The Agent Skills Directory

[PROMPT_INJECTION]: The skill relies on an autonomous loop where instructions are read from a local baton file (.stitch/next-prompt.md). This architecture presents an indirect prompt injection surface if the baton is influenced by untrusted data.
Ingestion points: .stitch/next-prompt.md in SKILL.md.
Boundary markers: The skill does not implement specific delimiters or instructions to treat the baton content as untrusted data.
Capability inventory: Bash shell access, filesystem Write access, and Stitch project management tools.
Sanitization: Not specified; the agent is instructed to parse and follow the baton content directly.
[COMMAND_EXECUTION]: The execution protocol involves running npx serve via Bash to perform visual verification of the generated content.
[EXTERNAL_DOWNLOADS]: The skill downloads assets (HTML and screenshots) from dynamic URLs provided by the Stitch MCP tools.

stitch-loop