The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads HTML templates and visual assets from the Google Stitch SDK.
Evidence: const htmlUrl = await screen.getHtml(); and const imageUrl = await screen.getImage(); which target Google's infrastructure.
[COMMAND_EXECUTION]: The skill utilizes Bash for environment validation and to run a local web server for visual verification of the generated site.
Evidence: ls node_modules/@google/stitch-sdk and npx serve site/public -p 3456.
[INDIRECT_PROMPT_INJECTION]: The skill implements an autonomous loop that reads instructions from a local file (.design/next-prompt.md) and processes external HTML content, creating a potential attack surface for embedded instructions.
Ingestion points: Reads from .design/next-prompt.md and remote HTML content fetched via the Stitch API.
Boundary markers: The baton file uses YAML frontmatter to separate metadata from the prompt body, but no explicit sanitization is defined for the downloaded HTML content.
Capability inventory: The skill has Write, Edit, and Bash capabilities allowed in its frontmatter.
Sanitization: No explicit filtering or sanitization logic is provided for external design content before it is processed by the agent.

design-loop