The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes ffmpeg and ffprobe for video processing and analysis. These commands are executed using subprocess.run with argument lists, which is a secure method for preventing shell injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data such as ASR results and background research. Although it does not use explicit boundary markers for this data, it implements a 'leakage scanner' in scripts/reference_check.py that utilizes n-gram matching and regular expressions to identify and block the inclusion of source-specific entities and phrases in the exported production reference, effectively mitigating injection-based exfiltration risks.
[DATA_EXFILTRATION]: The skill implements strict validation rules (R1-R8) to ensure a clean separation between raw video facts and reusable production methods. It specifically checks for and prevents the inclusion of absolute file paths, system metadata, or specific entity names in the final JSON output, protecting local environment details and source privacy.

video-reference