The Agent Skills Directory

[COMMAND_EXECUTION]: Automates the bypass of interactive safety prompts in the claude-zai tool. The skill launches the tool with the --dangerously-skip-permissions flag and then uses tmux send-keys to programmatically select the 'accept' option (sending '2') in the resulting security confirmation menu, effectively neutralizing the tool's built-in safety checks.- [COMMAND_EXECUTION]: Uses tmux to create windows and send arbitrary keyboard input to terminal panes. This provides a mechanism for the skill to control and execute commands within a sub-agent's shell environment via terminal automation.- [PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection (Category 8). * Ingestion points: Data from task descriptions and plan files stored in ~/.claude/plans/ are interpolated into the teammate's initial briefing. * Boundary markers: No delimiters or specific 'ignore embedded instructions' warnings are used when passing untrusted data to the sub-agent. * Capability inventory: The skill can launch external CLIs, create tmux windows, and send keystrokes to active terminal sessions. * Sanitization: No sanitization or escaping of dynamic variables (such as <taskslug> or TASK) is performed before they are sent via tmux send-keys, which could lead to command or instruction injection.

claude-zai-teammate