viettranx/3dviz-pro-max

3Dviz Pro Max – AI‑driven 3D scene authoring

What it is – A stand‑alone skill for Claude Code, Codex and similar LLM agents that lets the model turn a natural‑language intent (e.g., “build a small fantasy village with moving creatures”) into a complete, renderable Three.js scene. The skill supplies a ten‑step workflow, a searchable catalog of 223 “recipes”, 440 knowledge records, 22 ready‑made kit blueprints, style profiles, and Python helpers that capture the rendered frames for the agent to inspect.

How it works

1. Intent → reasoning – The agent parses the user request and decides which objects, styles and behaviours are needed.
2. Reuse / adapt kits – It selects a kit blueprint (e.g., a terrain, a cottage, a gear set) from the catalog, or author‑writes a new one.
3. Build first view – A Vite + Three.js scaffold is populated with the chosen assets, lighting rigs, camera settings, etc.
4. Run & capture – scripts/capture.py launches a headless Chromium instance (via Playwright), renders the scene on a real GPU, and writes PNG frames and logs.
5. Refine – The agent compares the observed frames with the intended description, adjusts parameters, and repeats until the report matches the evidence.
6. Report – The final output is a concise description of what was built together with the captured images.

What ships

Item	Quantity	Note
Recipes	223	24 thematic directions (environments, anatomy, robotics, etc.)
Knowledge records	440	Defaults for styles, lighting, camera, materials
Kit blueprints	22	Proven Three.js modules with SHA‑256‑verified captures (tiers T0‑T3; T4 not shipped)
Style profiles	37	Named looks with numeric defaults
Example studies	37	Fully runnable demos (heart anatomy, gear mechanics, fantasy village, …)
Showcase frames	21	Real GPU captures (1920×1080) stored in docs/demos/

Installation – The whole skills/3dviz-pro-max/ folder is the plugin.

# Claude Code (plugin directory)
ln -s $(pwd)/skills/3dviz-pro-max ~/.claude/skills/3dviz-pro-max
# or load as a zip
claude --plugin-dir 3dviz-pro-max-0.1.0-plugin.zip

For Codex, copy the folder into the Codex skill directory. Playwright (for Chromium) and optionally Blender 4.2+ are required only for the capture step; the rest runs with plain Node.

Running the demos

cd examples
pnpm install          # install Vite + Three.js deps
pnpm dev               # serves http://localhost:4180/<demo-id>/

Each demo folder contains the prompt, the generated scene, source references and limits.

Typical use‑cases

• Rapid prototyping of game‑level assets or scientific visualisations from a single sentence.
• Teaching: students ask an LLM to illustrate a concept (e.g., “show a gear pair with 3:2 ratio”) and receive a live Three.js model they can explore.
• Evaluation of LLM grounding: the skill’s built‑in retrieval‑case tests let researchers measure how well an agent can turn intent into observable graphics.

Honest limits

• The skill only captures frames; it does not guarantee that every recipe has been rendered successfully.
• Knowledge records are editorially reviewed, not derived from runtime observations.
• Tier T4 blueprints are not provided; the highest‑fidelity assets are the two Blender‑baked GLBs (tier T3).
• Evaluation scores are based on a single‑run per condition, so they are indicative, not conclusive.
• Without Playwright/Chromium or Blender, the capture step will abort and the agent must report the missing frames.

License – MIT for the repository’s code, data and authored media. Third‑party assets retain their original licenses (e.g., BodyParts3D meshes under CC BY‑SA 2.1 JP).

---

All details are taken directly from the repository’s README; no external assumptions have been added.