writing-plans

Map out the files to create or modify first, then decompose the work into tasks of 2-5 minute steps: write a failing test, run it, implement minimal code, verify it passes, and commit. Include exact file paths and complete code in every step.

Split the work by file responsibility with clear boundaries, then write each task as checkbox steps with actual test code, implementation code, run commands, and expected output. Avoid vague steps like 'add error handling' since agents cannot act on them.

Split when the spec covers multiple independent subsystems. Each plan should produce working, testable software on its own, ideally one plan per subsystem identified during the brainstorming phase.

Common failures include placeholders like TODO or 'add validation', missing code in steps, inconsistent function names across tasks, and references to types never defined. A self-review pass against the spec catches these gaps.

Yes. After saving the plan, you choose between subagent-driven execution, which dispatches a fresh subagent per task with review between tasks, or inline execution, which runs tasks in batches with checkpoints in the current session.