The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from the live editor (entity names, component values, and asset catalog metadata) which could contain adversarial instructions intended to influence the agent's behavior.
Ingestion points: Data is ingested via scene_state (entity list), entity_detail (component values), and search_catalog (external asset metadata).
Capability inventory: The agent can create, delete, and modify entities; write script files to assets/Scripts/; and execute scene mutations through the Creator Hub MCP server.
Boundary markers: No specific delimiters or instructions for isolating untrusted data from the command context are defined.
Sanitization: The skill lacks explicit instructions for sanitizing or validating entity names and property values retrieved from the editor.
[EXTERNAL_DOWNLOADS]: The place_smart_item tool downloads 3D models and configuration files from the Decentraland catalog to the local assets/asset-packs/ directory. This is standard functionality for the Creator Hub and involves vendor-controlled resources.
[COMMAND_EXECUTION]: The skill provides documentation for users to execute CLI commands (e.g., claude mcp add) to register the local MCP server. These commands are intended for manual environment setup and are not executed autonomously by the skill code.

creator-hub-mcp