The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses ollama.pull() to download Large Language Model (LLM) weights from remote registries to the local machine.
Evidence: examples/model-management.md demonstrates pulling models like 'llama3.1' with streaming progress.
[COMMAND_EXECUTION]: The skill provides patterns for programmatic control of the local Ollama daemon, including model management and an option to bypass SSL verification.
Evidence: ollama.delete(), ollama.copy(), and ollama.create() methods are used to manage local resources.
Evidence: reference.md mentions the insecure: true flag for ollama.pull(), which allows downloads from unverified servers.
[INDIRECT_PROMPT_INJECTION]: Several patterns involving data extraction from untrusted text or images create a standard indirect prompt injection surface.
Ingestion points: Text inputs in extractSummary (examples/structured-output.md) and image paths in analyzeImage (examples/embeddings-vision.md).
Boundary markers: Uses system messages for role definition, but examples lack explicit delimiters (e.g., XML tags or triple backticks) for separating untrusted user content from instructions.
Capability inventory: The skill can download, create, and delete local models (ollama.pull, ollama.delete, ollama.create).
Sanitization: Employs zod-to-json-schema and Zod.parse for strict output validation, which mitigates but does not fully eliminate the risk of model instruction override.

ai-infrastructure-ollama