The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes Playwright-based browser tools (prefixed with mcp__playwright__) for automation. This includes the use of browser_evaluate for running scripts in the browser to perform programmatic checks, which is standard for UI auditing tools.
[EXTERNAL_DOWNLOADS]: The skill navigates to external URLs to perform audits. While this involves processing external content, it is the primary intended function of the tool and no hardcoded malicious domains were found.
[SAFE]: The skill accesses project-specific design rules and tokens (e.g., .Codex/rules/components.md and packages/core/tokens/). This access is localized to the project environment and does not target sensitive system credentials or private configuration files.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface inherent to its function of processing untrusted web content.
Ingestion points: Processes external web content and accessibility trees via browser_snapshot in SKILL.md.
Boundary markers: No explicit delimiters are defined to separate ingested web content from agent instructions.
Capability inventory: The skill has access to extensive browser interaction tools including browser_click, browser_type, and browser_evaluate in SKILL.md.
Sanitization: No sanitization of ingested web data is mentioned before it is processed by the agent.

ui-audit-guide