The Agent Skills Directory

[Official SDK Usage]: The skill recommends installing official Google Gen AI SDKs for multiple programming languages. These packages are the primary interface for the platform and are provided by the vendor.
[Credential Safety]: Code samples demonstrate how to use environment variables and Application Default Credentials (ADC) to manage API keys and project identifiers, avoiding the risks associated with hardcoded secrets.
[Model-Generated Code Execution]: The skill features the ToolCodeExecution tool, which enables models to generate and run Python code. This is an intended capability of the API for performing complex calculations and data processing.
[Prompt Behavior Modification]: The instructions contain directives aimed at updating the agent's internal knowledge regarding model versioning, such as claiming legacy models are deprecated and knowledge is outdated. This is a common pattern in technical documentation for rapidly evolving AI platforms.
[Indirect Prompt Injection Surface]: By enabling the processing of external data (multimodal files, YouTube videos, and web search results) alongside powerful tools like function calling and code execution, the skill introduces a surface for indirect prompt injection.
Ingestion points: Multimodal inputs, batch processing files, and external URLs.
Boundary markers: Not explicitly shown in prompt templates, but the skill includes a dedicated safety configuration reference.
Capability inventory: Includes code execution, search grounding, and function calling across various modules.
Sanitization: Relies on the platform's configurable safety filters to manage content risks.
[Experimental Remote Tooling]: Documentation for the Model Context Protocol (MCP) includes a sample command that executes a remote tool via npx. This demonstrates the platform's extensibility but requires users to trust the source of the remote tool.
[Safety Threshold Configuration]: A dedicated safety reference explains how to adjust filters for dangerous content, harassment, and hate speech, allowing developers to align model behavior with their application's requirements.

gemini-api