The Agent Skills Directory

Indirect Prompt Injection (MEDIUM): The skill processes untrusted external data which creates an attack surface for indirect prompt injection.
Ingestion points: Loads scene-plan.json from the creative-director agent, video assets from video-asset-manager, and lyrics from external files.
Boundary markers: Absent. There are no delimiters or system instructions to ignore embedded commands within these data sources.
Capability inventory: The skill has the capability to execute shell commands (npm, npx, node) and perform file system writes to build the project.
Sanitization: Absent. Data from the scene plan and lyrics are interpolated directly into the rendering components without validation.
Dynamic Execution (MEDIUM): The workflow requires runtime compilation and bundling of React code via npx remotion render to generate the final video output. This execution of dynamically assembled code is a known risk vector for code injection if the source files are compromised.
External Downloads (LOW): The skill installs dependencies from npm (remotion, @remotion/cli). These are widely recognized packages and the risk is downgraded per the trust scope rule, though they still represent an external dependency.
Command Execution (LOW): Routine use of subprocesses to run helper scripts (convert_scene_plan.js) and build tools. While functional, it provides a vector for exploitation if arguments are not properly sanitized.

music-video-producer