The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill maintains an attack surface for indirect prompt injection by ingesting and processing untrusted data provided by users, such as visual references, URLs, and text descriptions.
Ingestion points: Data enters the agent context through user responses during the 'Briefing' and 'Reference Collection' phases in SKILL.md and references/collection-guide.md.
Boundary markers: The skill does not implement specific delimiters or 'ignore' instructions for external content.
Capability inventory: The agent can generate code (HTML/CSS/JS) in SKILL.md (Phase 6) and perform network operations via Google Stitch MCP tools in references/stitch-integration.md.
Sanitization: No explicit sanitization or validation logic is defined for the interpolated user content.
[EXTERNAL_DOWNLOADS]: The skill references and integrates with Google Stitch (stitch.withgoogle.com), a well-known service for UI prototyping.
The integration uses the official Google endpoint https://stitch.googleapis.com/mcp for its Model Context Protocol (MCP) server.
The skill provides clear instructions for users to manage API keys securely, emphasizing that they should not be committed to public repositories, which aligns with standard security best practices.

frontend-blueprint