A SKILL.md for driving AI agents through headless Blender (works alongside blender-mcp)

[capybala18]

Sharing something that's directly complementary to blender-mcp: a big
SKILL.md (~1300 lines) plus a supporting bpy Python library that gives
an AI coding agent a real workflow for building 3D scenes in Blender —
plan the components first, build against calibrated camera/lighting
presets, then self-review against measured render stats before calling
it done.

It covers the stuff that usually goes wrong when you just tell an LLM
"model this in Blender": automated geometry QA (interpenetrating parts,
floating objects, spacing outliers), exposure calibrated against real
render output, component checklists so nothing obvious gets skipped,
and a reference-photo reconstruction pipeline for matching real objects.

blender-mcp is what I point at when I want the interactive/debug side
of a Blender workflow that pure headless scripting can't give you —
live scene inspection, and especially its asset-library integration
(Poly Haven, Sketchfab, Hyper3D Rodin) for pulling in ready-made
environment pieces instead of hand-modeling everything from scratch.
My skill's default path is still headless bpy for the actual building,
but blender-mcp is the natural complement for that layer, not a
competitor to it — this project genuinely earned its 27k stars.

Everything in the gallery below was built with DeepSeek V4.1 Flash — a
small, cheap model with no 3D specialization — through my own agent's
headless Blender scripting path. Full .blend files for every scene are
in the repo's Releases if you want to open them and see how they're
actually built.

Repo: https://github.com/capybala18/capybala-bpy-skill