subagent-prompt-designlisted

# SubAgent Prompt Design Companion to `iterative-retrieval`. That skill solves _what context to gather_; this skill solves _how to commission the agent that does the work_. --- ## 1. The Core Problem: Subagents Are Born Amnesiac A SubAgent starts with **zero conversation history**. The only thing it knows is what you wrote in its prompt. The only thing that returns to you is its **final message** — everything intermediate is discarded. Consequences: - Anything you discussed with the user but did not include in the prompt **does not exist** for the SubAgent. - Any file content the SubAgent read but did not include in its return **is lost**. - Any reasoning the SubAgent did but expressed only mid-conversation **is silent**. Design every prompt as if you are handing a task to a contractor you have never met, via written brief, with no follow-up possible. --- ## 2. Anatomy of a Good SubAgent Prompt A complete prompt has five parts. Include all of them. ### 2.1 Task (one sentence) State the single concrete outcome. Use a verb + object + constraint. ``` ❌ "Look into the auth code." ✅ "Identify all call sites of `verifyToken()` in /Users/me/proj/src and return the file path and line number of each one." ``` The one-sentence limit forces you to split bloated tasks before you write the prompt, not after the agent wastes cycles. ### 2.2 Context (absolute paths + decided constraints) Include every piece of information the agent needs that is not on disk: - **Abs