The Agent Skills Directory

[SAFE]: The companion server (scripts/serve.mjs) is configured to listen strictly on the loopback interface (127.0.0.1), which prevents any external access to the brainstorming UI.
[SAFE]: Command execution is restricted to the legitimate purpose of opening the local server URL in the user's browser via open or xdg-open within scripts/serve.mjs.
[SAFE]: The skill implements a secure return channel where the browser does not communicate directly with the agent. Instead, it requires the user to manually copy and paste response prompts from the browser to the terminal, providing an inherent security boundary and audit trail.
[INDIRECT_PROMPT_INJECTION]: The skill harvests design language from local project files (e.g., DESIGN.md, CSS files, and components). While these are ingestion points for untrusted data, the risk is mitigated by the fact that the resulting HTML is rendered for user inspection and requires explicit user approval (Step 6 Gate) before any permanent changes are made to the codebase.
[SAFE]: The server logic is designed to be stateless and GET-only for content delivery, minimizing the attack surface for local interactions.

blueprint