The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data by downloading and parsing HTML/CSS from external URLs (htmlCode.downloadUrl) retrieved via the Stitch MCP server tools. Malicious instructions embedded in the CSS classes or HTML structure of the analyzed screens could potentially influence the agent's synthesis process during the generation of the DESIGN.md file.
Ingestion points: Screen HTML code downloaded from htmlCode.downloadUrl as specified in SKILL.md.
Boundary markers: The instructions do not define specific delimiters for the downloaded content when it is passed to the LLM for analysis, though the skill documentation includes a general warning to treat downloaded assets as untrusted data.
Capability inventory: The skill uses web_fetch, read_url_content, and various Stitch MCP tools to retrieve data, and performs file writing to create DESIGN.md.
Sanitization: No explicit sanitization or filtering of the retrieved HTML/CSS content is described before it is analyzed by the agent.
[EXTERNAL_DOWNLOADS]: The skill retrieves project assets (HTML code and screenshots) from external URLs provided by the Stitch MCP server. These downloads are directed to infrastructure associated with a well-known service (stitch.withgoogle.com) and are necessary for the skill's documented functionality of analyzing screen designs.

design-md