The 3dviz-pro-max project is an open-source agent skill designed to turn text ideas into explorable 3D scenes using Three.js and Blender. As of September 10, 2026, it provides AI agents with 223 recipes, 440 knowledge records, and 22 reusable kit blueprints to accelerate creative 3D visualization.
This repository provides a structured workflow for AI agents like Claude Code and Codex. It enables them to not only generate 3D assets but also to reason about objects, choose representations, and refine scenes based on captured frames. The project aims to bridge the gap between a high-level creative concept and a functional, interactive digital artifact.
What Assets Does the Project Provide?
The project ships with a large catalog of pre-authored assets to guide AI agents in their creation process. This includes hundreds of recipes and knowledge records across 24 distinct subject areas, from fantasy environments to molecular geometry. It also contains runnable templates and kit blueprints to jumpstart development.| Asset Type | Count (as of Sept 2026) |
|---|---|
| Recipes | 223 |
| Knowledge Records | 440 |
| Kit Blueprints | 22 |
| Runnable Example Studies | 37 |
The 24 "directions" cover a wide range of subjects, demonstrating the skill's flexibility. These categories include artistic domains like "Fantasy village" and "Puppet performance." They also cover technical fields such as "Molecular geometry," "Calculus and fields," and "Rigid mechanics." This structure provides agents with domain-specific guidance for modeling complex topics.
How Does the AI-Driven Workflow Function?
The skill utilizes a ten-step workflow that prioritizes creative direction before technical implementation. An agent begins with a user's intent, reasons about the necessary objects, and builds an initial version. The core of the process is a "run-and-inspect" loop where the agent captures and analyzes rendered frames to iteratively refine the scene.
This process is more than just code generation. It starts with defining the scene's key elements like silhouette and behavior. A Python script, `capture.py`, then drives the scene in a headless browser to produce actual images. The agent compares these observed frames to the original intent, enabling a refinement process that ensures the final output matches the user's vision.
What Are the Project's Current Limitations?
The repository maintains a policy of "honest reporting" regarding its limitations. While it offers many recipes, not all have been rendered into final scenes. The project also clarifies that its knowledge records are authoring guides, not guarantees of runtime performance, and some installation paths remained untested as of September 2026.The project's full potential relies on external tools. While Blender is not required, its absence limits asset creation to lower-fidelity tiers. Furthermore, controlled tests showed the skill improved a baseline score from 21 to 25 out of 30, a positive but thin margin that is not yet a definitive proof of superiority.
The documentation also notes that certain installation methods, particularly via the public marketplace, were untested at the time of the latest update. This transparency helps developers understand the current status of the project. Using a tool like the CLI for Claude Code could help developers manage such local skill configurations.
