The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill recommends installing the agent-browser utility from the Vercel Labs official GitHub repository as a fallback mechanism for capturing web screenshots.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface because it ingests 'reference packets' containing untrusted text such as annotations, overview descriptions, and UI copy, which are then interpolated into prompts for subagent reviewers. The skill effectively mitigates this risk by instructing the agent to treat this content as untrusted data for comparison purposes only and by requiring the redaction of all credentials, tokens, and sensitive information before processing.
Ingestion points: SKILL.md (Step 2 and Step 3).
Boundary markers: The instructions provide explicit warnings to disregard any instructions embedded within the reference data.
Capability inventory: The skill dispatches review tasks to parallel subagents using standard platform tools like task() or spawn_agent().
Sanitization: Explicit instructions are provided for the manual redaction of sensitive content prior to review dispatch.

visual-qa