The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes several utility scripts (e.g., scripts/ffmpeg-preflight, scripts/audio-inspect, scripts/media-intake, and scripts/extract-review-frames) that execute ffmpeg and ffprobe binaries via Python's subprocess.run and subprocess.Popen.
These executions are performed using list-based arguments rather than shell strings, effectively preventing shell injection vulnerabilities.
The scripts implement robust security controls, including execution timeouts, non-mutating inventory probes, and hard limits on output size (e.g., MAX_OUTPUT_BYTES_HARD in scripts/media-intake and MAX_DIAGNOSTIC_BYTES in scripts/extract-review-frames).
[INDIRECT_PROMPT_INJECTION]: The skill identifies an attack surface for indirect prompt injection as it is designed to process untrusted data, specifically local media files and transcript JSON artifacts.
Ingestion points: Data enters the workflow through tools like scripts/audio-inspect (via the --transcript argument) and scripts/media-intake (via file path scanning).
Boundary markers: The SKILL.md instructions mandate the use of explicit evidence classes (e.g., "Editorial heuristic" vs. "Observed artifact") and require a "needs_review" status for all editorial decisions. Templates like edit-decision-list.json include confidence and review status fields to maintain transparency.
Capability inventory: The skill possesses capabilities for file writing, frame extraction, and subprocess execution.
Sanitization: The skill follows best practices for sanitization, such as scripts/vision-review-handoff which uses opaque asset IDs to prevent leaking private paths, and scripts/render-edl which includes a dedicated escape_ffconcat_path function to sanitize paths in generated ffconcat files.
[SAFE]: The skill demonstrates a high degree of security maturity. It emphasizes non-negotiable safety checks, explicit stream selection, and the use of private workspaces for sensitive operations. It explicitly forbids the interpretation of identity, intent, or whole-program meaning from technical metadata alone, forcing human-in-the-loop verification for consequential editorial claims.

ffmpeg