The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data from users and external sources via prompts, tool calls, and multimodal inputs. It mitigates these risks by recommending structured output validation.
Ingestion points: User-provided messages in examples/core.md, prompt strings, image data processed in examples/embeddings-vision.md, and content retrieved via webFetch in reference.md.
Boundary markers: Uses role-based message separation (system, user, assistant) to structure the model's context.
Capability inventory: Provides tools for network communication, local file reading, and model lifecycle management (pull, delete, create).
Sanitization: Explicitly mandates using zod-to-json-schema and Zod validation for extracting data from LLM responses.
[SAFE]: The skill demonstrates secure credential handling by using environment variables (process.env.OLLAMA_API_KEY) for API keys rather than hardcoding them.
[SAFE]: All library dependencies (ollama, zod, zod-to-json-schema, openai) are well-known and appropriate for the skill's purpose.
[SAFE]: File system access is restricted to reading image data for legitimate vision tasks, consistent with multimodal LLM features.
[SAFE]: Network operations are directed at the local Ollama daemon or official service endpoints, representing standard usage patterns.

ai-infrastructure-ollama