How I use Claude Code: Separation of planning and execution

The author, Boris Tane, introduces a highly opinionated workflow for using Claude Code that deviates significantly from typical AI coding practices. His core principle is to never let the AI write code until a detailed plan has been reviewed and approved by a human. This approach, refined over nine months, is designed to prevent wasted effort, maintain architectural control, and achieve better results with minimal token usage by separating the 'thinking' (planning) from the 'typing' (execution).

• Phase 1: Research: Tasks begin with a directive for Claude to deeply understand a part of the codebase, requiring its findings to be written into a persistent research.md file. Keywords like "deeply" are crucial to prevent surface-level analysis. This artifact serves as the human's review surface to catch misunderstandings early.
• Phase 2: Planning: After research, Claude generates a plan.md outlining the implementation, including code snippets, file paths, and trade-offs. The author prefers custom markdown plans over Claude's built-in plan mode for greater control and persistence.
• The Annotation Cycle: This is the workflow's most distinctive part. The author reviews the plan.md in an editor, adding inline notes to correct assumptions, reject approaches, add constraints, or provide missing domain knowledge. Claude then updates the plan based on these notes, with the cycle repeating 1-6 times, explicitly guarded by "don't implement yet" until the plan is perfect.
• The Todo List: Before implementation, a detailed todo list is added to the plan, serving as a progress tracker for Claude to mark off tasks.
• Phase 3: Implementation: A standardized prompt initiates implementation, instructing Claude to execute the plan completely, mark progress, avoid unnecessary comments/types, and continuously typecheck. The goal is to make implementation a mechanical, "boring" process, as all creative decisions are finalized in the planning phase.
• Feedback During Implementation: The human's role shifts to supervision, with terse corrections for any deviations. For frontend issues, screenshots and references to existing UI patterns are used. If an approach goes wrong, the author reverts changes and re-scopes rather than incrementally fixing.
• Staying in the Driver's Seat: The author emphasizes maintaining control by cherry-picking from Claude's proposals, trimming scope, protecting existing interfaces, and overriding technical choices. The human makes the judgment calls; Claude handles mechanical execution.
• Single Long Sessions: The entire process (research, planning, implementation) occurs in one continuous session. Claude's context window management and the persistent plan document allow for sustained work without degradation.

This disciplined pipeline, devoid of "magic prompts," ensures that the human's judgment, product priorities, and architectural insights are injected at the right stages, leading to a more controlled and effective AI-assisted development experience where the AI executes a human-approved vision.