A Workflow That Actually Works with Claude Code

Boris Tane, an engineering lead at Cloudflare who previously founded Baselime (acquired by Cloudflare in 2024), has published a detailed breakdown of how he uses Claude Code effectively. His post quickly hit the front page of Hacker News with nearly 500 upvotes.

The Core Principle: Never Code Without a Plan

The foundation of Tane's approach is simple but powerful: never let Claude write code until you've approved a written plan. He enforces a strict separation between two modes of interaction.

Research & planning mode: Claude analyzes the codebase and proposes an approach — but touches no code
Implementation mode: Claude writes code only after the plan has been reviewed and approved by the human

The Three-Phase Workflow

The workflow consists of three distinct phases. In the research phase, Claude reads and understands the existing codebase. In the planning phase, Claude produces a written plan that the developer reviews and refines. Finally, in the implementation phase, Claude executes the approved plan.

This approach prevents the common failure mode of AI tools: generating large amounts of code in the wrong direction. Catching a bad plan before any code is written is far cheaper than debugging 500 lines of misguided implementation.

Why Separation Matters

Many developers find AI coding tools frustrating because of the cost of course-correction. Code gets generated quickly, but if the architectural direction is wrong, fixing it takes longer than writing it from scratch. By front-loading human judgment into the planning stage, Tane's method keeps the human in control of decisions that matter most.

The post resonated widely in the developer community, reflecting a maturing understanding of how to work effectively with AI coding assistants.

#workflow

How I Use Claude Code: Separating Planning from Execution