The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external websites and instructs the agent to analyze the resulting report. This file contains content extracted from the target page, such as element IDs, class names, font families, and console errors, which could be manipulated to include malicious instructions.
Ingestion points: The report.json file created by the capture script contains strings directly from the target site's DOM and console logs.
Boundary markers: No specific delimiters or instructions are provided to the agent to treat the report content as untrusted.
Capability inventory: The agent has the ability to execute shell commands (running the capture script) and modify project source code to implement design changes.
Sanitization: The capture script does not sanitize or filter extracted text before including it in the report.
[COMMAND_EXECUTION]: The skill requires execution of a vendor-provided script, scripts/capture.py, to perform site automation and image capture. It also involves running the Playwright browser installation command as a prerequisite.
[EXTERNAL_DOWNLOADS]: The skill relies on the installation of the playwright and pillow libraries. These are official, well-known packages used for browser automation and image processing.

visual-qa