Review a 50-file AI change by what it does for users. Map every place it touched, look at every changed screen before and after on one page, and decide only the calls that need you. The free AI starter kit we build our own products with makes that review page for you.
Why reading the diff stops working
AI writes code faster than you can read it, and every file looks reasonable on its own. The problems hide in the places it touched that you didn't ask about, like a shared helper that three other screens use, or a rule it copied instead of reusing.
So review from the outside in. Start with where the change reached and what your users will see, then read code only where something looks wrong.
Map every place the change touches
List your apps first: web, mobile, desktop, the website and the docs. Search each one for the feature's words, using the names the code uses as well as your own. Then open the hits, drop the false ones, and add places that draw the feature without naming it.
Order the list the way the change should land: shared code first, then each app, then the website and docs. Check open pull requests that touch the same files, because two changes in one place is where merges go wrong.
See every screen before and after
Capture every screen the change affects, before and after, and put them on one page. Pin notes on the exact spot that looks wrong, the way you'd comment in Figma. You'll spot a broken empty state or a shifted button in seconds, which no diff shows you.
Check the change against four rules
You don't need to read every line to check these:
- Each fact has one owner. If two places both decide the same thing, they'll drift.
- A rule true in two apps lives in shared code, written once and imported.
- One file, one job. A file that serves many features breaks many features.
- Imports point one way. Shared code never imports an app.
Any break is where you read the code.
Decide only the calls that need you
Most of a review is the AI applying rules you already set. Product calls are yours: anything that removes something users rely on, changes pricing or changes what a paid plan includes. Answer those once, write them down as rules, and the next review asks you less.
What to look for, and what to do
| What you see | What it usually means | What to do |
|---|---|---|
| A screen you didn't ask about changed | A shared component was edited | Check every screen that uses it |
| The same logic in two files | A copied rule | Move it to one owner |
| The change works on web but not mobile | It reached one app and missed another | Add the missing app to the map |
| Two open PRs touch the same file | A merge waiting to break | Land one first, then rebase the other |
| A feature quietly disappeared | Something removed to make a screen cleaner | Put it back, then decide |
Get the review set we use
The starter kit's review set maps every place a change touches, shows every screen before and after on one page, and asks you only the calls that need you. It works with Claude Code and Codex. Get the starter kit free on GitHub.
Questions
Can I review an AI pull request without reading the code?
Yes, for most of it. Check what changed for your users screen by screen, check where the change reached across your apps, and read code only where a rule is broken or a product call is needed.
How do I know which files an AI change really affected?
Search every app for the words the code uses for the feature, its names and types as well as yours. Then open the hits and add places that draw the feature without naming it, like a screen, a test or a store listing.
What should I look for first in a large AI change?
Anything decided in two places. When two files both work out the same thing, they drift apart, and that's where the next bug comes from.
Originally published at wednesday.is.
Top comments (0)