The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes JSON outputs from AI models to filter and categorize screenshots, which creates a potential surface for indirect prompt injection. However, this is mitigated by rigorous security logic:
Cryptographic Verification: The apply_vision_review.py and apply_evidence_review.py scripts re-verify the SHA256 hashes of every image file referenced in the model's response, ensuring no frames were altered or replaced by the model.
Non-Destructive Design: Original screenshots are never modified; all model-driven outputs are written to isolated new directories (_curated or _evidence_leads).
Sanitization: The application logic strictly validates paths using relative_to(root.resolve()) to prevent directory traversal attacks originating from malicious model instructions.
[EXTERNAL_DOWNLOADS]: The skill utilizes uv run to manage its execution environment. This may trigger the download of well-known Python packages like Pillow and rapidocr-onnxruntime from official package registries (PyPI). These downloads are standard operations for the tool's intended functionality.
[COMMAND_EXECUTION]: The skill invokes ffmpeg and ffprobe to handle video frame extraction. These subprocess calls are implemented using structured argument lists (not shell strings), which prevents command injection and aligns with the tool's primary purpose of video analysis.

video-screenshot