3D Model Skill
fletcherholt/3dmodel-skill
Modelagem 3D e visualizadores web interativos com feedback iterativo
Ver no GitHubSkill para Claude Code que automatiza modelagem 3D e criação de visualizadores interativos usando Blender headless e Three.js. Implementa um fluxo fechado de geração, renderização e refinamento iterativo, com padrões profissionais para web 3D, incluindo iluminação realista, materiais PBR e controles interativos (explodido, raio-x, voos de câmera).
Instalação
npx convez-ai fletcherholt/3dmodel-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerar modelos 3D parametrizados via Fusion 360 ou Blender
- Criar visualizadores web 3D interativos com Three.js para produtos
- Produzir vistas explodidas e raio-x automaticamente
- Refinar modelos gerados por IA (Tripo, Meshy, Rodin) com quality gates
Veredito

Skill robusto e maduro para pipelines profissionais de 3D na web. Excelente documentação técnica, cobre bem o ciclo completo (geração, refinamento, entrega web), mas requer familiaridade com Blender, Three.js e modelagem 3D — não é entry-level.
README
A tradução para o português deste README ainda está na fila.
3dmodel
A Claude Code skill for making 3D models and interactive 3D viewers from a coding agent.
It captures a repeatable way of working in 3D without being able to "see" a scene by intuition: generate geometry, render it, look at the image, critique it, fix the worst defect, repeat. Alongside the method it carries the practical knowledge that makes the output good rather than mocked up: Blender headless modelling, Three.js web delivery, realistic lighting and materials, and the UI patterns for interactive product, exploded and x-ray views.
What is inside
SKILL.mdthe skill itself: when to use it, the closed feedback loop, choosing a delivery target, realism techniques, interactive viewer patterns, and the pitfalls worth knowing before you hit them.references/ai-generation.mdthe fast lane: generate a base mesh with text/image-to-3D (Tripo, Meshy, Rodin, or self-hosted TRELLIS-2 / Hunyuan3D-2.1), pick best-of-N, then retopologise, bake and optimise it; plus photogrammetry and Gaussian-splatting for real-world capture, and when to generate versus hand-model versus parametric CAD.references/realtime-3d-web.mdreusable Three.js patterns with code: a scene skeleton with shadows and environment lighting, a spherical camera fly, an x-ray reveal, fixed-lane callouts with leader lines, colour-coded flow lines, the hide-the-overlay-while-moving rule, click versus drag, compressed-asset loaders (Draco/KTX2/Meshopt) with progress, and performance budgeting.references/blender-pipeline.mdheadless Blender modelling, the professional pipeline andbpyAPI surface, materials that survive glTF export, GLB export, and the export gotchas that otherwise cost an afternoon.references/fusion360.mdparametric CAD with Autodesk Fusion 360: the sketch-constrain-feature-component-joint workflow, design-for-change practices, the Python API with code (and the centimetre/radian units gotcha), driving Fusion from an agent via an MCP bridge, sourcing real existing CAD models, and the Fusion-to-Blender interop pipeline.
Install
Copy the folder into your Claude Code skills directory:
cp -r 3dmodel ~/.claude/skills/3dmodel
Then invoke it with /3dmodel, or just ask Claude to model or recreate something in 3D, build a web 3D viewer, make an exploded or x-ray view, export a GLB, or improve how a real-time scene looks.
The short version
- Decide the target first. Anything organic or one-off is fastest generated (text/image-to-3D) then refined. A precise manufacturable part points to Fusion 360 parametric CAD. A reusable asset or a photoreal still points to Blender and a GLB. An interactive web viewer points to Three.js where you control everything. Generate or capture the base, refine in Blender, engineer precision in Fusion, ship to the web with Three.js.
- Always work to a reference and match proportion, colour, layout and the small identifying details. Looking like a mockup usually means flat surfaces, no bevels, no shadows between parts, and wrong proportions.
- Close the loop every pass. Render or screenshot, read the image, fix one thing, repeat. Never ship 3D you have not looked at.
- For realism in real time the biggest wins are soft shadow maps, bevelled edges, environment lighting, ACES tone mapping, correct PBR, and baked texture detail.
- For interactive viewers keep labels in fixed lanes with leader lines, colour code the connections between parts, and hide the overlay while the view is moving so the lines never swing.
The skill is meant to grow. New techniques and pitfalls learned on later work get folded back in.