The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the manual configuration of a shell script (evaluate-session.sh) as a 'Stop' hook in the agent's environment, which executes every time a session ends to process transcript data.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8) because it ingests and processes untrusted session transcripts to identify 'learned skills'. * Ingestion points: evaluate-session.sh reads the session transcript file path provided by the platform via standard input. * Boundary markers: Absent; the script does not utilize delimiters or 'ignore' instructions when passing the transcript-related signal to the agent. * Capability inventory: The skill is designed to write new skill files to ~/.claude/skills/learned/ based on analyzed content, and the evaluator script performs file reads and directory creation. * Sanitization: Absent; the script uses basic text processing (grep) on the raw transcript and does not filter or escape potentially malicious instructions embedded in the user or model messages.

continuous-learning