The promptlibrary

Three things make an establishing shot read as cinematic rather than as stock footage: a stated altitude, a single motivated light source, and one atmospheric element. Naming the drift speed keeps the model from adding the fast orbiting move it defaults to. The explicit 'no text' matters — generative video will happily invent signage and it never survives a client review.

Scale needs a reference. A figure placed deliberately at the lower third does more work than any adjective about size. 'Ceiling visible' stops the model cropping into a mid-shot, which is its habit when asked for interiors.

Time-passage cutaways depend on the frame not moving. Saying 'identical camera position' and then listing what must stay constant is far more reliable than asking for a time-lapse, which tends to produce two unrelated images.

Generative video over-moves by default. Giving the move a real-world speed — a foot per second — and then explicitly ruling out zoom and shake is what keeps it restrained. 'Background compressing' asks for the optical effect of a long lens, which reads as a dolly rather than a digital push.

There is a difference between handheld and shaky, and models conflate them. Describing an operator — their distance, their position, the fact they are correcting — produces motion with intent behind it. 'No jitter' rules out the buzzing artefact these tools add when they hear handheld.

Reveals are the cheapest way to add production value, and models handle them well when the foreground is named as a specific object rather than described as 'a foreground element.' Stating 'continuous single take' prevents an invented cut halfway through.

The most common failure in generated portraits is soft, even, sourceless light — it reads as a render. Naming one source, its position and its quality, then explicitly refusing fill, is what produces a photographed look. 'Shadow side falling to near black' is the instruction that does the most work.

Practical lighting is what makes an interior feel like a real place. The key phrase is 'falling off naturally with distance' — inverse-square falloff is the thing models skip, and its absence is why generated interiors look flat and evenly bright.