The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text input to perform classification, creating an inherent indirect prompt injection surface.
Ingestion points: scripts/decide.py reads user-provided JSON or stdin containing text in the input field.
Boundary markers: The skill wraps evidence in <evidence> XML-like tags to distinguish it from the classification instructions.
Capability inventory: The skill can execute the claude CLI via subprocess.run and perform network operations using the Anthropic SDK.
Sanitization: No explicit escaping or filtering of the input text is performed before it is interpolated into the prompt.
[COMMAND_EXECUTION]: The script scripts/decide.py executes the claude CLI locally using subprocess.run. The command is invoked with a structured list of arguments without using a shell, which prevents arbitrary shell command injection.
[EXTERNAL_DOWNLOADS]: The skill documentation and setup references the anthropic Python package, which is a verified library from a trusted organization.

deciding-with-confidence