Begin by writing a PLAN.md breaking down the task into numbered atomic steps (1-2 sentences) with the goal, checklist, and instructions.
You don't need to complete all steps at one shot. Sub-agents can start from where you left off.
After each step, update PLAN.md with actions & outcomes - context for another agent to continue.
Use sub-agents to execute the work to save context. Delegate planning to better models (new, more thinking).