Blender Agent Skills for Claude Code: What EL4CTEO's 22-Skill Pack Covers and How to Check It

EL4CTEO's blender agent skills for Claude Code pack, blender-skills on GitHub, gives a coding agent 22 written Blender workflows to follow. You don't have to describe UV unwrapping or normal-map baking from scratch in every prompt. The README says the pack works with Claude Code, Codex, OpenCode, Gemini CLI, GitHub Copilot and Cursor, and that it targets Blender 5.2 LTS+.

The short version: the pack is a set of instructions plus bundled Python. It is not a model that makes art for you. It looks useful for solo developers who already run an agent. Its testing claim can't be checked from the README alone, so treat each skill as untrusted until it has run on a throwaway file.

What is EL4CTEO/blender-skills, and what does it claim?

The README says there are 22 skills covering "the full 3D pipeline": modeling, UVs, PBR, baking, rigging, animation, optimization, export, rendering, time of day, weather, simulation and QA. It also states: "Every script is tested in a real Blender."

The repo page showed this when it was read:

• An MIT license, about 430 stars, 39 forks and six commits on main.
• Folders for skills, shared, tools, tests, install, docs, assets and .claude-plugin, plus a GitHub Actions workflows folder.
• CLAUDE.md, AGENTS.md and CHANGELOG.md at the root.
• Topics including game-assets, bpy, mcp, godot, unity, unreal-engine and roblox.

A Nix overlay pull request, agents.nix PR #58037, packages the pack at version 1.1.0. Third parties are already shipping it through skills directories.

The Development section of the README lists two commands: tools/validate_skills.py for "spec + token budget checks" and tools/run_tests.py, which it describes as 50 headless Blender tests. That gives the testing claim a concrete basis. But no test logs or CI results were part of the material read for this post, so the claim is unverified here. With only six commits, the project is also very young.

Which 22 skills are in the pack?

The README's skills table lists every skill. Grouped by asset-pipeline stage, they fall out like this:

The README also links a with-and-without comparison page made by the maker. It shows outputs from the same model and prompts. It's a maker-produced demo, not an independent benchmark.

How do you install blender-skills in Claude Code and other agents?

For Claude Code, the README gives a plugin route:

/plugin marketplace add EL4CTEO/blender-skills
/plugin install blender-agent-skills@blender-agent-skills

For other agents, it says to clone the repo and run the installer:

git clone https://github.com/EL4CTEO/blender-skills
./blender-skills/install/install.sh          # macOS / Linux
.\blender-skills\install\install.ps1         # Windows

According to the README, the installer writes to ~/.claude/skills and ~/.agents/skills by default. It says Codex, OpenCode, Gemini CLI, Copilot and Cursor read .agents/skills. Flags include --agent to pick one agent, --project DIR to install into a single project, and --uninstall.

The README also says a Blender MCP server is optional. With one, the agent can work inside your open Blender session. Without one, the agent runs Blender in the background. Those two modes carry different risks, as covered below.

What a skill actually is

The Claude Code skills documentation describes a skill as a SKILL.md file of instructions in its own directory, such as .claude/skills/<name>/SKILL.md, with optional supporting files. The skill's description field decides when Claude invokes it. Lee Han Chung's first-principles deep dive calls skills prompt templates rather than executable code. It notes that bundled files can sit in folders like /scripts and can include Python or shell scripts.

This matters for risk. The SKILL.md text only steers the agent. Changes to your files come from the bundled scripts and from whatever Blender Python the agent writes and runs afterwards. For more on preparing a project so an agent can read it, see how to set up Claude Code for game development.

Does it work on older Blender versions?

The README requires Blender 5.2+, and blender-core mentions "5.x API rules". Reports on the 5.2 release explain why that matters.

• GameFromScratch describes 5.2 LTS as the release recommended for new projects, with two years of updates. It also lists an experimental node-based physics framework in Geometry Nodes that uses an XPBD solver.
• CG Channel reports a screen-space ray-tracing cleanup in Eevee that breaks compatibility in some scenes. It also reports 35 new compositor nodes and animation in-between operators that now work in Object Mode.

The analysis that follows: scripts written against 5.2 may call API features older versions lack, so pre-5.2 installs are a poor fit. Skills that touch rendering, lighting or weather may also look different in Eevee scenes carried over from older files.

How does it compare with other Blender and agent tools?

The site's 3D-Agent specimen covers a Blender add-on where the AI writes Blender Python from inside the app. blender-skills works the other way round. It is a set of instructions for an agent you already run in a terminal or IDE. A Blender MCP server is the optional bridge into a live session.

It also sits alongside the larger multi-engine packs. The site's 74-skill gamedev pack guide recommends installing only the skills you need, which fits here too. The --agent and --project flags make it possible to keep the pack scoped to one project.

How do you check a skill's steps before trusting it?

This checklist is drawn from the documented behaviour above. It is a procedure to follow, not a report of results.

1. Read before you install. Open the skill's SKILL.md and anything in shared and tools that it calls. Look for file deletes, network calls, and writes outside the project folder.
2. Work on a throwaway file. Use a blank scene or a copy of the .blend, with version control or a backup in place. Be most careful with an MCP connection to your open Blender, because changes land in the live session.
3. Confirm the version. Check that Blender is 5.2 or later. If the agent runs Blender in background mode, check that the bpy calls it uses exist in that version's Python API documentation.
4. Check the output, not the agent's summary. Depending on the skill:
   • UVs: overlapping islands, consistent texel density, seams where you expect them.
   • Normals and transforms: flipped faces, applied scale and rotation, correct units.
   • Bakes: enough margin, no cage artefacts, the right packing for ORM.
   • Rigs: deform weights on extreme poses, bone naming against the target skeleton.
   • Budgets: polygon counts, texture sizes and LOD steps against your own targets.
5. Re-import into the engine. Open the exported file in Unity, Unreal, Godot or Roblox and check scale, axis orientation, materials and animation clips. The README lists those export targets, but the parts read do not say which formats or engine conventions each one handles.
6. Reproduce the tests. Run python tools/run_tests.py with BLENDER_EXE pointing at your Blender, as the README describes. Read what the tests assert before treating "tested" as covering your use case.

A prompt that keeps the agent inside these limits might look like this:

Use the blender-retopology-uv skill on /scratch/crate_copy.blend only.
Do not save over any other file. Before running any script, show me
the full Python you intend to run. Afterwards, report UV island count,
overlapping islands and texel density per object.

Who is blender-skills for? A plain verdict

Likely fit: solo and small-team developers who already use Claude Code or Codex, work in Blender 5.2, and keep repeating the same steps: UVs, bakes, LODs, export checks. The pack's value is consistency. The agent gets a written procedure, not a fresh improvised one each time.

Poor fit: anyone on Blender older than 5.2, and anyone looking for one-click finished art. A skills pack gives an agent instructions. It does not generate models or textures by itself.

Caveats: the testing claim rests on the maker's own suite, which wasn't reproduced for this post. The project has six commits and a few hundred stars, so it is not an established standard. And agent-written Blender Python can overwrite files. Working on throwaway copies, scoping the install with --project, and running the 50-test suite yourself are the inexpensive ways to manage that.

More Blender coverage is under /tags/blender, and agent tooling news is under /tags/coding-agents.