The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute npx serve, which downloads and runs a local development server. This is used for visual verification of generated pages.
[PROMPT_INJECTION]: The skill implements an autonomous 'baton-passing' loop where it reads task instructions from a local file (.stitch/next-prompt.md) and uses them to drive its next actions. This pattern is susceptible to indirect prompt injection if the project files are influenced by untrusted external inputs.
Ingestion points: The .stitch/next-prompt.md file (Step 1).
Boundary markers: Absent. There are no instructions to the agent to treat the data in the baton file as untrusted or to ignore embedded instructions.
Capability inventory: The skill has access to high-privilege tools including Bash, Write, and Stitch MCP tools, providing significant control over the local environment and project files.
Sanitization: Absent. The page variable from the baton frontmatter is used directly to construct file paths for move and write operations, and the prompt body is passed to generation tools without validation.
[EXTERNAL_DOWNLOADS]: The skill fetches content (HTML and PNG screenshots) from dynamic URLs provided by the Stitch MCP tools and saves them to the .stitch/designs/ directory. While these are expected outputs of the service, it involves the retrieval and storage of external content.

stitch-loop