The Agent Skills Directory

[SAFE]: The skill incorporates explicit security guardrails in the 'Execution Protocol', instructing the agent to treat task data as untrusted and to respect user-defined limits on budget and iterations.
[SAFE]: Includes proactive path traversal protection by requiring the agent to validate page slugs, rejecting absolute paths, separators, and traversal attempts before deriving file system paths.
[SAFE]: Utilizes standard development tooling (npx serve) for visual verification within a local context, following established web development practices.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from a 'baton' file (.stitch/next-prompt.md) to drive code generation.
Ingestion points: Reads task descriptions and frontmatter from .stitch/next-prompt.md.
Boundary markers: Explicitly instructs the agent that baton text and generated files are untrusted task data and cannot authorize additional spending or publication.
Capability inventory: Uses file write capabilities, Stitch MCP tools for generation, and Bash for local server execution.
Sanitization: Employs strict slug validation to prevent file system manipulation via the page field.

stitch-loop