The Agent Skills Directory

[PERSISTENCE]: The SKILL.md file provides instructions to append an environment variable export (GEMINI_API_KEY) to shell profile files such as ~/.zshrc or ~/.bashrc. This allows the credential to persist across shell sessions and is classified as a persistence mechanism.
[INDIRECT_PROMPT_INJECTION]: The skill defines a large attack surface for indirect prompt injection by processing external data from multiple untrusted sources.
Ingestion points: The utility scripts load external image files (screenshots) and accept user-provided text briefs to drive LLM analysis and code generation.
Capability inventory: The agent possesses filesystem write access (via the -o or --output flags) and network access to the Google Gemini API.
Boundary markers: The script prompt templates do not include explicit markers or delimiters to isolate untrusted user data from system instructions.
Sanitization: Code blocks generated by the model are extracted using regular expressions and written directly to local files without secondary validation or sanitization.
[EXTERNAL_DOWNLOADS]: The skill requires the google-genai Python library. This is an official package maintained by Google, a well-known and recognized service provider.

gemini-visual