The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests metadata like screen titles from the Stitch platform to use in video overlays. Although this is an ingestion point for untrusted data, the structured processing via a JSON manifest and the intended design-focused use case make this a low-risk pattern.
Ingestion points: Screen metadata fetched from the Stitch project in SKILL.md via stitch:list_screens and stitch:get_screen.
Boundary markers: Data is stored in screens.json before being used in React code templates.
Capability inventory: The skill uses file writing and command execution (remotion render) via shell commands to process data into video files.
Sanitization: No explicit sanitization of the screen text content is performed during interpolation into JSX templates.
[EXTERNAL_DOWNLOADS]: The skill downloads image assets from Google Cloud Storage and retrieves packages from the NPM registry. These are well-known and reputable services necessary for the skill's operations.
[COMMAND_EXECUTION]: The skill executes shell commands for project initialization (npm install, npm create) and video rendering (remotion render). These actions are standard for programmatic video production and are limited to the project directory.
[DYNAMIC_EXECUTION]: The skill generates React components at runtime based on provided templates to build the video composition. This is a core feature of the Remotion framework and follows intended usage patterns.
[METADATA_POISONING]: A discrepancy exists between the author identified in the context (abhicodebugger) and the installation instructions in README.md (google-labs-code). This documentation inconsistency is likely an oversight and does not affect the security or functionality of the skill.

remotion