The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes external video and image data, creating a potential surface for indirect prompt injection where instructions could be embedded in the processed media.
Ingestion points: Video files and images are read via OpenCV in scripts/video_analyzer.py and scripts/model_trainer.py.
Boundary markers: No explicit boundary markers or instructions to ignore embedded instructions are present in the processing scripts.
Capability inventory: The skill utilizes Bash (python, pip, ffmpeg), file system operations (shutil.copy), and file writing (cv2.imwrite, json.dump).
Sanitization: No sanitization or validation of the visual content is performed to prevent interpretation of embedded instruction data.
[COMMAND_EXECUTION]: The skill requests permission to execute shell commands to facilitate model training and video processing.
Evidence: The allowed-tools section in SKILL.md specifies access to Bash(python*,pip*,ffmpeg*) which is used by the Python scripts to manage environment dependencies and manipulate video frames.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions for acquiring external software and research repositories.
Evidence: The skill references ultralytics on PyPI and a third-party repository github.com/abewley/sort.git for tracking functionality. While these are well-known resources in the computer vision community, they involve downloading external code at runtime.

computer-vision-pipeline