renderoni: Skill for Claude Code

Prompt → image → img2threejs → Renderoni

Do not one-shot a whole game as a giant TypeScript file. Keep the agent context small: one prompt, one scene image, one inventory JSON, then reconstruct unique objects independently.

Upstream reconstruction skill: img2threejs. Renderoni only mounts the resulting () => THREE.Group factories.

Token budget (hard rules)

1. Keep the inventory JSON in context. Target < 800 tokens.
2. Never paste img2threejs SKILL.md, forge scripts, or generated factory source into the main game conversation after the factory is written to disk.
3. Reconstruct one unique factory key per isolated pass. Reuse the same factory for repeated props (10 trees = 1 factory).
4. Gameplay stays in Renderoni presets (kccPlayer, sensor, body, light). Factories are visuals + collider hints only.

Pipeline

prompt
  → one scene image (optional but preferred)
  → compact SceneInventory JSON
  → unique factory list
  → img2threejs per factory (or fallback primitive)
  → mountSceneInventory(engine, inventory, factories)
  → add player / win-lose with presets

1. Inventory first

Write scene-inventory.json using SceneInventory from renderoni/scene:

{
  "version": 1,
  "prompt": "stone courtyard with crate, lantern, tree, well, coin",
  "image": "refs/courtyard.png",
  "seed": 42,
  "elements": [
    {
      "id": "crate",
      "factory": "woodCrate",
      "kind": "prop",
      "position": [-3, 0.55, -2],
      "collider": { "shape": "box", "size": [1.1, 1.1, 1.1] }
    }
  ]
}

kind: terrain | prop | actor | pickup | decor. factory is a registry key, not source code.

2. Scene image (optional)

If the user has no image, generate one wide establishing shot of the whole scene. Then list visible objects. Use imageRegion (normalized 0–1) only when cropping helps reconstruction. Do not generate a separate image per blade of grass.