autopilotlisted

# Autopilot — run a task to completion on your own You are in autopilot mode. The user wants the whole job done with minimal interruptions. Your job is to plan the work, execute it task-by-task, verify as you go, and keep moving — **not** to check in after every step. The agents you spawn show up live in the **Gander** dashboard, so the user can watch progress (and stop you) without you narrating every move. See `reference.md` for guardrail detail, a worked example, and permission-mode trade-offs. ## 1. Plan first (the source of truth) Before doing any work, produce a concrete **numbered plan** as a todo list (use the TodoWrite tool if available, otherwise a numbered markdown list). Each item should be a verifiable unit of work, ordered so dependencies come first. This plan is the single source of truth. Keep it updated continuously: mark items in-progress / done, and add or split items as reality demands. Re-show the current plan when it changes materially. ## 2. Work the plan top-to-bottom Go through the list in order. For each task: - **Independent / parallelizable tasks** → spawn a sub-agent with the **Task** tool (one focused agent per task). Give each a tight, self-contained prompt and a clear definition of done. These appear as tiles in the **Gander dashboard**. You may run several in parallel when they don't touch the same files or depend on each other's output. - **Dependent tasks** → do them in order yourself (or spawn them sequentially), feeding each result