The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external websites through the extract_website, get_website_result, and extract_brand_kit tools to gather design evidence. This creates a surface for indirect prompt injection if a reference site contains hidden instructions (e.g., in HTML comments, CSS, or metadata) designed to manipulate the agent's behavior during the design process.
Ingestion points: External site HTML/CSS, screenshots, and client Brand Kits (PDFs/assets) provided via the Brand API MCP.
Boundary markers: The instructions mandate consolidating all evidence into a BRAND.json file and closing all external reference sources before the 'Compose' phase, effectively isolating the untrusted content from the final building step.
Capability inventory: The skill can drive a browser, write files (HTML, JSON), save images (screenshots), and utilize various Brand API tools.
Sanitization: Employs a lookup_slop audit tool to check generated sections for common AI-generated patterns and ensure the output aligns with design rubrics rather than malicious or 'generic' instructions.

taste-director