Teaching your AI agent to actually finish a 3D model

What it does This repo packages professional Blender workflows — modeling, UVs, PBR materials, baking, rigging, animation, export, even weather effects — as Agent Skills for Claude Code, Codex, Gemini CLI, Copilot and Cursor. A router skill (blender-core) handles Blender 5.x API rules and dispatches to the right specialist. Every script is claimed to be tested in a real Blender, backed by 50 headless Blender tests in the repo’s tooling.

The interesting bit It’s essentially curated prompt engineering with receipts: instead of hoping your agent remembers the difference between ORM packed textures, it gets a skill document per discipline. The repo includes a before/after comparison page showing the same model and prompts with and without the skills — a rare willingness to show evidence rather than just assert quality.

Key highlights

• 22 skills spanning the full pipeline, from hard-surface bevels to ARKit 52 blendshapes
• Export targets include Roblox, Unity, Unreal, Godot, web and USD
• A QA skill that catches z-fighting, floating parts, bad normals and weights
• Works with an optional Blender MCP server for live sessions, or runs Blender headless in the background
• Geometry Nodes driven from simple JSON — procedural modeling without node spaghetti

Caveats

• Requires Blender 5.2 LTS+, so older pipelines are out of luck
• The comparison page is the only quality evidence offered; no independent benchmarks

Verdict If you use an AI agent for game-asset or 3D work, this is a genuinely useful shortcut — the boring parts (naming conventions, texel density, export settings) are exactly where agents flail. If you don’t touch Blender, move along.