The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external, untrusted data including product URLs and style reference media (images/videos). This data is used to generate prompts for AI video and image models, creating a surface where malicious instructions could be injected indirectly via the external content.
Ingestion points: Processes PRODUCT-BRIEF.md, user-provided URLs, and style reference files.
Boundary markers: The skill lacks explicit instructions to delimit or ignore instructions within external data.
Capability inventory: Access to shell commands (ffmpeg, yt-dlp), file system operations, and AI generation APIs.
Sanitization: No explicit sanitization or filtering of external content is defined.
[COMMAND_EXECUTION]: The skill provides instructions to run various shell commands, including yt-dlp for downloading content, ffmpeg for media manipulation, and perl for process management. These represent a significant execution surface.
[DYNAMIC_EXECUTION]: The skill suggests using a perl wrapper to execute arbitrary commands with a timeout (exec @ARGV). It also describes a process for generating and assembling media assets dynamically at runtime.

vid-ad