UI Worst-Case Data: test results

Verdicts

Date

2026-10-29

Strong · claude-sonnet-5-5 (Claude Code alias "sonnet")

Right on all 3 cases, read by hand. On a listing row with an enum and a required callback it named the callback in its own list, kept the price a whole number from zero up, used all three status values and no other, and wrote the six awkward titles and the lists of none, one and a hundred. On a member badge with optional props it kept the score between 0 and 100, put the joining date at the far past, the far future, the offset +14:00 and a minute before midnight, and left out the score and the badges in their own fixtures. On a component with one boolean prop it wrote exactly two fixtures.

Weak · claude-haiku-5-5 (Claude Code alias "haiku")

Right on all 3 cases, read by hand. It produced the same fixture set as Sonnet in every case: the callback only in its list, no negative or decimal price, no score outside 0 to 100, the date edges and the absent optional props, and two fixtures for the single boolean. One detail worth noting: both models reused the large number 4000000000 from the example in the skill, which is harmless because it is a valid large value.

Note

Light test: three cases (two planted, one clean), one run per model; no comparison without the skill. The two component types were written by us; one has an enum, a required callback and a prop documented as a whole number from 0 up, the other has a bounded score and optional props, and the control has a single boolean prop. The checks validate every fixture against the declared type and look for each worst value; they cannot see that only one prop changes per fixture, so that rule was read by hand. Facts come from public MDN, Unicode and IANA pages read on 29 October 2026; there are no owner-measured facts.

What was not measured

• Models other than the two named above were not run.
• Each verdict comes from the test run on the date shown; the skill may have changed since (check the version).
• Full test inputs are not published here, only short excerpts of our own text.
• Results on your own texts, languages and domains can differ.