The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill utilizes evaluate_script to inject custom JavaScript into a live browser session. Specifically, scripts/01-install-hook.js redefines the XMLHttpRequest.prototype.open and send methods to intercept and store JSON responses from target websites.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process large datasets from external, untrusted research websites. While the instructions mandate using filePath to store data on disk rather than passing it through the LLM context, the ingestion and assembly process constitutes an attack surface.
Ingestion points: Browser network responses captured in scripts/01-install-hook.js and subsequently processed by scripts/assemble.py.
Boundary markers: The SKILL.md instructions explicitly forbid returning raw data rows to the model context, requiring the use of the filePath parameter for disk storage.
Capability inventory: Includes arbitrary JavaScript execution in the browser via evaluate_script, local Python execution via scripts/assemble.py, and the ability to write files to the workspace.
Sanitization: The skill relies on standard JSON parsing and basic integrity checks (unique IDs and row counts) without deep content sanitization of the extracted data.
[COMMAND_EXECUTION]: The skill requires running a local Python script scripts/assemble.py to finalize the data capture and verify integrity. It also involves driving a browser via Chrome DevTools Protocol (CDP) and executing scripts to handle pagination and UI interaction.
[PROMPT_INJECTION]: The instructions in SKILL.md advise the agent to minimize safety and legal warnings regarding Terms of Service (ToS) violations. It explicitly instructs the agent to mention the risk only once and characterizes subsequent warnings as "nagging," which could lead the agent to assist in activities that violate site terms despite initial safety indicators.

ui-json-capture