The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core purpose is to analyze video content, which is an untrusted data source. It performs visual analysis on frames to score criteria like legibility and word counts, creating a surface where malicious instructions could be embedded in the video content to attempt to influence the agent's evaluation.
Ingestion points: The skill reads multiple metadata files (planes.txt, stats.txt, gestures.txt, cuts.txt) and image files (still-*.png, beat-NN.png, phone-390.png) generated by a bash script from a user-provided video (SKILL.md, references/review-pack.md).
Boundary markers: The agent workflow mandates a specific sequence (restate message, check hard-fails, score rubric) and requires explicit frame references for every finding, which constrains the impact of potential adversarial content (SKILL.md, references/verdict-format.md).
Capability inventory: The agent uses standard file-reading tools and CLI utilities (ffprobe, ls) to interact with the review pack; it does not have capabilities for network operations or unauthorized file writes outside the project directory.
Sanitization: The review process utilizes signal-processing extractions (such as area-weighted color planes and motion energy peaks) rather than generic OCR of all visible text, which narrows the attack surface.

motion-review