The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, creating a potential surface for indirect prompt injection.
Ingestion points: The skill ingests untrusted data from external sources via import_media (user-provided URLs for video, audio, and images) and get_media_context (transcripts and metadata extracted from uploaded media), as described in references/tool-map.md.
Boundary markers: While the skill emphasizes following the 'intent' and 'brief' (e.g., in references/acceptance-template.md), it lacks explicit instructions to treat embedded text in transcripts or media metadata as untrusted instructions that should not be obeyed.
Capability inventory: The agent has the ability to perform network operations (import_media), modify project state (apply_composition, update_project), and store persistent preferences (remember_taste), as detailed in references/project-policy.md and references/tool-map.md.
Sanitization: The instructions do not specify sanitization or escaping protocols for content derived from transcripts before it is interpolated into captions or other visual overlays.

composing-video