The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The evaluate_image.py script fetches pre-trained model weights from official repositories for evaluation purposes. Specifically, it downloads the openai/clip-vit-base-patch32 and Salesforce/blip2-opt-2.7b models using the transformers library, which is standard practice for ML evaluation skills.
[COMMAND_EXECUTION]: The skill includes a Python script designed for local execution by the user to analyze image quality. The script uses argparse to handle user-supplied paths and prompts safely, and it performs all heavy computation locally using libraries like torch and opencv-python-headless.
[DATA_EXFILTRATION]: Analysis of the source code confirms that images are processed locally. Images are read from the filesystem and converted to base64 strings only for the purpose of embedding them into a local HTML report. No network calls are made to send this data to external servers.
[PROMPT_INJECTION]: The skill handles user-supplied prompts which are then interpolated into model queries and an HTML report. The script includes a dedicated _escape function that sanitizes characters like <, >, and &, effectively mitigating risks of Cross-Site Scripting (XSS) or indirect injection within the generated evaluation report.

image-evaluation