The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources including YouTube links, audio/video transcripts, and OCR output from images.
Ingestion points: Data enters via user-provided media files, YouTube links, and receipt/screenshot images as defined in the SKILL.md file.
Boundary markers: The skill requires the use of structured schemas such as transcript_boundary/v1 and prepared-vs-observed boundaries to isolate extracted content from the agent's instructions.
Capability inventory: The skill can route tasks to adjacent workflow components via skill_view and utilize the omh runtime CLI for status recording (found in SKILL.md).
Sanitization: Explicit safety rules are included to mask sensitive personal information (PII) like credit card and ID numbers during receipt OCR processing.

omh-media-input