The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions direct the user to install the claude-in-chrome extension from the official Anthropics GitHub repository to enable the required browser automation capabilities.
[COMMAND_EXECUTION]: The skill utilizes standard shell commands, specifically curl, to verify that the target development server is responding before initiating the QA pass, and mkdir to establish the output directory for report files.
[DYNAMIC_EXECUTION]: The agent is instructed to use a javascript_tool to execute predefined script templates within the browser context. These scripts perform systematic audits of the DOM, such as accessibility checks, touch target measurements for mobile viewports, and the generation of deterministic finding IDs using the browser's cryptographic APIs.
[INDIRECT_PROMPT_INJECTION]: The skill has an inherent surface for indirect prompt injection due to its core function of processing untrusted external data.
Ingestion points: Untrusted data enters the agent context through the browser automation tools read_page, read_console_messages, and read_network_requests, which capture content directly from the target web application (SKILL.md).
Boundary markers: The instructions do not specify the use of delimiters or provide explicit instructions for the agent to ignore potentially malicious commands embedded within the web page content or console logs.
Capability inventory: The skill maintains capabilities including filesystem write access (for generating JSON reports) and network navigation via the browser MCP (SKILL.md).
Sanitization: There is no mention of filtering, escaping, or sanitizing the data captured from the browser before it is analyzed by the agent.

candid-chrome-qa