The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The Layer 3 semantic QA pass (described in references/layer3-semantic.md) uses a multimodal AI model to review the video. The prompt construction for this model includes the interpolation of an 'Original editing request/instructions' field containing user-supplied content. This creates a surface for indirect prompt injection, where a user could provide a brief designed to mislead the AI, suppress error reporting, or attempt to override the QA instructions. While the prompt includes calibration rules and enforced JSON output, there is no evidence of strict sanitization or delimitation for this external input.\n- [COMMAND_EXECUTION]: The skill makes extensive use of system commands to perform technical video analysis. It uses ffmpeg and ffprobe across multiple layers (references/layer1-technical.md, references/layer2-transcript.md, references/layer4-inspection.md) and includes Python helper scripts (scripts/rms_scan.py, scripts/draw_markers.py) that invoke these tools via the subprocess module to analyze audio waveforms and render visual markers.\n- [EXTERNAL_DOWNLOADS]: The skill requires external dependencies and suggests using npm install for its internal engine and npx for running transcription tools like hyperframes. These commands fetch and execute packages from the public npm registry.

video-qa