After nine months using Claude Code as my primary development tool, I've developed a workflow that radically differs from typical AI coding approaches. Most developers jump straight to implementation, typing prompts and fixing errors iteratively. My method centers on one principle: never let Claude write code until you've reviewed and approved a written plan.

This separation of planning and execution prevents wasted effort, keeps me in control of architecture decisions, and produces better results with minimal token usage. The workflow starts with deep research—I ask Claude to thoroughly understand relevant codebase sections and document findings in a persistent markdown file. I use explicit language like "deeply" and "in great details" to ensure comprehensive analysis rather than surface-level reading.

The annotation cycle is the workflow's most distinctive element. After Claude generates a plan, I review it in my editor and add inline notes directly into the document—correcting assumptions, rejecting approaches, or providing domain knowledge. This cycle repeats 1-6 times, with the markdown file serving as shared mutable state between me and Claude. The plan document acts as a structured specification I can review holistically, unlike trying to reconstruct decisions from chat messages. Before implementation, I always request a granular task breakdown to track progress.