Sub-agent delegation

Delegation is authorised standingly (see CLAUDE.md). This skill covers how to do it well.

Briefing a sub-agent

• Instruct them to return only key information
• Define clear boundaries per agent. Specify which files each agent owns
• Include "you are one of several agents" in instructions
• Set explicit success criteria. Combine small updates to prevent over-splitting
• Sub-agents can compete and erase each other's changes - ensure no overlap
• If the task is simple and does not require careful consideration, reasoning or creativity (for example summarising simple web searches) you may use the sonnet model

Named vs forked

Named (standard) sub-agents have their own context window - good for parallel research, inspection, or separate features. Default to a named sub-agent.

A fork inherits the main session's full conversation history, system prompt, tools, and model. Output isolation is preserved (only the final result returns) but input isolation is lost.