# Rich environment evaluation prompt

Use this prompt with an explicitly supplied author model name. Replace the
bracketed fields before running. This is an authoring task, not a scored result.

> Read AGENTS.md and docs/artifacts.md. Create a new artifact by [EXACT MODEL NAME]
> under [MODEL ID]/[NEW ARTIFACT ID], using the scaffold command. Build an original,
> richly furnished garden pavilion that feels inhabited and rewards exploration.
>
> Preserve a clear centre at least 5 m across and a believable human scale. Put
> useful objects into coherent activity areas: a reading corner, a tea setting,
> planted seating, a water feature, and a small sculpture display. Build a layered
> foreground, middle distance, and surrounding landscape. Include at least 12
> distinct object families and 40 intentionally placed visible objects; repeated
> pebbles or leaves alone do not meet the richness requirement. Add construction
> details such as canopy slats, trim, cushions, book spines, vessel rims, stems,
> and contrasting material roughness. Prioritize composition over polygon count.
>
> Implement at least five discoverable, reversible object interactions. Each
> must visibly change the actual rendered scene, support pointer and keyboard
> activation, provide a concise status update, and reset to its initial state.
> Examples: switch lantern lights, open a hinged journal, fill cups, change basin
> water level, and turn a sculpture. A text-only response is not an interaction.
> Provide close-up and whole-scene views. Do not autoplay sound, animation, or
> camera motion. Avoid constant render loops when the scene is idle.
>
> Supply original procedural geometry, preview renders, the generator and its
> seed, a named object inventory, an asset-size/geometry report, dependency
> versions and licenses, and browser tests. Keep code, media and generation
> scripts in the artifact directory. Generate a GLB for browser interaction and
> a USDZ for the native environment when the available toolchain permits it;
> both should depict the same furnished layout. Do not replace shared site chrome.
>
> Use runtime capability detection and explicit user actions for immersive entry.
> A static USDZ does not implement object interactions. If using static native
> immersion, clearly scope the interactions to the browser scene. If implementing
> interactive immersion instead, supply a real WebXR render loop and spatial
> input, with session cleanup; do not substitute a capability check for rendering.
> Preserve a useful still-image fallback without WebGL, JavaScript, network access,
> or spatial APIs. Make the sandboxed gallery preview usable with local assets.
>
> Run npm run check and validate the USDZ. Visually inspect the full scene and
> interaction states. Check mobile, keyboard access, reduced motion, load failure,
> repeat/reset behavior, and gallery sandboxing. Record measured results and
> limitations in README.md. Separately record actual Safari/visionOS versions and
> device tests; mark spatial rendering unverified when no headset was tested.

## Review rubric (0–4 each; leave untested items unscored)

| Dimension | Evidence for a strong result |
| --- | --- |
| Spatial richness | Coherent activity areas, useful objects, depth, detail visible in close-up |
| Visual craft | Readable composition, consistent scale, distinct materials, plausible construction |
| Interaction | Five visible state changes, direct picking, keyboard equivalents, repeatability, reset |
| Compatibility | Flat fallback, sandbox preview, runtime probes, honest native vs interactive scope |
| Reproducibility | Original assets, generator, seed, versions, licenses, sizes and geometry inventory |
| Validation | Passing checks, inspected screenshots, failure handling, separately recorded device evidence |

The Living Pavilion artifact is one worked example, not an asset set every model
must copy. When comparing models, give each the same prompt, tool access, runtime
budget, and device-testing opportunity; retain independent artifact directories.