The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides strict instructions to the agent to stay within the boundaries of the provided documents. It explicitly requires running a retrieval script before answering and mapping every claim to a specific source, which acts as a safeguard against hallucinations and outside instruction interference.
[COMMAND_EXECUTION]: The skill utilizes a local Python script scripts/search_evidence.py for document retrieval. While a subprocess.run call is present in tests/test_policy_examples.py, it is restricted to calling the local search script using the current Python interpreter for testing purposes, which is a standard and safe development practice.
[DATA_EXFILTRATION]: No network access, telemetry, or exfiltration patterns were detected. The retrieval script only reads from the local references/corpus directory and does not access sensitive system paths.
[REMOTE_CODE_EXECUTION]: The skill does not download external scripts or execute code from remote sources. All dependencies are from the Python standard library.

z-expense-policy-qa