The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external text content through a JSON variable to populate the display of the carousel cards.
Ingestion points: The script in cuboid-carousel.html ingests untrusted data from the cardsJson variable via the window.__hyperframes.getVariables() call.
Boundary markers: Text content from the JSON is interpolated directly into the canvas rendering context (e.g., ctx.fillText) without boundary markers or specific delimiters intended to prevent instructions from being followed by downstream agents.
Capability inventory: The skill is restricted to WebGL and Canvas rendering and does not possess capabilities for file system access, shell execution, or non-whitelisted network operations.
Sanitization: Input is parsed as JSON and basic type validation is performed on fields (color, title, etc.), but no sanitization of the text strings themselves is performed.
[EXTERNAL_DOWNLOADS]: The skill references and downloads external assets and tools from well-known and vendor-owned sources.
Evidence: Fetches web fonts from Google Fonts domains (fonts.googleapis.com and fonts.gstatic.com).
Evidence: Installation instructions in SKILL.md describe using npx to install the hyperframes package, which is a resource owned by the vendor (heygen-com).

cuboid-carousel