The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes a user-supplied 'world brief' to generate a structured world_spec JSON, which is then used as input for code generation and rendering tools.
Ingestion points: The initial 'world brief' provided by the user is the primary source of untrusted data processed by the skill.
Boundary markers: The instructions mandate separating 'explicit_constraints' from 'inferred_details', which provides a logical structure, but does not include formal delimiters to prevent instruction injection within the brief.
Capability inventory: The threejs_world tool writes multiple files to the local file system (index.html, world.json, world-runtime.js) and generates executable JavaScript for the Three.js runtime.
Sanitization: There is no evidence of automated sanitization or escaping of user input before it is mapped to the world specification and subsequent code generation.
[EXTERNAL_DOWNLOADS]: The skill utilizes the threejs_asset_catalog tool to install 3D model catalogs into the project workspace from external sources. The instructions emphasize licensed and rights-safe content but verify the external fetch mechanism is active during the build phase.
[COMMAND_EXECUTION]: The workflow involves executing Python-based tools (ThreeJSWorld) and asset management tools to perform construction and rendering operations.
[DYNAMIC_EXECUTION]: The build process dynamically generates a world-runtime.js file, which contains the logic for constructing the 3D scene. This script is assembled based on parameters derived from the user's brief, representing a dynamic script generation pattern.

threejs-world-generation