The Agent Skills Directory

[COMMAND_EXECUTION]: The skill configures a local bash script (evaluate-session.sh) to run automatically as a 'Stop' hook whenever a session ends. This script is used to count messages and signal the agent to begin pattern extraction.
[DATA_EXFILTRATION]: The skill reads local session transcript files to identify reusable patterns. This data is handled within the agent's local environment and is not exfiltrated to external domains.
[PROMPT_INJECTION]: The skill processes session transcripts which contain untrusted user input. This represents an indirect prompt injection surface where a user could potentially 'teach' the agent malicious patterns. However, the risk is mitigated by the agent's internal safety guardrails and the fact that extraction is intended for the user's own future sessions.
Ingestion points: The transcript_path variable in evaluate-session.sh points to the session history file.
Boundary markers: None identified in the provided scripts.
Capability inventory: The agent uses its internal file-writing capabilities to save learned skills to the ~/.claude/skills/learned/ directory.
Sanitization: No explicit sanitization of transcript content is performed by the shell script; it relies on the agent's interpretation of the data.

continuous-learning