The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates and writes executable code files, including index.html, world-spec.js, and world-runtime.js, to the local filesystem during the world construction process.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input in the form of 'world briefs' to define the 3D environment. This data is interpolated into a structured specification used to generate executable scripts.
Ingestion points: User-provided text or structured briefs are processed in SKILL.md to extract constraints.
Boundary markers: The skill requires the agent to maintain separate lists for explicit_constraints (user-supplied) and inferred_details (agent-supplied).
Capability inventory: The threejs_world tool writes multiple files and copies assets to the local projects/ directory.
Sanitization: The instructions do not specify sanitization or escaping of user strings before they are embedded into the generated world.json and .js files.
[METADATA_POISONING]: The references/worldclaw-principles.md file cites a source paper (WorldClaw) with a future publication date of 2026 and an arXiv ID (2608.05248v1) that is currently invalid or hallucinated, which constitutes misleading metadata.
[EXTERNAL_DOWNLOADS]: The skill uses the threejs_asset_catalog tool to download and install 3D model catalogs to the local project directory. The skill explicitly requires local copying to prevent remote asset fetching at render time.

threejs-world-generation