The Agent Skills Directory

[Robust Input Sanitization]: The skill implements a mandatory security model for all external inputs (URLs, files, images). This includes file size caps (50KB for docs, 5MB for images), path safety checks to prevent directory traversal, and stripping of prompt injection patterns (e.g., 'ignore previous instructions') before model processing.
[Network Fetching Security]: Tools like WebFetch are used to extract brand data from URLs. The skill enforces strict outbound HTTPS-only policies, blocks private/loopback IP ranges, limits redirect hops, and implements DNS rebinding defenses.
[Command Execution Safeguards]: While the skill uses shell commands for file management and preview opening, these are generally scoped to the brand/ directory and temporary directories. The orchestrator mode specifically disables automatic browser opening to prevent unexpected behavior during automated builds.
[Indirect Prompt Injection Surface]: As the skill ingests external data (Figma, Canvas Apps, web code, design docs), it is technically susceptible to indirect prompt injection. However, it explicitly mitigates this by wrapping untrusted content in <untrusted_user_content> tags and providing instructions to the LLM to treat the content as data only.
[Dependency Integrity]: The skill uses established external services like Figma's REST API and MS Learn MCP for enrichment. It requires environment variables for secrets like FIGMA_TOKEN rather than hardcoding them, following secret management best practices.

design-system