'Claude did most of the modeling itself': Blender scene built with Opus 5.5 draws 1,900 likes, and the method is 'give it references and watch'

Japanese creator Yasuhiro Arafune (@arafune) posted a 'Blender × Claude Opus 5.5 test' video on Nov. 6 (about 1,900 likes, 80,000 views). He said he allowed the 3D generator Tripo but Claude used it for only about 5% and modeled most of the scene itself. ChatGPT made a few textures such as text, shaders were finished almost entirely procedurally, audio came from Tencent's HunyuanVideo-Foley, and Claude even cut the making-of.

He followed Kamome Ashizawa's Oct. 14 note post on modeling a photoreal café with Claude Opus 5.5 and Blender MCP. There the human supplies references (three ChatGPT images and a written description of the café) and reviews progress, explaining what to fix and why. Claude Code drives Blender through the official Blender MCP, writes four design documents, builds the scene in 13 stages and renders each from the reference angle to overlay and compare. Rules live in CLAUDE.md plus subagents for design, modeling and visual QA.

Ashizawa told it never to use outside 3D models and to write that into the rules, and when results weren't photoreal, to research Blender photoreal techniques, write them up, plan, get approval, then execute. The final scene has about 1.1 million faces and renders in 7-11 seconds per frame.

After a reply (26 likes) said the video proves nothing without the pipeline, prompts and time spent, Arafune linked the method and corrected himself: not every texture was procedural, some generic materials came from ChatGPT and only the foreground trees were assets.