Efficient Workflow Using Claude Code: Planning Before Execution

Boris Tane describes his approach to using Claude Code for development by emphasizing the importance of separating planning from execution. This methodology revolves around a key principle: never allow Claude to write code before reviewing and approving a written plan, which helps avoid wasted efforts and maintains architectural integrity.

Core Phases of the Workflow

Phase 1: Research

Every task begins with a deep research phase where Claude is directed to understand the relevant parts of the codebase thoroughly. This understanding is documented in a research.md file to ensure the AI’s comprehension before any planning begins. This helps in identifying potential systemic issues like ignoring existing caching layers or inconsistent API implementations.

Phase 2: Planning

Post-research, a detailed implementation plan is crafted in a plan.md file. This plan includes a thorough explanation of the approach, necessary code snippets, and file modification paths. Tane highlights using specific open source code as references for features to guide Claude in generating effective plans.

The Annotation Cycle

Critical to Tane's workflow is an annotation cycle where he reviews Claude’s generated plans in his text editor, allowing for meticulous edits and annotations before Claude proceeds to code generation. This meticulous approach guards against the production of isolated implementations that could disrupt surrounding code.