The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides numerous Python scripts (e.g., cut-from-edl.py, probe-media.py, detect-segments.py) that interface with the ffmpeg and ffprobe binaries. Analysis of these scripts confirms that they utilize subprocess.run with argument lists instead of raw shell strings, which is a secure implementation that prevents shell injection attacks.
[INDIRECT_PROMPT_INJECTION]: The skill documents an 'Edit-as-Code' workflow where the agent is expected to read and reason over transcripts generated from media files (references/edit-as-code.md). This represents a potential surface for indirect prompt injection, as adversarial text contained in a video's audio or metadata could influence the agent's logic during the selection or verification phases.
[DATA_EXFILTRATION]: While the skill includes network-adjacent tools like yt-dlp for downloading media, it does not contain any patterns for exfiltrating local data. The documentation explicitly advises users to strip metadata (-map_metadata -1) when processing untrusted uploads to prevent accidental data exposure.
[PROMPT_INJECTION]: The skill's instructions in SKILL.md include directives for the agent to follow 'Doctrine: probe first' and 'Footguns' rules. These are functional instructions intended to ensure operational success and output quality, rather than attempts to bypass safety guardrails or override system instructions.

ffmpeg-ops