Blender Mentor

Turn Claude Code / DSH / Cursor into a senior Blender mentor + technical director.

It diagnoses your level, teaches a structured lesson, grades your work with a 0–3 rubric, and runs machine acceptance on your .blend — assets, curves, renders, ESD secondary animation, and biomechanics.

English | 简体中文

What it looks like

you    › Review this walk cycle. Grade it with the rubric and give me a P0/P1/P2 fix list.

mentor › L2 · Review mode · score 2.1 / 3 — critical: foot sliding
         P0  Lock contact frames 1 / 13 / 25; add heel roll.
         P1  Reduce hip Y translation 18%; clean shoulder counter-rotation.
         P2  Add toe flex on push-off; re-check the 12-frame cadence.
         acceptance  asset_gate PASS · curve_audit 3 warnings · render_probe PASS

Not a Blender add-on. The skill gives your AI agent the mentor's playbook. It can generate commands and, when the host has shell/Blender access (or an MCP bridge), run blender -b ... --python ... automatically. Without that access it cannot operate Blender's UI for you. You still need Blender installed to open .blend files and run acceptance checks (see Requirements).

Rubric at a glance

Score	Meaning	Walk-cycle example
0	A critical item vetoes the delivery	foot sliding → 0, not deliverable regardless of the other items
2	Shippable	contacts locked, no interpenetration
3	Production-ready	contacts locked + heel roll + reusable cadence

Full scoring anchors and self-review template: references/rubric.md.

Quick start

DSH — clone and install (Windows / Linux / macOS / WSL):

git clone --depth 1 https://github.com/Memory175/blender-mentor.git
cd blender-mentor

# Windows
powershell -NoProfile -ExecutionPolicy Bypass -File dsh/install.ps1

# Linux / macOS / WSL
bash dsh/install.sh

Claude Code — install as a plugin:

/plugin marketplace add Memory175/blender-mentor
/plugin install blender-mentor@blender-mentor-marketplace

Cursor / other Agent Skills clients — clone the repo and point your client at SKILL.md (in Cursor, mention @SKILL.md; for an always-on rule, add a project rule that references it).

Then open a new agent session and try:

Use blender-mentor to walk me through M1 modeling and topology.

What you get

Component	What it does
Mentor behavior	L0–L3 level diagnosis; explain / coach / accept / firefight / review modes
D1–D4 delivery loop	deliver → self-review → P0/P1/P2 fix list → explicit next-step question
0–3 rubric	critical items, scoring anchors, and a self-review template
Engineering precheck (ENG-001)	stability, wall thickness, draft, overhang/support, interference, clearance
M0–M6 curriculum	modeling & topology → UV/materials/lighting/render → animation/rigging → physics/geometry nodes/pipeline
24 executable tools	6 .blend acceptance scripts + engineering precheck + module governance (preflight/lint/conflict/claim audit/self-audit index) + source governance (audit/refresh) + regression/routing tools
Delivery QA gate	floors/coordinates, mesh/UV, camera/export, professional-boundary checks
INT-001 (experimental)	Chinese-style two-bedroom apartment interior; declares styles/layouts; sources pending
INT-002 (experimental)	modern luxury large flat; program from user plan; dimensions pending
Interior rule	interior modeling requires a dedicated INT-* module with styles/layouts; style/layout mismatches are surfaced
Knowledge base	130 common misconceptions + 40 red lines

Requirements

Requirement	Needed for
Blender 4.2 LTS / 5.x	.blend acceptance scripts (blender -b ... --python scripts/...)
Python 3	selftest.py, persona_lint.py, load_router.py, preflight.py
Optional: shell or Blender MCP/RT bridge	Automatic script execution; without it, copy commands into Blender Scripting or a terminal manually

How it works

SKILL.md       routing + D1–D4 delivery loop + machine-acceptance commands
references/    14 reference files (9 persona docs + module format + authoring guide + model QA + INT modules), loaded on demand
scripts/       24 executable tools
assets/        ESD templates, load plans, deny-word example, source registry template

1. SKILL.md routes the request and sets the delivery loop.
2. preflight.py classifies it: proceed (module-backed), proceed_core (teaching/operations, no module needed), ask, or block (reference data required but not covered).
3. load_router.py builds a per-task load plan and only loads modules selected by preflight; proceed_core tasks get core docs, never modules.
4. source_audit.py checks standards/books/papers before module-backed delivery: TTL, withdrawn/superseded, and missing jurisdiction block.
5. Acceptance scripts run against your local .blend through headless Blender.
6. Your files stay local; the acceptance tools do not upload anything.

