The Agent Skills Directory

Local Script Execution: The skill instructions specify the execution of a Python script (scan_sensitive_content.py) from the adr-author skill. This script serves as a security control, scanning content for sensitive information such as tokens or connection strings before it is committed to a repository.
Indirect Prompt Injection Surface: The skill processes external data during state reconstruction and artifact validation. It addresses this by incorporating defensive instructions: (1) Ingestion points include scanned artifacts in references/flow-state.md and inventory files in references/session-state.md; (2) Boundary markers are present, instructing the agent to treat content as data rather than instructions; (3) Capabilities include script execution and file writing; (4) Sanitization is performed by the sensitive content scanner.
State Management: The skill reads and writes session data to a hidden directory (.copilot-tracking) to maintain continuity. This is a standard mechanism for providing a consistent coaching experience across project sessions.

data-science-engineering-foundation