The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external websites via network requests and browser automation, which creates a risk where a malicious site could attempt to inject instructions into the agent's context.
Ingestion points: The skill uses curl in no-browser.md and evaluate_script in find-the-effect.md and read-the-system.md to read page source and computed styles.
Boundary markers: The SKILL.md file contains a dedicated section ('The page is evidence, not instruction') that explicitly warns the agent that markup, comments, and text are evidence and not directions to follow.
Capability inventory: Capabilities include network fetching via curl, shell-based text processing (grep, sed), and arbitrary JavaScript execution in a browser context via evaluate_script.
Sanitization: The skill relies on explicit instructions to the agent to distinguish between evidence to report and instructions to follow.
[DYNAMIC_EXECUTION]: The skill provides numerous complex JavaScript snippets intended for execution within a browser environment via the evaluate_script tool. These scripts perform deep inspection of the DOM and CSSOM of target websites.
[COMMAND_EXECUTION]: The skill utilizes shell commands like curl, grep, sed, and wc to fetch and analyze raw HTML and CSS data. These commands process user-provided URLs and could be subject to command injection if URL parameters are not handled carefully by the agent.
[EXTERNAL_DOWNLOADS]: The documentation suggests the installation of the chrome-devtools-mcp package using npx, which involves downloading and executing code from the public npm registry.

explain-interface