The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided video and image files alongside natural language descriptions to perform edits, creating a potential surface for indirect prompt injection.
Ingestion points: Local media paths or external URLs provided through --param videos and --param images in SKILL.md and references/scenarios.md.
Boundary markers: Uses natural language templates and bracketed placeholders for user input, but does not specify robust delimiters or instructions to ignore potential commands embedded in media content.
Capability inventory: Orchestrates remote model calls via anycap video generate and executes local analysis using ffmpeg and ffprobe as described in references/qa.md.
Sanitization: No specific content validation or sanitization for ingested media files is mentioned.
[COMMAND_EXECUTION]: The skill uses local command-line utilities for video analysis and generation tasks.
Evidence: Execution of ffprobe for metadata verification and ffmpeg for frame extraction in references/qa.md.
Evidence: Use of the vendor's anycap CLI tool for model discovery and video editing operations throughout the skill documentation.
[EXTERNAL_DOWNLOADS]: The skill references external resources and allows for processing of remote content.
Evidence: Links to official documentation on ai.google.dev and the vendor's site at anycap.ai.
Evidence: Support for remote URLs in the anycap video generate command parameters.

anycap-gemini-omni-video-edit