The Agent Skills Directory

[Indirect Prompt Injection] (MEDIUM): The skill is designed to take user-provided requirements and pass them to an external LLM (Gemini) for code generation. This creates a surface where malicious instructions embedded in design requirements could influence the output.
Ingestion points: The PROMPT argument in the mcp__gemini__gemini tool call (referenced in SKILL.md).
Boundary markers: Partially mitigated by the instruction OUTPUT: Unified Diff Patch ONLY, which attempts to constrain the model's output format.
Capability inventory: Uses mcp__gemini__gemini to generate UI/UX code (React, Vue, CSS). The skill explicitly recommends sandbox=True to limit the execution environment's risk.
Sanitization: No input sanitization or validation of the provided prompt is defined within the skill itself.
[Command Execution] (LOW): The skill utilizes the mcp__gemini__gemini tool. While this tool can generate and potentially lead to the execution of code, the skill's recommendation to enable sandbox=True is a standard security best practice for this capability.

gemini-delegation