Advanced installation

Claude.ai

Zip the skill folder (SKILL.md + references/, scripts/, assets/) and add it to your Skills library.

DSH plugin market

This repository is also a DSH skill plugin: package.json declares dsh.bundle.patch → cordis.patch.yml, which mounts the bundled skills/blender-mentor/. After the dsh.so listing is approved:

dsh plugin --profile web add github:Memory175/blender-mentor

Downloadable installers

dsh-install.ps1 and dsh-install.sh are attached to the latest release; download and run them if you prefer not to clone.

Repository layout

.
├── SKILL.md                     # Agent Skill entry point (required)
├── README.md / README.zh-CN.md  # English / 简体中文
├── LICENSE                      # MIT
├── CHANGELOG.md
├── package.json                 # DSH plugin manifest (dsh.bundle.patch)
├── cordis.patch.yml             # DSH Cordis patch (skill-filesystem provider)
├── lib/index.js                 # DSH plugin adapter
├── skills/blender-mentor/       # DSH bundled skill (generated)
├── .claude-plugin/marketplace.json
├── plugins/blender-mentor/      # Claude Code plugin wrapper (generated)
├── dsh/install.ps1 / install.sh # DSH one-line installers
├── references/                  # 14 reference files (9 persona docs + module format + authoring guide + model QA + INT modules)
├── scripts/                     # 24 executable tools
├── assets/                      # ESD templates, load plans, deny example
├── docs/                        # banner + example session image
└── tools/build_skill.py         # maintainer build tool

FAQ

Is this a Blender add-on? No. It is an Agent Skill: instructions + scripts that an AI client loads. Blender itself is a separate install.

Do I need Blender installed? For teaching, critique, and planning: no. For opening .blend files, rendering probes, and machine acceptance: yes, Blender 4.2 LTS / 5.x.

Does it control the Blender GUI? No. It writes and runs bpy/headless scripts and reviews the artifacts you provide; it does not click through the interface.

Is my .blend uploaded? No. Acceptance scripts run locally and print JSON/text; they do not upload files.

How do you keep standards/books current? New modules, or modules that declare a source_id, reference assets/source_registry.json. source_audit.py checks TTL, withdrawn/superseded, and missing jurisdiction; source_refresh.py queries official catalogs/Crossref but only proposes changes. All shipped module data now use source_id; the built-in registry maps them to module reference sections as internal / metadata-only, so external standards/books still need page-checking. “Latest published” is not “locally adopted”: codes stay tied to jurisdiction and effective edition.

Which clients are supported? DSH, Claude Code, Claude.ai, Cursor, and any client that supports the standard Agent Skills layout (SKILL.md + resources).

What do the two version numbers mean? The persona suite is V5.16.1; the packaged skill release is v5.16.1.

Development

python -m compileall -q scripts tools
python scripts/persona_lint.py . --allow-no-deny   # structure gate
python scripts/selftest.py .                       # 28 pass + 1 skipped

Blender-side checks:

blender -b file.blend --python scripts/asset_gate.py -- --json gate.json
blender -b scene.blend --python scripts/biomech_check.py -- --eccentric 1,4 --support 1-4

Rebuild the skill directory:

python tools/build_skill.py --source <V5.16.1 suite> --out <repo parent> --force

License

MIT — see LICENSE. Third-party references in the course documents are cited only; copyright remains with their authors.

---

If this saves you time, a ⭐ helps others find it · Report an issue · Releases