The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a continuous learning loop that processes session transcripts—containing untrusted user input—to generate new skills. This architecture is susceptible to indirect prompt injection or poisoning. A malicious user could intentionally demonstrate behaviors or patterns in a session that, when extracted, include hidden instructions or backdoors for future sessions.\n
Ingestion points: evaluate-session.sh reads the session transcript path from the platform's hook input, which contains raw conversation history.\n
Boundary markers: Absent; the instructions do not provide delimiters or instructions for the agent to disregard potentially adversarial content within the transcript history during the evaluation phase.\n
Capability inventory: The skill facilitates the creation and execution of new files in ~/.claude/skills/learned/ based on patterns extracted from transcripts.\n
Sanitization: No automated sanitization or validation logic is present to filter malicious instructions from the agent's proposed 'learned' skills before they are saved to the filesystem.\n- [COMMAND_EXECUTION]: The evaluate-session.sh script executes local shell commands including grep, mkdir, jq, and sed to parse session metadata and manage skill storage directories.

continuous-learning