The Agent Skills Directory

[SAFE]: The skill provides legitimate developer guidance for working with Google's Gemma models, including model selection criteria and deployment templates for various platforms.
[EXTERNAL_DOWNLOADS]: The skill references documentation and model repositories hosted on official Google (ai.google.dev) and Hugging Face (huggingface.co) domains. These sources are considered well-known and trusted within the developer community.
[COMMAND_EXECUTION]: Code assets include standard implementation patterns for model inference using Hugging Face Transformers and Google Cloud Vertex AI. These operations are intended for local or cloud-based AI application development and follow best practices.
[PROMPT_INJECTION]: The application templates in the assets directory process external user input for LLM inference. While this establishes a surface for indirect prompt injection at runtime, it is the expected behavior for the provided boilerplate code.
Ingestion points: assets/gradio-app.py (message parameter), assets/transformers-js-app.js (question input), and assets/vertex-ai-app.py (user input).
Boundary markers: Absent in templates.
Capability inventory: Capabilities are limited to text generation and model inference.
Sanitization: No input sanitization is present in the boilerplate code.

gemma-dev