The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external style definitions (style.md) from user-provided reference sets, creating a surface for indirect prompt injection.
Ingestion points: Reference set directories containing style.md files (e.g., generations/refs/brand/style.md).
Boundary markers: The instructions explicitly direct the agent to treat these files as untrusted style data and ensure they do not override user instructions or access private files.
Capability inventory: The skill performs file system writes to the generations/ directory and initiates network requests to Google's API.
Sanitization: The agent is instructed to manually review style data before inclusion in the generation prompt.
[EXTERNAL_DOWNLOADS]: The skill documentation references the installation of the official google-genai Python package and communicates with Google's Generative AI infrastructure.
Details: Fetches content from Google's official API at generativelanguage.googleapis.com and references documentation at ai.google.dev.
[COMMAND_EXECUTION]: The reference files contain example shell commands (curl) and Python scripts for interacting with the Gemini API.
Details: These examples involve reading environment variables for API keys and writing binary image/video data to the local filesystem within the generations/ folder.

generate-nanobanana