The base template
Four slots, one sentence each.
[Place and surface], [lighting and weather], camera [movement] at [pace]. Audio: [ambience] plus [one specific sound]. A viewer here could [interaction].
A repeatable workflow: four prompt elements, one steering discipline, and the mistakes that cost the most consistency.
Works today with world and video models, and transfers to Genie 4.
Most weak prompts describe only scenery. A world model needs something to simulate, which means motion, sound, and a possible interaction.
If you only have time for one improvement, add motion. A static description gives the model nothing to predict.
Each step is small on purpose. Big prompts drift; small prompts compose.
One sentence on surface, light, and weather. This becomes the anchor you refer back to later.
State the camera move and the pace. "Drifting forward at walking speed" is more useful than "cinematic".
Name two or three sounds. Ambience plus one specific detail reads better than a list of ten.
Describe what is possible in the space. This is what the agent loop plans against when you intervene.
Change one thing at a time, and change it late. Early edits compound drift.
Fill in the brackets, or copy a complete example and adapt it.
Four slots, one sentence each.
[Place and surface], [lighting and weather], camera [movement] at [pace]. Audio: [ambience] plus [one specific sound]. A viewer here could [interaction].
Emphasise materials and slow motion.
[Landscape type] under [weather] at [time of day], [surface detail] catching the light, camera [orbiting/drifting/tracking] at [pace]. Audio: [wind/water/insects] plus [one distinctive sound].
Emphasise structure and light sources.
[Interior or street type] with [architectural detail], lit by [light sources], camera moving [direction] at [pace]. Audio: [room tone or traffic] plus [mechanical detail]. A viewer could [interaction].
Emphasise dialogue and reaction.
A [character description] in [setting], doing [action], camera [framing and movement]. Audio: [ambience], and the character says [line]. They react to [stimulus].
Complete, tested structure.
A windswept coastal cliff at golden hour, tall grass bending in gusts, ocean swell breaking far below, camera slowly orbiting the headland at a steady pace. Audio: surf and wind, plus gulls calling. A viewer could walk toward the edge.
Complete, with dialogue.
A narrow night market aisle under strings of bulbs, steam rising from food stalls, camera tracking forward at shoulder height. Audio: overlapping voices and sizzling oil, and a vendor calls out an order. A viewer could step up to a stall.
Complete, quiet and static.
A dim archive room lined with paper files, one desk lamp, dust drifting in the air, camera slowly pushing toward a single open folder. Audio: quiet page turns and a clock ticking. A viewer could lift a document out.
Complete, with a character.
The interior of a slowly rotating orbital station, sunlight sweeping across the deck every few seconds, a crew member floats past and greets the camera, camera holding steady. Audio: low life-support hum and the click of a handrail.
Each of these maps to a documented characteristic of the model, not to prompt superstition.
With no motion, there is nothing to simulate and nothing to steer. Add a camera move and one thing that changes.
Google reports drift after many edits. Make fewer, bigger changes rather than a stream of small ones.
An intervention near the start compounds through the whole session. If you know a change is coming, put it in the initial prompt.
Signage and labels are unreliable unless explicitly prompted. Do not design a scene around readable text.
Real places are not reproduced precisely. If exact layout matters, do not rely on this.
Ordinary interactions hold up better than unusual ones. Test before you build a workflow on a specific interaction.
Practical details on prompting and steering.
The prompt composer appends your framing and length choices automatically, so you can focus on the description.