Have Claude Review the Diff, Not the Whole File

A review pattern where you give Claude the diff — the lines that actually changed — plus enough surrounding context to judge them, and ask it to review only the change, not to critique the whole file.

Hand Claude a whole file and it reviews everything, including untouched code, burying the findings that matter under stylistic nitpicks about lines nobody changed. A diff scopes attention to the decision under review. Adding just enough context — the function's signature, the callers — lets Claude judge whether the change is correct in situ rather than in the abstract.

Reviewing a pull request or a commit before you push. Any time the unit of interest is 'what changed', not 'is this whole file good'.

When you actually want a holistic audit of a file or module — then the whole file is the right input. And a diff with too little context can hide a break in the unchanged code that calls it; include the neighbours.

Review this change. Here is the diff:
<paste diff>

For context, here are the relevant surrounding pieces (signatures, callers): <paste>.

Review only what changed. Focus on correctness, edge cases the change introduces, and whether it breaks any caller shown. Ignore pre-existing issues in unchanged lines unless the change makes them dangerous. List findings most-serious first.

You paste a 20-line diff plus the two functions that call the changed one. Claude spots that the new early-return skips a cleanup step the callers depend on — a real bug — instead of, as it would with the whole file, complaining about a variable name three functions away.

Feed Claude the diff together with the PR description or the ticket, and ask it to check the change actually does what the description claims — no more, no less. Scope creep and missing pieces are easiest to catch when the stated intent sits next to the real diff.

• Pasting the diff with zero context, so Claude can't tell whether a change breaks its callers.
• Letting it drift into reviewing untouched code, reintroducing the noise you were trying to avoid.
• Reviewing a huge, multi-concern diff in one pass — split it first so each review has a single subject.