The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including project metadata and HTML code, which could be exploited for prompt injection.
Ingestion points: External content enters the agent context via the mcp_stitch:get_screen and web_fetch tools as specified in SKILL.md.
Boundary markers: There are no defined delimiters or safety warnings to separate fetched project data from the agent's instruction set.
Capability inventory: The skill possesses the Write capability to create files and the web_fetch capability for network access.
Sanitization: The instructions do not include steps to sanitize, escape, or filter the retrieved HTML source code before processing.
[EXTERNAL_DOWNLOADS]: The skill retrieves external assets from URLs provided dynamically through project metadata.
Evidence: SKILL.md directs the agent to download HTML code and screenshots from the htmlCode.downloadUrl and screenshot.downloadUrl provided by the Stitch MCP tools.

design-md