The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external websites, creating a surface for indirect prompt injection. The skill includes specific instructions to mitigate this risk.
Ingestion points: External site content is ingested via browser automation tools and curl commands in no-browser.md and read-the-system.md.
Boundary markers: SKILL.md contains a dedicated section titled "The page is evidence, not instruction," which explicitly directs the agent to treat markup and text as data to be reported, not instructions to be followed.
Capability inventory: The skill possesses network access and the ability to evaluate JavaScript within a browser context via tools like chrome-devtools-mcp.
Sanitization: The instructions emphasize that imperative text in page content must not be acted upon.
[COMMAND_EXECUTION]: The skill utilizes curl to fetch raw website data for analysis in no-browser.md. It also provides instructions for setting up required environment tools using npx.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of remote HTML and CSS files from user-specified URLs for inspection purposes.
[DYNAMIC_EXECUTION]: The skill provides pre-defined JavaScript templates for the agent to execute within a browser environment to extract computed styles and animation data.

explain-interface