The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes ffmpeg and ffprobe to perform technical video analysis, such as detecting scene cuts and measuring audio loudness. These commands are invoked securely in scripts/lib.py using list-based arguments in subprocess.run, which prevents shell injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data including video files and metadata JSON files (such as ASR results and character research) to generate production references. This constitutes an ingestion surface for potentially untrusted data.
Ingestion points: The agent reads video files and various JSON understanding artifacts (e.g., asr_result.json, background_research.json) from a user-specified working directory.
Boundary markers: The skill provides clear instructional boundaries to separate facts from methods. The reference_check.py script enforces these boundaries programmatically via a suite of validation rules (R1-R8).
Capability inventory: Across its scripts, the skill has the capability to execute shell commands (via ffmpeg/ffprobe) and perform file system read/write operations.
Sanitization: The skill includes a dedicated validation component in scripts/reference_check.py that scans for entity names and verbatim transcript segments using regex and n-gram matching to prevent source data leakage.

video-reference