How I Use Claude Code

A detailed personal workflow for using Claude Code that centers on a strict research-plan-implement pipeline. The core principle: never let Claude write code until a written plan has been reviewed and approved. The workflow has three phases — deep codebase research written to a markdown file, iterative plan annotation where the developer adds inline corrections and constraints, and a single implementation command that runs without interruption. Key practices include using persistent markdown artifacts instead of chat, an annotation cycle of 1–6 rounds to inject domain knowledge, a granular todo checklist for progress tracking, and terse corrections during implementation. The author argues this approach prevents the most common AI coding failure mode: implementations that work in isolation but break the surrounding system.