Have Claude Code Plan First, Then Approve the Plan

A workflow where you ask Claude Code to produce a written plan — the files it will change and the approach — and get your sign-off before it edits anything. You review the intent, which is cheap, rather than the diff, which is expensive to unwind.

The costly failure mode of an agentic coding tool is confidently doing the wrong thing across many files, which you then have to detect and revert. A plan surfaces a wrong direction — misread requirement, wrong file, missed constraint — while it's still one paragraph, not fifteen edited files. Redirecting a plan costs a sentence; redirecting a completed change costs a review and a rollback.

Multi-file changes, refactors, anything touching unfamiliar code, and any task where 'the wrong approach' is plausible. The bigger the blast radius, the more the plan gate pays off.

Tiny, unambiguous edits — fix this typo, add this log line — where writing and reading a plan costs more than just doing it.

Before changing any code: explore the relevant parts of the repo and give me a plan — which files you'll change, what each change does, the order, and anything you're unsure about. Don't edit yet.
I'll approve or redirect, then you execute exactly that plan and tell me if reality forces a deviation.

Asked to add rate limiting, Claude Code's plan reveals it intended to add it in the route handlers; you redirect it to the shared middleware instead — a thirty-second correction that would have been a sprawling multi-file revert if it had just started editing.

For larger work, have it keep the plan as a checklist it updates as it goes, marking steps done and noting deviations. You get a live view of progress and a diff between planned and actual you can audit at the end.

• Approving a plan you skimmed — a rubber-stamped plan is no gate at all.
• Letting the agent silently deviate from an approved plan without flagging it.
• Demanding a plan for trivial edits, adding friction where there's no risk.