World Model Prompt Architect
Write prompts that generate walk-in, explorable simulated worlds.
This skill teaches you to design prompts for real-time world models like Genie 4: scene anchors, interaction verbs, and consistency constraints that keep a generated world coherent as you explore it.
Install it in 3 steps
No sign-up. No CLI tools. Pick your AI, copy one line, paste it over.
Prefer doing it yourself? Buy the skill ($4.99), download the ZIP from your checkout confirmation, unzip it into the directory shown above, restart your AI session — same result. Already bought it? Your AI assistant can download it for you.
What's inside
One SKILL.md (the playbook) + one starter template. A preview:
World Model Prompt Architect
The idea in one paragraph
Real-time world models turn a text prompt into a world you can walk into — you type as you explore, and the world changes as you go. That only works when the prompt gives the model something solid to hold onto: where you are, what you can touch, and what should stay the same while you move. This skill shows you how to build those prompts layer by layer, using Genie 4's announced capabilities — real-time simulation, keyboard exploration, and typing-to-change — as the design target.
What it is built on
- Google DeepMind announced Genie 4 on October 27, 2026 ("A new generation of world simulation"), per the DeepMind blog post by Raia Hadsell, Diego Rivas, and Hang Qi: a prompt produces a real-time simulated world you explore with a keyboard and reshape by typing while walking.
- DeepMind describes four advances over the prior generation: stronger photorealism, smoother motion, native generated audio, and longer coherence — about 2 minutes of continuous consistency per TechCrunch, roughly twice Genie 3.
- A Gemini 3.8 Flash guiding agent works in the loop: it looks at frames, breaks a goal into steps, and verifies each step.
- Genie 4 is a research preview for trusted testers only, with a waitlist; DeepMind publicly acknowledges limits including drift, object permanence, physics, and legible text.
Workflow
- Anchor the scene (input: a one-line idea -> output: a scene brief). Name the location, era, time of day, weather, and camera feel. A world without anchors drifts within seconds, so write these down before anything else.
- Pick the interaction verbs (input: scene brief -> output: 5-8 allowed verbs). List what the explorer can do — open, climb, push, ask, light — and nothing else. Verbs are the contract between you and the world.
- Set the consistency constraints (input: scene brief -> output: constraint list). Fix the facts that must never change: gravity behaves normally, a red door stays on the north wall, written signs stay legible. Reference the known limits (drift, object permanence) and constrain around them.
- Draft the start state (input: brief + verbs + constraints -> output: start-state paragraph). Describe exactly what the first frame looks like and where the explorer stands. The guiding-agent idea from Genie 4 — a target broken into steps — starts here: define where the explorer begins and what "interesting" means.
- Script two change triggers (input: start state -> output: typed commands + expected changes). Write the exact phrases the explorer will type mid-walk ("make it rain", "open the red door") and what the world should do in response. Keep each trigger to one change at a time.
- Run a drift check (input: full prompt -> output: revised prompt). Re-read the prompt after a simulated walk-through and ask: did anything contradict an anchor, did an object move on its own, did text become unreadable? Fix what drifted.
Template preview
World Model Prompt Worksheet
Fill this in, then assemble the pieces into your final prompt.
Assemble: paste the filled rows top to bottom into a single prompt paragraph and run the drift check.
