The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interacts with the caveman CLI to generate reports, simulate fixes, and manage long-term memory via commands like caveman learn report, caveman learn simulate, and caveman mem remember.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data from token sink reports and project configuration files (e.g., CLAUDE.md). While this presents an attack surface for indirect injection, the skill implements robust mitigations: it enforces a mandatory SHA256 hash check (content_sha256) to verify the integrity of text blocks before they are moved to memory or edited, and it requires per-edit user approval.
[DYNAMIC_EXECUTION]: The skill includes a 'Skill Distillation' feature that can generate new agent skills based on repeated tool usage patterns. This is managed as a controlled experiment requiring user consent and longitudinal performance monitoring (via caveman learn experiment) to ensure the generated instructions improve performance without introducing errors.

caveman-learn