The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements video playback and automated frame analysis features, which create an attack surface for instructions embedded in media content.
Ingestion points: External media URLs and content processed via AVPlayer and AVPlayerViewController (referenced in SKILL.md and references/avkit-patterns.md).
Boundary markers: The implementation patterns do not utilize delimiters or specific instructions to disregard content found within the media data.
Capability inventory: The skill specifically enables and configures allowsVideoFrameAnalysis for types such as .text and .machineReadableCode in references/avkit-patterns.md.
Sanitization: The provided patterns lack logic for sanitizing or validating information extracted from the video frames.
[METADATA_POISONING]: The skill metadata and reference sections use documentation links pointing to a parody domain rather than official vendor documentation.
Evidence: Multiple links in the References section of SKILL.md redirect to https://sosumi.ai/documentation/... instead of the expected official vendor documentation sites.

avkit