The Agent Skills Directory

[COMMAND_EXECUTION]: Employs benign shell commands like pwd and git rev-parse to ensure the agent is operating in the correct project context.
[DATA_EXFILTRATION]: Instructs the agent to list necessary credentials and third-party dependencies during the planning phase. This is used for requirement gathering and specifically prohibits requesting actual secrets during the implementation phase.
[PROMPT_INJECTION]: Contains instructions to analyze external sources like GitHub issues and project documentation. While these represent an indirect injection surface, the skill's structure requires human approval of any proposed plan, mitigating the risk of automated obedience to embedded instructions.

think