The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill ingests untrusted data from user-provided documents and external clinical trial databases. This creates a surface where embedded instructions in external content could potentially influence the agent's behavior. This is an area to review, though typical for research tools. Evidence Chain (Category 8): (1) Ingestion points: Data enters the context via references/00-initialize-intervention.md (initial_context) and references/01-research-protocols.md (external search results). (2) Boundary markers: The instructions do not currently include explicit delimiters or warnings to ignore instructions found within the clinical research data. (3) Capability inventory: The skill possesses capabilities for local file writing (via cat) and local script execution (scripts/sample_size_calculator.py). (4) Sanitization: There is no evidence of specific sanitization or filtering applied to external clinical data before it is processed by the agent.
[Local Script Execution]: The skill executes a bundled Python script (scripts/sample_size_calculator.py) to perform statistical calculations. This is a controlled use of local code with numeric parameters derived from the user session.
[External Data Acquisition]: The workflow performs research by accessing authoritative government domains such as fda.gov and nih.gov. These references are documented as part of the core functionality and target established, trusted organizations.
[Command Execution]: The skill uses standard shell commands like cat for document assembly. These operations are intended for concatenating generated protocol sections into a final document within the local workspace.

clinical-trial-protocol-skill