Guides
Can a Claude skill enforce your studio's naming, scale and export rules?
Build a Claude skill that applies your 3D asset standards: naming, scale, pivots, polycount and export settings, with a report you can verify.
6 min read

Quick answer
Can a Claude skill enforce 3D asset standards?
It can apply them and check them, but not guarantee them. A standards skill holds your naming, scale, pivot, polycount and export rules, and tells Claude to fix files to match and then report what passed. Pair it with a script for hard checks, and still review the report before assets ship.
The problem it solves
Every 3D team has standards. Most live in a wiki page that nobody opens, so they're applied from memory and drift with every new hire. Assets turn up as Cube.017 with unapplied scale and the origin floating a metre off the ground.
A Claude skill can't make people read the wiki. It can make Claude apply the wiki every time it touches an asset. When Claude is the one importing, cleaning and exporting, through Blender MCP or a 3D generation server, the standards get applied by default instead of by memory.
Apply, check, report
A standards skill should do three distinct things, in this order:
- Apply the rules Claude can fix safely: apply transforms, set origins, rename, sort into collections.
- Check the rules it shouldn't fix silently: polycount over budget, too many materials, missing textures, non-manifold edges.
- Report both lists in a form you can verify: object names, numbers, pass or fail.
The report is the part people skip, and it's the part that makes the skill trustworthy. Claude follows written instructions well, but a skill is instructions, not a validator. A report with numbers lets you catch the times it didn't.
What goes in the standards
Start from what your team already argues about. A typical game-art set:
Scale and orientation
- 1 Blender unit = 1 metre.
- Reference heights: door 2.1 m, character 1.8 m, table 0.75 m.
- Front of the asset faces -Y. Origin at bottom centre, except doors (at the hinge) and weapons (at the grip).
Naming
- Static meshes
SM_<Asset>_<Variant>, skeletal meshesSK_<Asset>, materialsM_<Asset>, texturesT_<Asset>_<Map>. - No spaces, no
.001suffixes in exported names. - A deeper convention you can adapt: AI 3D asset naming conventions.
Budgets
- Triangles per asset class: small prop, hero prop, character, environment piece.
- Materials per asset, and texture resolution per class.
- Background on setting these: AI 3D asset poly budget planning.
Export
- Format per engine (FBX or GLB), and the exact exporter settings: Apply Scalings, Forward and Up axes, Apply Transform, Smoothing, Add Leaf Bones, Path Mode with Embed Textures.
- An output path relative to the
.blend, such as//export/<Asset>/.
Style
- The look that keeps a library coherent: palette, how stylised, edge wear, texel density. This is harder for Claude to check, but worth writing down for generation prompts. See AI 3D asset style guides.
How to structure the skill
Keep SKILL.md short and put the standards in their own files, so Claude reads them only when needed:
.claude/skills/asset-standards/SKILL.md: the procedure.claude/skills/asset-standards/standards.md: scale, orientation and naming.claude/skills/asset-standards/budgets.md: numbers per asset class.claude/skills/asset-standards/export.md: settings per engine.claude/skills/asset-standards/scripts/report.py: a tested script that prints the facts
The SKILL.md itself can be this short:
---name: asset-standardsdescription: Use when importing, cleaning up, checking or exporting 3D assets for this project, or when asked whether an asset meets our standards.---1. Get the scene summary. Identify the assets in scope and their class (small prop, hero prop, character, environment).2. Read standards.md. Apply transforms, set origins and rename. List every rename.3. Run scripts/report.py and compare its numbers with budgets.md. Do not reduce polycount without asking.4. If exporting, follow export.md for the target engine.5. Reply with two lists: Fixed, and Needs a decision. Include the numbers.
Writing tips for each part are in How to write a SKILL.md for Blender.

Why a script for the checks
Claude can count triangles by writing Python each time. A tested report.py is better for three reasons:
- Consistency. It counts the same way every time: triangles after modifiers, or before, as you decide.
- Speed. One call returns every number, and only the output enters the conversation.
- Trust. You've read the script. You know what "pass" means.
Have the script print, per object: name, triangle count, material count, whether scale is applied, UV map count, and image textures missing from disk. Claude then compares those numbers with your budgets and writes the report.
Skill, rules file or both?
In Claude Code, CLAUDE.md is loaded in every session. That makes it right for two or three project-wide facts ("This is a Unity project. Units are metres. Asset standards are in the asset-standards skill") and wrong for a 200-line checklist that would sit in context during unrelated work.
The skill holds the detail and loads only when you're working on assets. Mentioning the skill in CLAUDE.md makes it easier for Claude to reach for it.
Making it a team standard
The real win is that everyone's Claude follows the same rules:
- Commit the skill folder in the project repository. Changes to standards go through review like code.
- Put a version line at the top of
standards.mdso you can see which version produced an export. - When someone finds a rule Claude keeps missing, fix the skill, not the chat.
Sharing Claude skills across a 3D team covers distribution, including claude.ai uploads for people who don't use Claude Code.
Standards for generated assets
Generated models need standards more than hand-made ones, because they arrive fast and from many sources. Extend the skill with an intake step for AI output:
- Check the triangle count. Raw image-to-3D output is often tens of thousands of triangles. Vi3W's real TRELLIS tests landed around 31,000–32,000 before retopology.
- Bring it to budget. In Vi3W, Clean shape rebuilds the mesh to a chosen size (500 to 20,000 faces) and bakes the textures onto it. In Blender, use Decimate or Remesh on a copy.
- Apply scale and origin, and rename to the convention.
- If it's a character, rig it and check the deformation. Vi3W's Add bones handles humanoids and, with It's an animal, creatures.
With the Vi3W MCP server connected (live from today), the skill can tell Claude to run the clean-up in Vi3W before the model reaches Blender. See Claude skills for 3D AI.
What it can't enforce
- Taste. It can check that a material is named
M_Crate, not that the crate looks right. - Work done without Claude. Files exported by hand skip the skill. For a hard gate, run
report.pyin CI or as an import check in your engine. - Perfect obedience. Instructions are followed most of the time, not always. That's why the report exists.
Start small
Write down five rules your team breaks most often, put them in a skill with a report step, and use it for a week. Then add budgets. For a steady supply of assets to test it on, generate some props in Vi3W.
Frequently asked questions
claude skills vs rules
Rules files are always-on instructions loaded at the start of every session; skills load only when a task matches. Put two or three project-wide basics in rules, and the detailed asset standards in a skill so they don't crowd every chat.
skill.md vs claude.md
CLAUDE.md is Claude Code's project memory, read every session. A SKILL.md is read only when its description matches the request. Keep "we use metres and Unity" in CLAUDE.md, and the full asset checklist and export recipe in a skill.
Last reviewed

