The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies extensively on calling external binaries, specifically ffmpeg and ffprobe, to perform technical analysis. Commands are used for black frame detection, freeze detection, audio stats, and generating contact sheets.
[EXTERNAL_DOWNLOADS]: The skill instructions direct the agent to install dependencies using npm install and to execute tools via npx (e.g., npx hyperframes), which involves fetching and running code from the npm registry.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted media content (video and audio) using a multimodal LLM (Gemini). This creates an attack surface where malicious instructions could be embedded in the video or audio to deceive the agent during the QA process.
Ingestion points: Rendered video files (proxy.mp4) and transcription data are uploaded to the Gemini API for analysis as described in references/layer3-semantic.md.
Boundary markers: The LLM prompt includes a manifest summary of intentional regions and the original user brief to provide context and attempt to distinguish between errors and deliberate creative choices.
Capability inventory: The skill has the capability to execute shell commands (ffmpeg/ffprobe), write diagnostic reports to the file system, and perform network requests to the Gemini API.
Sanitization: The skill utilizes responseSchema at the API level to enforce structured JSON output from the LLM, which mitigates the risk of the model producing rogue instructions for the agent to execute.
[DYNAMIC_EXECUTION]: The Python script scripts/rms_scan.py uses subprocess.run() to execute ffmpeg with arguments constructed at runtime, including the path to the video file provided by the user.

video-qa