The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for using the agent-tmux CLI tool to assign tasks to workers (e.g., agent-tmux claude assign {safe-name} {repo-dir} {prompt-file}).
The instructions explicitly direct the agent to write task content to a prompt file rather than interpolating it directly into the shell command, which is a key security measure against command injection.
[INDIRECT_PROMPT_INJECTION]: The templates are designed to ingest and process potentially untrusted data from local repositories or external documentation (in the SEARCH, IMPLEMENT, and RESEARCH templates).
Ingestion points: The templates take {repo/dir}, {files/module}, and {official docs/repo/spec} as input for subagents.
Boundary markers: The skill uses explicit headers (GOAL:, SCOPE:, ACCEPTANCE:, REPORT:) to define clear task boundaries for subagents, reducing the risk of subagents following instructions embedded within the data they process.
Capability inventory: Subagents are granted the capability to search, read, and modify files within a specified scope via the agent-tmux environment.
Sanitization: The skill mitigates risks by requiring the dispatcher to resolve paths and write prompt files before execution, ensuring the subagent receives a static, validated instruction set.

delegation-templates