The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external URLs or HTML files provided as arguments.
Ingestion points: The agent navigates to an external resource ($0) using Chrome DevTools and analyzes the resulting state through screenshots and console message listings.
Boundary markers: The instructions lack explicit delimiters or "ignore instructions" wrappers to prevent the AI from following commands embedded in the visual text of the screenshot or in console logs.
Capability inventory: The skill has access to shell execution (Bash), file reading (Read), and browser-side script execution (mcp__chrome-devtools__evaluate_script).
Sanitization: No sanitization or filtering is performed on the content rendered by the browser before it is presented to the multimodal model.
[COMMAND_EXECUTION]: The skill uses the Bash tool to execute shell commands.
Evidence: The pipeline instructions include ts=$(Bash(date -u +%Y%m%dT%H%M%SZ)). While currently limited to a benign date utility for timestamping, the use of a general-purpose shell tool presents an execution surface.
[DYNAMIC_EXECUTION]: The skill executes dynamic JavaScript within the browser context via an MCP tool.
Evidence: The skill defines a significant block of asynchronous JavaScript code for auditing slide overflow and DOM element visibility, which is passed to mcp__chrome-devtools__evaluate_script at runtime.

verify-artifact