Route to the specialized skill

Task	Skill
Units, collections, naming, linking, asset libraries	`blender-scene-setup`
Hard-surface modeling, bevels, booleans, normals	`blender-hardsurface-modeling`
Sculpting, multires, remesh	`blender-sculpting`
Retopology, UV unwrap, texel density	`blender-retopology-uv`
PBR materials, color spaces, texture sets	`blender-pbr-materials`
Baking high→low maps (normal, AO, curvature, ...)	`blender-texture-baking`
Geometry Nodes, procedural modeling	`blender-geometry-nodes`
Body rigging, Rigify, weight painting	`blender-character-rigging`
Face rig, shape keys, ARKit, visemes	`blender-facial-shapekeys`
Keyframes, actions, NLA, baking, root motion	`blender-animation`
Poly budgets, LODs, draw calls, cleanup	`blender-game-optimization`
Export: glTF/FBX/USD to Roblox, Unity, Unreal, Godot, web	`blender-export-pipelines`
Cycles/EEVEE settings, compositor, batch render	`blender-rendering`
Lighting, HDRI, camera, turntable	`blender-lighting-lookdev`
Time of day, night, sunset, fog, mood (cyberpunk, horror...)	`blender-time-of-day`
Rain, snow, wet surfaces, puddles, lightning, dust/embers	`blender-weather-fx`
Cloth, rigid body, fluids, hair, GN physics	`blender-simulation`
Writing add-ons / extensions	`blender-addon-development`
Batch/headless pipelines, file conversion	`blender-pipeline-automation`
Checking a mesh/asset before delivery (always)	`blender-asset-qa`
End-to-end playbooks (prop, character, Roblox, product shot)	`blender-workflow-recipes`

Execute

Live (blender-mcp execute_blender_code): work in small steps and print only what you need. To call a skill script, import it:

import sys, importlib; sys.path.insert(0, r"<skills_dir>/blender-asset-qa/scripts")
import qa_check; importlib.reload(qa_check); print(qa_check.run(objects=["Chair"]))

Each MCP call runs in a fresh namespace: re-import in every call. That's cheap because modules stay cached in sys.modules. reload only after editing a script. Headless (any agent with a shell): this prints only the script's JSON (@@JSON@@ lines):

python <skills_dir>/blender-core/scripts/run_in_blender.py --blend in.blend <script.py> -- --objects Chair --save out.blend
python <skills_dir>/blender-core/scripts/run_in_blender.py --expr "import bpy; print(len(bpy.data.objects))"

Raw form: blender -b in.blend --factory-startup --python-exit-code 1 -P x.py -- args. Every script here exposes run(**kw) -> dict and the same flags after --. First call in a session: scripts/env_report.py, which returns the version, engines, GPU, add-ons, units, color management and scene counts.

Rules

Data API over operators: use bpy.data, bmesh, mesh.attributes, and foreach_get/set with numpy. Operators depend on context and selection, and many need a 3D View, which background mode doesn't have.
If an operator is unavoidable: with bpy.context.temp_override(object=ob, active_object=ob, selected_objects=[ob], selected_editable_objects=[ob]): bpy.ops.... Check that the result is {'FINISHED'}.
Edit mode isolation: ob.data isn't updated while in Edit Mode. Use bmesh.from_edit_mesh + bmesh.update_edit_mesh, or switch to OBJECT mode before reading.
Nodes: find them by type (n.type == "BSDF_PRINCIPLED"), never by name, because names are localized. Sockets by English name or identifier are fine.
Enums: read the valid values with [i.identifier for i in T.bl_rna.properties[p].enum_items]. Dynamic enums (render.engine, view_transform, look) under-report, so assign them inside try/except TypeError (the message lists the valid values).
Engines: BLENDER_EEVEE, CYCLES, BLENDER_WORKBENCH. Cycles device: scene.cycles.device = "GPU" + preferences compute_device_type.
Evaluated data: use ob.evaluated_get(depsgraph), then .to_mesh(), then always .to_mesh_clear(). Call view_layer.update() before reading matrix_world after a change.
Freed data: applying or removing a modifier, or removing an object or ID, frees it. Collect names first ([m.name for m in ob.modifiers]) and never touch the old Python reference afterward, because it returns garbage or crashes.
Never trust default names ("Cube.001"). Keep object references, rename explicitly, and look objects up via bpy.data.objects[name].
Scale: 1 unit = 1 m. Apply rotation/scale (transform_apply) before bevels, modifiers with distances, rigging, physics, baking and export.
Paths: bpy.path.abspath("//textures/x.png") resolves relative to the .blend. Save explicitly; use save_as_mainfile(filepath=..., copy=True) for backups.
Live sessions are the user's work: never read_factory_settings, never delete data you didn't create, and save a backup copy before destructive batch edits.
GPU module in background: call gpu.init() first (5.2+). bgl no longer exists.
Big meshes: no per-vertex bpy.ops. Use numpy foreach_get (see references/bpy-patterns.md).

5.x API changes that break old tutorials/LLM memory

action.fcurves → channelbags · scene.node_tree → scene.compositing_node_group · material.use_nodes is deprecated (new materials already have nodes) · mod["Socket_2"] → mod.properties.inputs.Socket_2.value (5.2) · BLENDER_EEVEE_NEXT → BLENDER_EEVEE · Boolean FAST → FLOAT · bone layers → bone collections · bl_info add-ons → extensions (blender_manifest.toml) · bpy.props values are no longer dict-accessible (scene["cycles"]). Full table: references/api-changes-5x.md.

Verify (required before saying "done")

Numbers, not adjectives: report tris, objects, materials, dimensions (ob.dimensions) and the output files written.
Geometry changed → run blender-asset-qa qa_check.py (z-fighting, detached parts, non-manifold, flipped normals, scale).
Look at it: MCP look (mode="angles", shading material/wireframe). Headless: blender-asset-qa/scripts/review_shots.py renders contact sheets.
Re-read your own output: confirm that exported files exist and that their size is non-zero.

References

references/api-changes-5x.md: 4.x→5.2 migration table. Load it when code fails with AttributeError/KeyError.
references/bpy-patterns.md: verified snippets for bmesh, attributes, numpy, temp_override, depsgraph, drivers, append/link, handlers and timers.

blender-core

Blender Core (5.2 LTS+)