worldbench

A single prompt for generating a self-contained Three.js world.html of a floating biome island, plus a small site for browsing and interacting with what different models produce from it.

There is no scoring, no schema, no CLI, just the prompt and your own eyes.

The model only ever sees the natural-language prompt. It is never shown outputs/create.html. That file is a hand-tuned ideal world for humans: a qualitative target when judging outputs and when refining the prompt. Models invent their own scale and layout; a result can differ or be better if the ecological relationships and climate logic hold.

Running the site

npm install
npm run dev

Three pages:

• Home: a minimal floating-island animation, then "See Recent Tests" (up to 4 pinned tests).
• About: what worldbench is, the reasoning it tests, the biome legend image, and a placement graph of how the biomes relate to each other.
• View Tests: a grid of every test, searchable by name, model, or provider and filterable by provider badge, each card opening into its own page with an intro clip, the test date, and the live, interactive world.html embedded below it. Under that world, Compare models opens a grid that puts it side by side with the worlds from up to five other tests, added one panel at a time.

Adding a test

Drop a new folder in public/tests/<slug>/ with:

• meta.mdx: frontmatter (title, model, provider, date, pinned, plus optional summary, xPost) followed by optional notes. provider keys into the company name and logo; leave it out for non-model entries like the hand-tuned reference.
• world.html: the model's raw output
• world-preview.html: the same world with its legend / HUD overlays commented out of the render, which is what the comparison grid loads. Comparing islands side by side means comparing islands, not four different HUDs at once.
• optionally intro.mp4 / intro.webm / intro.gif

No code changes needed, the site reads public/tests/ at build time.

Using the prompt yourself

1. Open prompts/prompt.md (see also prompts/README.md).
2. Copy the entire file and paste it into a model.
3. Save the raw output (no surrounding prose) as world.html.
4. Open it in a browser and look around.
5. Optionally open outputs/create.html yourself as a qualitative reference.
6. Repeat with other models and compare the results side by side.