The plan is your checkpoint
Before executing a nontrivial task, Claude Code can lay out the steps it intends to take — which files it will touch, what approach it will use, what commands it expects to run — so you can catch a wrong assumption before any code changes. This is the cheapest point in the whole session to correct course: rejecting or adjusting a plan costs nothing, while rejecting a finished set of edits after the fact means the agent has to undo work and redo it. Reading the plan is where you verify the agent understood the actual problem, not just words that sounded like the problem.
A plan is also where you notice scope creep before it happens. If you asked for a bug fix and the plan includes refactoring three unrelated modules "while I'm in there," that's the moment to say no, not after the diff lands. Treating the plan step as optional and skipping straight to execution throws away the cheapest opportunity you have to steer.