The five-slot video prompt checklist
If the same prompt keeps producing a different video every time, the prompt is missing information. Each missing field becomes a field the model fills in for itself.
Why prompts feel like a lottery
Language models write video prompts reliably for two things — the subject and the scene — and systematically omit three others: what the subject isdoing, how the frame is composed, and how the camera moves. Every dimension you leave unspecified is a dimension the model gets to choose for you. That is the mechanism behind the "lottery" feeling: it is not randomness, it is underspecification.
Filling all five slots is the single most effective change we have made to generation hit rate.
The five slots
| Slot | What it controls | What to write |
|---|---|---|
| 1. Subject | Who or what the frame is about | Age, build, clothing, skin tone for people; colour and material for products. With multiple subjects, state which is primary. |
| 2. Subject motion | What the subject does, in time order | "First she does X, then Y, then Z." Interaction between subjects counts here. |
| 3. Scene | Where it happens and under what light | Location, time of day, and lighting (natural window light, side light, neon). Add environmental motion such as passing crowds or traffic. |
| 4. Spatial / composition | How things are arranged in frame | Shot size (close-up, medium, wide), subject position (centre, off to the right), foreground/midground/background, and how the layout changes over the shot.This is the slot most often missing, and the usual cause of strange framing. |
| 5. Camera | How the camera moves | Height, angle, focus behaviour, stability, and the movement itself (push in, pan, tilt). "Slow push from 1m to 40cm, steady handheld feel, focus locked on the face." |
Before and after
Weak — covers roughly two of the five slots:
A woman uses a facial mist, her skin becomes dewy, the background is an office.
Composition, camera and motion order are all missing, so the model chooses them. This is where wrong-looking faces, odd framing and inexplicable camera moves come from.
Strong — all five slots filled:
Subject: woman in her early 30s, black business attire, dry flaky skin, holding an unbranded white mist bottle.
Motion: first turns a dry cheek toward the camera, then closes her eyes and sprays toward her face, mist fills the air, she pats her cheek and smiles.
Scene: modern office desk by day, natural light through a window, laptop and papers on the desk, blurred colleagues moving behind.
Spatial: opens on a medium shot with the subject slightly right of centre; pushes in to a facial close-up as she sprays, mist particles in the foreground.
Camera: slow continuous push from about 1m to 40cm, steady handheld feel, focus locked on her face.
Self-test while writing
Cover the other four slots and read only one. Can you picture the shot from that slot alone? If not, that slot is not written yet.
Pre-submit checklist
- Subject: appearance, skin tone, clothing all stated? Primary subject clear?
- Motion: does it have a time order (first X, then Y)? Any interaction described?
- Scene: location and lighting both given? Environmental motion included?
- Spatial: shot size, subject position, foreground/background, and how the layout changes? Check this one twice.
- Camera: height, angle, movement and focus specified? Clearly separate from slot 4?
- Consistency: does it carry over subject, scene and lighting from the previous shot?
- On-screen text: is the frame kept free of text, UI and screen recordings? Generate text as an overlay afterwards instead — models render lettering unreliably.
- Length: is the prompt in the model's comfortable range for a primary shot?
- Deliberate variation: when running a batch for variety, which single slot are you intentionally leaving open? Vary one, not all five.
Reducing variance further
A text prompt leaves all five slots to be interpreted. A start-frame image locks three of them — subject, scene and composition — immediately. That leaves only motion and camera for the prompt to control, which is a large drop in variance. If hit rate matters more than variety, lock with an image first and write a prompt for the remaining two slots.