The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external media files (video and audio) which could contain embedded text or speech designed to influence the behavior of the agent reading the resulting observation reports.
Ingestion points: The video parameter in SKILL.md accepts external files for decoding, frame extraction, and transcription.
Boundary markers: The instructions lack explicit directives to use delimiters or ignore instructions found within the analyzed content, though it specifies that the skill should return 'observations without edits or verdicts'.
Capability inventory: The skill reads local files, performs transcription (via a Whisper backend), and writes reports, manifests, and image files to the local file system (skills/test-runs/).
Sanitization: The skill does not describe any sanitization or escaping of extracted text (e.g., ASR transcripts or OCR from frames) before including it in the observation.md report.

watch