low angle makes him larger than the viewer; a high angle would make him smaller. This single change does more than any adjective, because it determines what the frame means rather than what it contains.
Pass two: say what the light is doing
A low-angle medium shot of a man walking down a street at night, lit from behind by a shop window, his face in shadow.
"At night" is a time, not a lighting plan. Light has a direction, a source and a quality, and naming them is most of what separates a flat image from one that looks photographed. Backlight separates him from the background and hides his face — which, conveniently, also hides the facial drift that plagues multi-shot work.
Pass three: say what the lens does
A low-angle medium shot on a 35mm lens, a man walking down a street at night, lit from behind by a shop window, his face in shadow, background softly out of focus.
Focal length and depth of field are the two properties people are usually reaching for when they type "cinematic". A long lens compresses and isolates; a wide lens exaggerates space. Shallow focus tells the viewer where to look. Naming them gets you the look directly instead of hoping the word summons it.
Pass four: say what moves
A low-angle medium shot on a 35mm lens, tracking slowly backwards as a man walks down a street at night, lit from behind by a shop window, his face in shadow, background softly out of focus.
This is video, and camera movement is the property that still images do not have. A static frame, a slow push, a track alongside — each reads completely differently. Unstated, you get whatever the model feels like, which is frequently a drifting wobble that marks the clip as generated.
Pass five: cut what is not earning its place
Read it back and remove anything that does not change the picture. Long prompts are not better prompts; models drop the ends of them, and the detail you cared most about is often the one that goes.
| Keep | Cut |
|---|---|
| Shot size and angle | "Cinematic", "masterpiece", "8K", "award-winning" |
| Light direction and source | "Beautiful lighting" |
| Focal length, depth of field | "Professional camera" |
| Camera movement | "Dynamic" |
| Subject and action | Repeating the subject three ways |
The left column describes decisions. The right column describes hopes. Only one of them is actionable by a model.
Change one thing at a time when you are learning a tool. Rewriting four parts at once and liking the result teaches you nothing about which part did the work — and that is how most prompt folklore gets made.
The order matters
Put the shot size first. Models weight the beginning of a prompt more heavily, and shot size governs framing, which governs everything else. Then light, then lens, then movement, then subject.
Writing in that order also happens to be the order a crew would make the decisions, which is not a coincidence — the prompt is a shot description, and shot descriptions had a grammar long before any of this existed.