Codex builds and refines a pelican-on-a-bicycle Blender scene

Simon Willison used Codex mode in the ChatGPT macOS app to ask Blender to render a pelican riding a bicycle. He gave the agent a short instruction to use the installed Blender application, then later made the setup easier to repeat by specifying Blender’s background command and a Python scene script. The first pass produced an editable .blend file and a Python script after 2 minutes 39 seconds.

Willison then asked for a background and more visual detail. That pass produced a festival version after 3 minutes 51 seconds. A further request to improve the scene took 5 minutes 59 seconds and generated a final project file and script. The article links the generated artifacts in a public repository, so this is a documented creative build rather than a hypothetical capability.

The useful workflow is to let the coding agent operate a local creative tool, preserve the generated source and project file, and refine the scene through specific visual follow-ups. The source does not report a professional review, a print, or a game deployment; the result is an editable rendered scene. Willison later created a local Blender skill from what Codex learned and reused it for image-based experiments, but the article does not document the output of the separate faverge-egg prompt mentioned in the related Hacker News comment.

Prompt
Original prompt

Source: Blog · Simon Willison