The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that explicitly override standard agent behavior regarding user interaction. The "Investigation Flow" section directs the agent to "Do NOT break flow by asking clarifying questions," "don't ask for permission," and "Derive information from logs, databases, or file state — don't ask the user to be your instrumentation layer." These instructions encourage the agent to perform actions without user oversight or confirmation.
[DATA_EXFILTRATION]: The skill instructs the agent to access and read potentially sensitive data from private application directories. It provides paths and commands for reading SQLite databases within macOS sandbox containers (~/Library/Containers/<bundle-id>/Data/Documents/), which store user-private data for system applications. It also suggests using log show and log stream to extract detailed system telemetry and logs.
[COMMAND_EXECUTION]: The skill provides instructions for sensitive and destructive system commands. Specifically, it guides the agent to use tccutil reset to modify system permissions and rm -rf to delete application containers (~/Library/Containers/<bundle-id>/). These operations can lead to significant data loss or changes in the system's security posture if executed without careful review.

systematic-debugging