A good world-model prompt describes four things: the place, the motion, the sound, and a possible interaction. Most weak prompts describe only the first. That is the single biggest difference between a prompt that produces a watchable scene and one that produces something you can steer.
Why world prompts differ from image prompts
An image prompt describes a still. A video prompt describes a shot. A world prompt has to describe something that continues to exist after the description runs out.
That means the model needs to know:
- What the environment is made of, so it can stay consistent
- How it should change over time, so there is something to predict
- What it should sound like, because audio is generated with the scene
- What is possible in the space, because an agent will plan against those affordances
If you omit motion, the model has nothing to simulate. If you omit audio, it invents a default. If you omit interaction, you get a backdrop rather than a world.
The four elements
Place
Surface, lighting, weather, time of day. This becomes the anchor you refer back to in later steering instructions.
A rain-slick cobblestone alley in Kyoto, lantern light pooling in puddles, late evening, light drizzle.
Motion
Camera behaviour and pace. Film language works well here because it is well understood.
Camera drifting forward at walking pace, slight sway, passing closed shopfronts.
Sound
Name ambience and one specific detail. A short list reads better than an exhaustive one.
Audio: distant bicycle bell, water dripping from eaves, low city hum.
Interaction
What a viewer could do in this space. This is the element most people skip and the one that makes the world feel real.
A viewer could step under the awning on the left or continue toward the crossing.
A template you can copy
[Place and surface], [lighting and weather], camera [movement] at [pace].
Audio: [ambience] plus [one specific sound].
A viewer here could [interaction].Fill it in:
A highland ridge above the clouds at first light, wet grass and exposed rock,
camera panning right along the crest at a slow, steady pace.
Audio: steady wind and distant eagle calls, plus the sound of boots on loose stone.
A viewer here could walk along the ridge or step toward the edge.That is roughly sixty words. It is enough to define a world and short enough that you can still make targeted edits later.
Six mistakes that cost consistency
These map to documented characteristics of the model, not to prompting superstition.
1. Describing a still, not a world
With no motion, there is nothing to simulate and nothing to steer. Always add a camera move and one thing that changes.
2. Stacking too many edits
Google reports drift after many edits. Make fewer, bigger changes rather than a stream of small corrections. If you find yourself on your eighth adjustment, the first prompt was probably underspecified.
3. Editing early
An intervention near the start compounds through the whole session. If you already know a change is coming, put it in the initial prompt instead of steering to it.
4. Expecting text to render
Signage and labels are unreliable unless explicitly prompted. Do not design a scene whose meaning depends on readable text.
5. Assuming geographic accuracy
Real places are not reproduced precisely. If exact layout matters — a real building, a specific street — do not rely on this. Use a reference image and accept approximation.
6. Trusting complex physics
Ordinary interactions hold up better than unusual ones. Test the specific interaction your workflow depends on before you build on top of it.
Steering discipline
Once the world is running, how you intervene matters as much as how you prompted.
Change one thing at a time. Multiple simultaneous edits make it impossible to tell which one caused a problem.
Change late. Early edits have longer to compound.
Prefer additive changes. Adding weather or a character tends to hold up better than moving existing geometry.
Stop when it is good enough. Every intervention is a small consistency cost. If the scene works, save it rather than polishing it into drift.
What about image and video starts?
DeepMind describes turning 2D archival photographs into immersive simulations. Starting from a still is the strongest way to control composition, because the layout is already decided — you are asking the model to extend and animate something, not to invent it.
If you have a reference frame, use it. Prompt-based layout control is always approximate; image-based control is much tighter.
Practise with a composer
The prompt composer on this site appends your scene type, aspect ratio, and clip length automatically, so you can concentrate on the description. It also includes a library of twelve prompts covering natural, urban, interior, and speculative environments.
It does not generate anything itself — generation is handed to a partner tool. That separation is deliberate: the site's job is helping you write a better prompt, not pretending to run a model it does not have.
Related reading:
- How to use Genie 4 — the step-by-step workflow
- Genie 4 features — why audio and interaction cues matter
- Genie 4 limitations — the constraints behind these rules
Source for capability claims: Google DeepMind's Genie 4 announcement.
