The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill script dynamically loads utility modules from computed paths at runtime. It includes a search mechanism that checks environment variables and scans standard plugin installation directories for its peer files.
Evidence: The script scripts/ask_gemini.ts uses await import(resolve(PLUGIN_ROOT, "utils.ts")) and shared.ts after attempting to resolve the plugin root by reading ~/.claude/plugins/installed_plugins.json and checking plugin cache directories.
[INDIRECT_PROMPT_INJECTION]: The skill accepts user-provided text and image files and transmits them to an external LLM (Gemini 3.1 Pro) for analysis. This creates a surface where malicious instructions hidden in the content being reviewed could influence the agent's behavior.
Ingestion points: The skill reads command-line text arguments and local image files (PNG, JPG, etc.) provided by the user.
Capability inventory: The script performs file system reads (for images) and makes network calls to the Google Gemini API.
Boundary markers: There are no explicit delimiters or instructions to the model to ignore potential directives embedded within the images or text strings provided for review.
Sanitization: Input content is not filtered or sanitized before being passed as a 'part' to the Generative AI SDK.

ask-gemini