The mental : Claude is a brilliant junior colleague. Capable, fast, mostly correct. But brilliant junior still means review-before-merge.
Every time Claude proposes a change, it shows you the diff before applying. Read it. Even briefly. Even when you trust the result.
The two failure modes the diff catches:
- Wrong scope — Claude edited 14 files when you wanted 1. Easier to catch in the diff than to undo later.
- Subtle wrong fix — Claude “fixed” a bug by suppressing the symptom instead of finding the cause. Often invisible if you don’t read the code.
Trust grows over time as Claude learns your (via CLAUDE.md and ). It never grows to “auto-accept everything”. The diff is cheap; the cleanup isn’t.
Get a second opinion that isn’t compromised
Reading the diff yourself catches scope and obvious wrongness. For anything substantial, add a machine reviewer — but not the that wrote the code, which will pass its own work almost every time. Open a fresh session and hand it the diff cold. See the review-in-a-fresh-session tip.
has this built in as /code-review, with effort levels you can turn up when the change deserves it.