Cinematography

Cinematography for generated video

The camera and lighting decisions, with what each one does to the viewer — because a video model will happily render a technically correct shot that lands emotionally wrong.

Full menus (shot sizes, angles, all camera moves, optics, composition, lighting setups, palettes, film stocks, styles): prompt-expander/references/visual-language.md. This skill is the decision layer on top of them.

1. Shot, lens and aperture by purpose

Pick the row that matches the job of the shot, not the one that sounds most cinematic.

A wide shot has two jobs, and the frame decides which. Subject large in frame with the space around them reads as context. Subject small and swallowed by the space reads as isolation. Both are wides; only the second one is lonely. Say which you want.

Lens to feel: ultra-wide (≤24mm) expands space and exaggerates depth · wide (28–35mm) holds subject and environment · normal (50mm) is human-eye natural · telephoto (85mm+) compresses depth and flatters portraits.

Depth of field:

2. Camera movement — pick for the effect, name exactly one

Rules that are not optional:

• Exactly one move per clip. Two named moves, or a vague one, makes the model default to a near-static drift.
• Some single moves have two-word names, and they are still one move. dolly zoom, slow push-in, slow pull-out, whip pan, crash zoom. The rule above is about two independent moves ("pan while craning down"), never about word count.
• Focus and duration are not moves. rack focus (reveal), long take (suspense) and slow motion describe something other than where the camera goes. A clip carrying only one of them still has no camera move and still needs one, or static shot.
• A locked frame must be stated: static shot, optionally static shot, no camera movement. Omitting the camera line does not give you a still frame.
• Never write "locked" at all — not "locked tripod", not "locked shot", not "locked off". "Locked tripod" is misread: the model adds motion anyway and has been observed rendering a literal tripod in frame. static shot says the same thing and is safe.
• Never combine contradictory terms ("drone shot zooming into close-up"). Pick one perspective.

2b. Angle — the cheapest way to change how someone reads

Where the camera sits relative to the eyeline changes who has power in the frame, before a word is spoken. An unstated angle renders eye-level: fine as a decision, a defect as an accident.

The trap is using the angle that contradicts the beat — a low angle on someone being humiliated makes them look like they are winning. Pick the angle for who should have power in that second, not for what looks impressive.

3. The over-the-shoulder rule (stops the wrong face appearing)

In a dialogue scene between two characters, the listener is where generated clips go wrong — the model renders their face, badly, in the corner of frame.

• Close-ups (85mm, f/1.8): only the speaking character is visible. Do not mention the other one at all.
• Medium (50mm/35mm) and wider: show the back of the other character's head at the edge of frame. Describe them only from behind — "back of head", "back of shoulders", "shirt collar seen from behind". Never describe their face, expression, eyes or glasses.

Wrong — this renders a face:

Man frozen in shock visible at frame edge left in soft focus. Balding with semi-rimless glasses. Completely stunned expression, gaze fixed on her.

Right — this renders an over-the-shoulder:

The back of his head and shoulders visible at frame edge left in soft focus. His balding crown and the back of his light blue-grey shirt collar seen from behind. He is facing away from camera toward her.

Same idea generalises: in a close-up, keep exactly one character in focus. Others are soft-blurred, off-screen, or speaking from off-screen.

4. Lighting

Colour temperature:

Standard lighting line:

Lighting: [quality] [source] at [angle]. [detail]. Color temperature [X]K. [grade]. 24fps.

Examples:

Lighting: Bright harsh fluorescent office lighting from multiple ceiling panels.
Color temperature 4100K cool white. Very slight green cast typical of office
fluorescents. 24fps.

Lighting: Warm natural afternoon sunlight from a window at camera-left. Soft golden key
on the face. Color temperature 5000K. Warm cosy domestic grade. 24fps.

One lighting motif per clip. And never write a mid-clip lighting transition into the look line — a light that changes is a cut, not a description.

5. Framing for the aspect ratio

9:16 (vertical): centre the subject · leave headroom at the top for platform UI · keep everything important inside the centre 80% of frame · text overlays read best at top and bottom.

16:9 (landscape): rule of thirds · more room for environmental context · better for wide shots and scenery.

If the interface already sets the ratio, do not also write it in the prompt text.

5b. Composition — where the subject sits inside the frame

In 9:16 you buy centre framing whether you want it or not. Section 5 tells you to centre the subject for platform reasons — under-the-UI safety, not meaning — and the frame still reads as control. To signal instability in vertical you have to work for it: a dutch angle, or the subject pushed hard to one side against deliberate negative space.

Frame-within-a-frame is easy to use by accident. Anything shot through a window, a doorway or a windscreen is one, and it says trapped even when the scene is about freedom. Check that the meaning is the one you want before shooting the whole film through glass.

6. Where the shot fits the story

One clip is one beat. Choose the coverage to match:

Standard sequence: establishing wide → medium → detail/close. Reveal, then specify.

Identity holds best in calm close and medium shots and drifts in fast action and wide shots — so put identity-critical beats in calm coverage and let action shots run loose. See identity-and-likeness.

7. Emotion -> technique (reverse index)

Every table above runs technique -> effect. This one runs the way a writer actually arrives: knowing how the beat should feel and needing the camera that delivers it.

[FIELD PRACTICE]. These pairings are live-action editorial convention, not vendor documentation, and nothing in this repo yet proves a video model renders "low angle" as authority rather than merely as a low camera. Use them to choose, then let a shoot confirm. Source and the gap analysis they came from: docs/camera-agent/emotion-technique-map.md.