The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcript data and media recordings during the editorial discovery phase, creating a surface for indirect prompt injection. This risk is mitigated through a non-negotiable multi-stage approval process (Gates 1-8 in intents/storyline.md), including mandatory human review of all verbatim clips and integrity checks to ensure original meaning is preserved. Ingestion points include external transcripts and media paths; capabilities include local file generation and media metadata probing; sanitization is achieved through manual verification and strict JSON schema enforcement.\n- [COMMAND_EXECUTION]: The skill includes a local helper script (scripts/render_video.py) that invokes ffmpeg and ffprobe. The script implements security best practices by using list-based subprocess execution to prevent shell injection, resolving executable paths safely via standard library functions, and strictly validating all canvas dimensions, color codes, and timestamps against predefined patterns before passing them to the rendering engine. It also includes path resolution logic that prevents the tool from overwriting source media or the approved editorial script.

podcast-production