The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a 'baton-passing' pattern where the agent reads instructions for its next iteration from a local markdown file (next-prompt.md). This creates a surface for indirect prompt injection as the agent's behavior is dictated by file content that could be influenced by external data or previous iterations.
Ingestion points: Task instructions are parsed from next-prompt.md, and project roadmap context is read from SITE.md.
Boundary markers: The skill uses YAML frontmatter to separate metadata from the prompt body, but lacks specific 'ignore' directives to prevent instruction overrides within the markdown body.
Capability inventory: The agent has access to Bash, Write, Read, and the stitch toolset.
Sanitization: No sanitization or validation of the task prompt is performed before it is passed to the generation tools.
[COMMAND_EXECUTION]: The execution protocol utilizes the Bash tool to start a local development server using npx serve. This is a standard procedure for visual verification of web content but involves executing CLI commands.
[EXTERNAL_DOWNLOADS]: The skill is designed to download generated HTML and image assets from URLs (downloadUrl) provided by the Stitch MCP tool. These assets are retrieved and saved to the local file system in a staging directory (queue/).

stitch-loop