Atomic Planning

File decomposition first

Before defining tasks, map out ALL files:

• What files will be created?
• What files will be modified?
• What is each file's single responsibility?
• What are the interfaces between files?

Lock in decomposition decisions here. Units should have clear boundaries.

Task sizing

Each task: 2-5 minutes of AI work. If it cannot be described in 2-3 sentences, split it.

Right-sized examples:

• Add a database column and migration
• Add a single UI component to a page
• Update a server action with new logic
• Add a filter dropdown

Too big (must split):

• "Build the dashboard" → schema + queries + UI components + filters
• "Add authentication" → schema + middleware + login UI + session handling

XML plan format

<task type="auto" model="sonnet">
  <n>Task title</n>
  <files>file1.ts, file2.ts</files>
  <skill>skill-name</skill>
  <test-first>Write test: [description]. Watch fail (RED).</test-first>
  <action>[Implementation]. Watch pass (GREEN).</action>
  <verify>Run tests. All pass. Typecheck passes.</verify>
  <done>Completion criteria.</done>
</task>

Model routing

• haiku: mechanical 1-2 file tasks
• sonnet: standard multi-file integration (default)
• opus: architecture decisions, complex refactors, review

Wave grouping

• Wave 1: all independent tasks (run in parallel)
• Wave 2: tasks that depend on Wave 1 results
• Wave N: tasks that depend on Wave N-1

Never put dependent tasks in the same wave.