The Agent Skills Directory

[DATA_EXFILTRATION]: The skill incorporates strict guardrails prohibiting the transmission of clinical data to any external endpoints. It explicitly requires that patient data and extraction results remain on the local machine to prevent unauthorized disclosure, addressing risks associated with cloud storage and external APIs.
[COMMAND_EXECUTION]: The skill uses run_shell_command to execute a local NLP script. It includes specific mitigations for command injection, such as instructions to handle shell metacharacters and constraints on where output files can be written (Current Working Directory or user-specified local paths only).
[PROMPT_INJECTION]: The skill identifies the risk of instructions embedded within clinical notes influencing agent behavior. It mitigates this via a 'Clinical Decision Boundary' guardrail, which ensures the agent treats input text as data to be structured rather than clinical guidance or executable instructions.
Ingestion points: Clinical text processed via the --text argument or read from files (SKILL.md).
Boundary markers: Prompt-based boundaries defined in the 'Security Guardrails' section.
Capability inventory: Command execution (run_shell_command) and file reading (read_file).
Sanitization: Explicitly addressed through the 'Command Injection Prevention' and 'File Access Scoping' guardrails.

medical-terms-hardened