The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes sensitive medical records (conditions, medications, lab results, clinical notes) retrieved from the HealthEx MCP server. This external data represents an ingestion surface where instructions could be hidden by a malicious actor within the health record data (e.g., in a clinical note).
Ingestion points: healthex mcp-call is used to fetch data for subcommands like get-medications, get-conditions, and get-labs in SKILL.md.
Boundary markers: The instructions command the agent to explicitly state "Based on your HealthEx records..." when presenting insights, which provides a source attribution boundary.
Capability inventory: The agent can execute subcommands of the healthex binary to query and summarize medical data.
Sanitization: The instructions strongly emphasize not fabricating clinical facts ("No Real Data -> Never Fabricate") and strictly handling process handles/placeholders, though they do not specify how to sanitize natural language text within the records themselves.

healthex