Reviewing AI-generated frontend PRs: a reviewer's guide
What changes when the author of a frontend PR is Claude Code or Cursor instead of a teammate, and why the review process you already trust wasn't built for it.
Articles
Musings and measurements
from UXSense.
How to test Claude Code frontend changes before you merge
Claude Code will happily ship a frontend change that passes every test and still breaks the thing your users actually do. Here's a practical checklist for catching that before merge.
We gave a coding agent a real redesign task. Tests passed. Review passed. Visual regression passed. Here's what none of them caught.
Every check in a frontend pipeline validates the diff against the ticket. None of them asks whether the change still works for someone who already knew how to use the product — so we ran a real agent-authored redesign past all of them and published what they missed.
Visual regression vs. behavioral regression: what's the difference?
Visual regression testing catches pixels that changed. It says nothing about the pixels that stayed the same while what people do with them changed underneath.
What tests don't catch when AI rewrites your interface
Tests passing and code review passing don't mean your interface still works the way people expect. Here's the category of failure that slips through both.
An AI that watches your sessions is still guessing
A model that watched ten thousand sessions has ten thousand opinions. Why watching isn't measuring, prompts are specs that rot, and a finding that's never graded stays exactly as trustworthy as the day it was written.
Every PR check should have to find out if it was right
Almost everything anyone builds has a trigger and an action, and no proof, no memory, and no stop condition. Why the three missing parts are the whole failure, and what we built instead.