The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external 3D scene data (GLTF/WebGPU assets) which represents a potential attack surface for indirect prompt injection via embedded metadata or structure.
Ingestion points: Assets are processed through GLTFLoader, DRACOLoader, MeshoptDecoder, and KTX2Loader as described in the loading pipeline.
Boundary markers: The instructions explicitly require the agent to validate scene structure and asset URLs against an allowlist.
Capability inventory: The skill utilizes file system tools (write_file, edit_file, create_file) and command execution (execute_command) for rendering implementation.
Sanitization: The security guardrails mandate that the agent must not allow embedded GLTF extensions, such as scripts or cameras, to auto-execute.
[EXTERNAL_DOWNLOADS]: The skill references standard industry dependencies for its primary functionality.
Fetches and manages Three.js, React Three Fiber, and the postprocessing library (pmndrs).
References specialized decoders like Draco, Meshopt, and KTX2, which are typically fetched during the build or asset loading phase.
[DYNAMIC_EXECUTION]: The skill utilizes WebAssembly (WASM) binaries for high-performance asset decoding.
Uses WASM-based decoders (Draco/Meshopt) to handle compressed geometry and textures.
The author explicitly notes the risk of executing foreign code via these binaries and instructs the agent to bundle them from pinned, verified sources and reject untrusted origins.

render-webgl-threejs