The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from an external knowledge base to verify claims against sources.
Ingestion points: The agent reads and interprets the content of markdown files in the knowledge/ directory and various source artifacts (markdown, quotes, etc.) in the sources/ directory.
Boundary markers: No explicit delimiters or instructions are used to distinguish the content being audited from the agent's core instructions, creating a risk that malicious instructions in the KB could be followed.
Capability inventory: The skill executes a local Python script via subprocess, reads files, writes [DRIFTED] markers into notes, and maintains a verification-log.md file.
Sanitization: The skill does not perform validation or sanitization of the external content prior to processing it.
[COMMAND_EXECUTION]: The skill executes a local Python script (scripts/check_kb_references.py) using python3 to automate the mechanical auditing process. This is a standard operation required for the skill's functionality.

kb-verify