メニュー
Read one reference image and produce a structured scene plan (terrain, objects, palette, lighting, camera). Use when build-world calls into Phase 0.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Top-level orchestrator. Take a reference image from input/ and assemble a full stylised 3D scene in Blender via the autonomous vision loop. Use when the user says "build me a world" / "make a scene" / "blast this into Blender".
Generate one atomic 3D asset via fal Tripo P1 (textured, PBR auto-included, $0.50). Polycount tuned per object class — sweet spot 5,000-15,000 faces. No fallback (Trellis 2 quality is worse for stylised cases; retry P1 with a better reference plate instead).
Generate one tiled PBR material via fal PATINA. BaseColor + Normal + Roughness + Metalness + Height maps from a text prompt. ~$0.06-0.10 per material at 1024px. Use for procedural surfaces (sand, water, grass, dirt, rock) — NOT for objects (those come textured from Tripo P1).
The autonomous vision loop. Sort objects by size descending, place in batches of 5, render the same 3 angles as the nano-banana control views, compare six-up, micro-adjust, then continue. Eevee renders — fast.
Build the procedural ground in Blender — primitive surfaces (sand, water, grass), GN displacement for terrain shape, then tiled PBR materials via PATINA. Pastes a known-good template rather than composing GN nodes from scratch (sidesteps GN-API-drift).
Read one reference image and produce a structured scene plan (terrain, objects, palette, lighting, camera). Use when build-world calls into Phase 0.
| name | analyze-reference |
| description | Read one reference image and produce a structured scene plan (terrain, objects, palette, lighting, camera). Use when build-world calls into Phase 0. |
Use the available Codex tools for image viewing, file inspection, and file edits. In this migrated skill, $1 means the world slug chosen by the user or derived by build-world.
You are analysing one reference image and writing a scene plan to worlds/$1/plan.json.
Read the reference image. Use the Read tool on worlds/$1/source/reference.png — Use Codex's available image input or view_image on the local PNG. No external vision API.
Identify the style anchor. Match against docs/style-references.md anchors:
"custom" if no clear match.Extract style metadata:
palette: dominant 4-5 hex colours (sand, water, foliage, wood, sky)level_of_stylization: "high" (Banjo) / "mid" (Pokémon SV) / "low" (Horizon FW)shading: "toon" / "flat" / "pbr-light"lighting: "golden-hour" / "noon" / "dusk" / "overcast"Identify the terrain shape. Read the image and decide:
"island" (round, surrounded by water)"shoreline" (water on one side)"hills" (no water, vertical relief)"flat" (no relief)"crater" (recessed centre)
Plus: ground type (sand/grass/dirt/rock/snow), water level if applicable, approx size in metres (default 12×12 for an island).Enumerate the objects. Look at the image and list every distinct object class you can identify. For each:
id: lowercase-slug (e.g. palm-tall, beach-hut, rock-mossy)name: human-readable description suitable for a Tripo gen prompt (e.g. "Tall stylised palm tree with curved trunk")count: how many instances are in the reference (or how many you'd add to look right)approx_positions: list of [x, y] in metres, anchored at scene originscale_hint: "large" / "medium" / "small" / "tiny"priority: "hero" (visually critical, use Tripo P1 textured) / "filler" (generic, use Trellis 2 or re-instance)face_limit: integer in [5000, 15000]. Pick from the table in generate-3d/SKILL.md (hero structural 12-15k, hero organic 10-13k, mid-detail 7-9k, filler 5-6k).Total instance count across all objects MUST be ≤ 25 by default. Each instance = $0.30-0.50 in fal spend. The cap keeps a single scene under $15 of asset generation.
How to fit:
ids (object classes).count for repeating items (palms, rocks, shells). Same .glb, different positions — costs $0.50 once.count across all objects = total instances. Keep this ≤ 25.If the reference clearly NEEDS more than 25 instances to read correctly:
Reference suggests ~N instances. Default cap is 25 (≈$<estimated cost> in fal spend). Raise the cap?Camera framing. Estimate the reference's camera:
framing: "three-quarter aerial" / "side" / "top-down" / "hero-low-angle"focal_length_mm: usually 35 for the Nintendo aerial-three-quarter look, 50 for closer, 24 for widerWrite worlds/$1/plan.json with the structure shown in docs/PIPELINE.md:
{
"world_slug": "$1",
"style": { "anchor": "...", "palette": ["#..."], "level_of_stylization": "...", "shading": "...", "lighting": "..." },
"terrain": { "shape": "...", "ground": "...", "water": { "present": true, "level": -0.05 }, "size_meters": [12, 12], "feature_notes": "..." },
"objects": [
{ "id": "palm-tall", "name": "...", "count": 3, "approx_positions": [[2,1], [3,0.5], [1.5,2]], "scale_hint": "tall", "priority": "hero", "face_limit": 12000 },
...
],
"camera": { "framing": "three-quarter aerial", "focal_length_mm": 35 }
}
count) ≤ 25, confirm all priority: hero objects have a clear name suitable for asset generation, confirm every object has a face_limit in [5000, 15000]. Fix in-place if needed.Returns the plan path: worlds/$1/plan.json. Logs a one-line summary of the analysis (anchor, terrain, object count) for the orchestrator.