Guides
A Claude skill for game-ready asset checks: poly count, UVs, scale
Build a Claude skill that checks game assets in Blender: triangle budgets, UVs, scale, normals, non-manifold edges, textures and naming.
6 min read

Quick answer
What is a Claude skill for game asset QA?
It is a checklist Claude runs on a 3D file before it goes into a game engine: triangle count against budget, materials, UV maps, applied scale, origin, normals, non-manifold edges, missing textures and naming. Through Blender MCP, Claude runs the checks on the open scene and returns a pass/fail report with the numbers.
Why QA is the best first skill
Asset problems are cheap to fix in Blender and expensive to find in the engine. A prop with unapplied scale imports 100 times too big. A mesh with two UV maps confuses a lightmapper. A texture path pointing at someone's Downloads folder shows up as a pink material in a build.
None of these checks need judgement. They're facts you can read off the file, which makes them exactly the work to hand to Claude with a skill. It's also the safest skill to start with, because a QA pass reads and reports without changing anything.
What the checklist covers
Group the checks so the report is easy to scan. A solid default for a PC or console game:
Geometry
- Triangle count per object and per asset, against the budget for its class.
- N-gons: faces with more than four sides, which triangulate unpredictably on export.
- Non-manifold edges: in Blender, Edit Mode, then Select, Select All by Trait, Non Manifold. These break some tools and all 3D printing.
- Loose geometry: stray vertices and edges with no faces.
- Normals: faces pointing inward. They show as holes or black patches in-engine.
Transforms
- Applied scale and rotation: scale should read 1, 1, 1 before export.
- Origin: at the bottom centre, or wherever your standard says.
- Location: at the world origin for single assets.
UVs and textures
- A UV map exists on every mesh that has a texture.
- UV map count: one for textures, plus a second only if your engine needs a lightmap channel.
- Missing images: textures whose file isn't on disk.
- Texture resolution against your budget.
Materials and naming
- Material count per asset.
- Names match your convention, with no
.001suffixes and no defaultMaterial.
Rig (characters only)
- An armature is the parent of the mesh.
- Every vertex is weighted to at least one bone.
- Bone count is within your engine's or platform's limit.
The skill files
.claude/skills/asset-qa/SKILL.md: the procedure.claude/skills/asset-qa/budgets.md: budgets per class and per platform.claude/skills/asset-qa/scripts/qa_report.py: a tested script that prints every number
The SKILL.md:
---name: asset-qadescription: Use when checking whether 3D assets are game-ready, before export or handoff, or when asked to QA, review or validate a model.argument-hint: [pc | mobile]---1. This is read-only. Do not change the scene.2. Run scripts/qa_report.py on the selected objects, or all meshes if none are selected.3. Read budgets.md. Use the platform passed as an argument; default to pc.4. Report per asset: PASS or FAIL, then each failed check with the actual and allowed value.5. End with the three fixes that would clear the most failures.
Step 1 matters. Keep QA read-only, and let a separate skill or a direct request do the fixing. That way you can run QA on anyone's file without worrying. More on writing these: How to write a SKILL.md for Blender.
Why the numbers come from a script
Claude can work out each number by writing new Python every run. A fixed script is better for QA:
- the same file gives the same report every time
- you decide once what counts (triangles after modifiers, say)
- one call returns everything, and only the printed output enters the conversation
Ask Claude to write the first version of qa_report.py, then test it in Blender's Scripting workspace on files where you already know the answers. After that, the skill runs your tested copy. Can Claude write Blender Python scripts? covers how to review what Claude writes.

Running it
With Blender MCP connected (setup) and the file open:
/asset-qachecks the selection against PC budgets./asset-qa mobilechecks against mobile budgets.- "Is the Props collection game-ready?" works too, because the description covers "game-ready", so Claude loads the skill.
A useful report reads like this, with your own names and numbers:
- SM_Barrel_A: FAIL. Triangles 31,850 (budget 3,000). Scale not applied.
- SM_Crate_B: PASS.
- SM_Lantern: FAIL. Missing texture
T_Lantern_BaseColor.png. - Top fixes: reduce SM_Barrel_A, apply scale on 4 objects, relink 1 texture.
Adding a fix pass
Once you trust the report, add a second skill, asset-fix, that fixes only the safe items: apply transforms, set origins, recalculate normals, and rename. It should ask before anything that changes the look, such as reducing polycount, merging materials or deleting loose parts.
For polycount, ask Claude to work on a duplicate with Decimate (Collapse, with a ratio) or Remesh, and show a screenshot of both. For a cleaner result on generated props, run the reduction where the model was made. Vi3W's Clean shape rebuilds the mesh at a chosen size from 500 to 20,000 faces and bakes the textures onto the new mesh. With the Vi3W MCP server connected (it launched today), the fix skill can request that step directly. The trade-offs are covered in Decimate vs retopology.
Per-engine notes worth adding
- Unity: FBX scale and axis settings, and whether you rely on the Import Settings scale factor.
- Unreal: the
SM_andSK_prefixes, and that a collision mesh uses theUCX_prefix if you author your own. - Godot: GLB is the usual choice. Check that materials survive the import.
Keep these in budgets.md or a separate engines.md, so the skill loads them only when needed.
What a QA skill can't catch
- Visual problems: bad silhouettes, texture seams in obvious places, a wrong art style. Those need a person, or at least a screenshot and a clear question.
- Engine-side issues: import settings, shaders or LOD set-up after the file leaves Blender.
- Assets exported without Claude. For a hard gate, run
qa_report.pyas part of your build or import process too.
A complementary human checklist is in AI 3D game asset approval checklist.
Try it on generated assets
Generated models are the best test for a QA skill, because they fail in consistent ways: dense meshes, a single big material, and scale that's arbitrary. Make a few in Vi3W, download GLB, and run your skill on them.
Frequently asked questions
claude skills with scripts
Yes, a skill can include scripts in its folder. Claude runs them and only their output enters the conversation. For asset QA that's ideal: a tested script prints the numbers, and Claude compares them with your budgets.
claude skills with parameters
In Claude Code, skills accept arguments: declare them with the arguments and argument-hint frontmatter fields and pass them after the command, for example /asset-qa mobile to check against mobile budgets instead of PC ones.
Last reviewed

