Pull request reviews assume a human wrote the code line by line, so reviewing it line by line makes sense. That assumption is now wrong most of the time, and it is making reviews slower and shallower.
Review the cause, not the effect
When code is generated, the meaningful review surface is the prompt and the spec. A reviewer who reads "add rate limiting: 100 req/min per API key, return 429 with Retry-After" understands the change in seconds — and can catch a wrong intent that 800 lines of plausible code would hide.
What vb diff shows
vb diff renders three panes side by side: how the spec changed, how the prompts changed, and how the code changed. The code is still there — but now it is the evidence, not the entire case.
vb diff main..feature/rate-limit
spec + 100 req/min per key, 429 + Retry-After
prompts + "use a sliding window, not fixed buckets"
code ~ middleware/limiter.ts (+47 -3)
+ middleware/limiter.test.ts (+61)Faster reviews, deeper ones
Teams that review the prompt first report two things: reviews get faster, because intent is dense and code is verbose — and they get deeper, because reviewers finally argue about what the software should do instead of rubber-stamping how it does it.
“The best review comment on a generated PR is almost never about the code.”
