Writer / Reviewer split
Get reviewer eyes that aren't biased toward the author.
Official basis Fresh context improves review — Claude won't defend code it just wrote. docs ↗
The prompt
You didn't write this code. Review <file> for edge cases, race conditions, and consistency with our existing patterns. Be skeptical and assume there's at least one problem. Why it works
Running review in a separate session (or telling the model it isn’t the author) recreates the value of a second engineer. “Assume there’s a problem” pushes it past a lazy “looks good to me.”
Fill in
<file>— the code under review.
Follow-ups
- “Rank your findings by likelihood of causing a real incident.”