Documentation

Three.js WebGPU Skill Pack documentation

Install, route, and verify the Three.js WebGPU Skill Pack with task-focused guides for Codex, Claude Code, and existing projects.

Use these docs to install the complete skill pack, choose the smallest relevant skill set, and integrate it without replacing your project's existing rules. Installation makes the guidance available; runtime WebGPU and evidence claims still require explicit verification.

Choose the task you need to complete

These pages are operating documentation, not another product tour. Start with the task that is currently blocking adoption:

For the contract presented to coding agents, continue to the agent documentation. For an existing renderer, material, or prompt workflow that must change, use the migration guides.

Understand how the pack is intended to work

The installable product is the skills/ tree. Each published skills/<name>/ directory contains its SKILL.md plus any bundled references or agent metadata it needs. Repository examples, labs, reviews, and generated media remain repository-only and are not part of the installed payload.

Install the complete pack when you want broad routing. threejs-choose-skills selects from the installed inventory, so it is not the recommended standalone installation target. For a focused request with a known owner, invoke that subject skill directly instead of asking the router to rediscover an obvious route.

The current repository targets Three.js r185 and pins the package to 0.185.1. Canonical implementations use WebGPURenderer, TSL, NodeMaterial families, and node-based rendering APIs. Those are source constraints, not proof that an arbitrary browser, device, or existing application already satisfies them.

Keep project ownership separate from skill ownership

The pack owns specialist Three.js rendering guidance. It does not own a project's framework, router, application state, data transport, accessibility, deployment, or source-asset pipeline. Existing repository instructions remain in force, and a skill must not silently replace them.

When a required capability has no installed specialist owner, the correct result is an explicit gap or reduced scope. The router must not stretch a nearby procedural or image-pipeline skill over unsupported application work merely to produce a complete-looking answer.

Verify before making technical claims

Installation proves that instruction files are available. It does not prove the native WebGPU backend, a performance envelope, a fixed bug, or a production scene's correctness.

Before implementation, record the resolved Three.js package, runtime THREE.REVISION, initialized renderer backend, required capabilities, target devices, and product-defined error or performance gates. After implementation, inspect the relevant final image and diagnostic artifacts directly. Use the demo registry only as a published catalog; current claims require a matching source hash, source-bound evidence record, and required artifacts.

The installed skills/threejs-choose-skills/SKILL.md file defines routing fields and completion conditions. Verify each route against the current project's installed inventory and acceptance bounds. The native WebGPU validation harness defines a backend, artifact, resource, lifecycle, and readback evidence shape; confirm its current source-bound record before use. Neither contract substitutes for running the relevant checks inside the project being changed.

Know the practical requirements

  • A coding-agent runtime that can discover local skill folders or follow explicit local-file instructions.
  • Node.js for the preferred installer, repository scripts, examples, and installed-source inspection.
  • The project's actual Three.js dependency and lockfile, not a remembered version label.
  • A browser and device with the required WebGPU capabilities for full canonical runtime validation.
  • A reproducible scene state, camera, data input, or seed for any visual acceptance claim.

If one of these requirements is missing, document the narrower result that remains supportable. Do not present fallback behavior or an attractive screenshot as proof of the canonical route.

Relevant skills

Relevant demos and evidence

Sources and correction path