Explore, Plan, Code, Commit
Anthropic teaches a simple, repeatable rhythm for working with Claude Code, and it holds up across almost any task. The phases are explore, plan, code, commit. Skipping the first two is the most common reason a session goes sideways.
Explore
Before writing anything, have the agent read the relevant code and report back what it found. You are checking that it understands the problem the same way you do. If it does not, you correct course now, when it is free.
Plan
Ask for a step-by-step plan and review it. This is where you catch a wrong approach in one comment instead of in a hundred lines of diff. Many people use a dedicated plan mode for exactly this.
Code
Now let it implement, ideally in small steps it can verify as it goes. Keep the success condition in front of it: tests green, types clean, the feature working.
Commit
Have it write a clear commit message and, if you like, open a pull request. A small, well-described commit is easy to review and easy to revert.
