The Agent Skills Directory

[Prompt Injection] (HIGH): The skill implements a feedback loop that processes untrusted session data into permanent agent instructions.
- Ingestion points: Reads session history via the CLAUDE_TRANSCRIPT_PATH environment variable in evaluate-session.sh.
- Boundary markers: None; there are no delimiters or logic to distinguish between user intent and data content.
- Capability inventory: The skill is designed to write new executable or instructional files to the local filesystem (e.g., ~/.claude/skills/learned/).
- Sanitization: Absent; the script triggers the 'learning' process based solely on the number of messages in a session.
[Command Execution] (MEDIUM): The evaluate-session.sh script executes shell commands and filesystem operations (mkdir) using paths derived from config.json, which could be used for directory traversal or path manipulation if the configuration is altered.

continuous-learning