The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests local JSON files containing clinician decisions and interventions. While these natural language fields could theoretically contain instructions intended to influence the agent (e.g., during transcription), the risk is well-mitigated by the skill's narrow documentation-only scope and lack of network or execution capabilities.\n- Ingestion points: Processes local manifest files like 'source_fact_manifest.json' and 'clinician_authored_intervention.json'.\n- Boundary markers: Implements a 'Hard safety boundary' and 'Required visible notice' to strictly limit agent behavior to formatting and validation tasks.\n- Capability inventory: Operations are restricted to local file reads and writes using standard library tools; no network requests, subprocesses, or dynamic evaluation are present.\n- Sanitization: Scripts perform exhaustive structure and type validation but do not sanitize natural language content for downstream agent consumption.

treatment-plans