blog
A verdict on rendered pixels, not a vibe
Design review that cannot be measured is opinion. The review step measures the rendered page — contrast math, layout logic, known tells — and returns PASS or the exact fix list, with the builder never grading its own work.
Why the builder cannot be the reviewer
Ask an agent to check its own page and it will tell you the page is good. This is not dishonesty; it is the same bias a person has about their own work. It knows what it meant to do, so it reads the intent instead of the output.
The fix is structural. The thing that grades the page is not the thing that built it, and it does not accept an argument. It starts from the page as rendered — the pixels and the DOM, not the plan — and it reports what it finds.
What gets measured
Contrast first, because it is arithmetic. Every text and background pair on the page has a real ratio, and a ratio either clears the floor or it does not. No amount of taste turns a failing ratio into a passing one.
Then layout logic: whether the grid the page claims to use is the grid it actually renders, whether tap targets are big enough to hit, whether the type scale is the scale in the direction or a set of one-off sizes that drifted in during the build.
Then the tells — the patterns that mark a page as generated rather than designed. They are named, not vague. A named tell can be fixed. "This feels generic" cannot.
PASS, REVISE, REJECT
Three verdicts, and each one means something specific. PASS means the page clears the floors and the direction it committed to. REVISE means the structure is sound and there is a list of concrete fixes — the failing pair, the value it needs, the element responsible. REJECT means the page did not ship against its own direction, and patching it would be dishonest; it gets rebuilt.
The value of a verdict is that it can be acted on without a conversation. A fix list that names the element and the required value is a work order.
The tells we name
Gradient hero behind a centered headline. Glass panels over nothing. Three identical cards with icon, title, and two lines of padding. Purple as a substitute for a decision. Motion added because a library was available rather than because something moved. Each of these is a specific pattern, which means each is a specific edit.
The library exists for the other half of this: the cards carry what to steal, so the fix is not only subtraction. You remove the tell and replace it with a named move from work that earned it.
What this does not do
It does not decide what your product is, and it does not replace your taste. It grades the parts of a page that have a right answer — contrast, hit areas, scale discipline, the presence of known filler — and leaves the parts that do not have a right answer to you.
The review reads the evidence your agent collects; it does not look at screenshots or operate the page itself, so a PASS covers the evidence supplied, not every interaction.
That is the whole discipline: measure what can be measured, name what cannot, and never let the builder mark its own exam.