Lee Sun-woo

The layout passed. The page was still wrong.

On November 7, 2026, this site’s article list passed a horizontal-overflow check and was still visibly broken. In a 390px viewport, twelve articles had landed in the 88px column intended for their section label. The adjacent content column was 246px wide.

The check was accurate: nothing extended beyond the viewport. It simply answered a narrower question than whether the list was laid out correctly. That distinction is a useful starting point for an AI design critique.

Describe the relationship that failed

Adding an introductory note changed the order in which the grid placed its children. The note occupied the content column, and the following list was automatically placed in the label column on the next row. The correction explicitly assigned the list to the content column.

“The mobile layout needs polish” would have left too many possible responses. A useful finding identified the list, the incorrect column and the consequence: titles and descriptions wrapped into an unnecessarily narrow strip.

That description also set the limit of the fix. It did not call for a smaller font, shorter titles or a new archive layout. The intended relationship between the section label and its content already existed in the neighboring groups.

Reconstructed layout. The list belongs in the content column, even when both placements fit the screen.

Make evidence match the claim

The follow-up check compared the lists’ left edges and widths across the page, including the group with an introductory note. At 320, 390, 768 and 1280px, the lists shared the intended content column. Horizontal overflow remained a useful separate check.

A critique can use the same discipline at different levels:

  • For alignment, name the elements that should share an edge.
  • For hierarchy, point to the competing elements and the intended priority.
  • For task completion, describe the action and inspect its result.
  • For an assumed user benefit, state what still needs to be tested with people.

These are different claims. A screenshot can reveal the narrow column; it cannot establish that a revised archive helps people find an article faster.

Use a rubric to direct attention

In its work on long-running application development, Anthropic reports overly generous self-evaluation and separates the generator from an evaluator with explicit design criteria. The authors also note that a separate evaluator can remain lenient.

My application is to ask for a finding I can accept, reject or test: a location, an observed problem, its likely consequence and a bounded correction. A rating such as “8/10 for craft” gives me much less to act on.

The archive failure changed the relevant review question. “Does the page fit?” remained necessary. “Do these lists occupy the same intended column?” was what would have caught this particular mistake.