20 October 2026 · Frontend Design revision check
A 75% screenshot win rate does not settle adoption.
Decision: Keep the current pinned listing. Do not treat a proposed skill rewrite as proven better than a no-skill brief on your project.
In an open Anthropic skills PR, a contributor reports that a proposed Frontend Design rewrite beat an older original in 21 of 28 decisive comparisons. The author says 30 prompts were sampled from a bank of 50 across Claude 4.5 Haiku, Sonnet, and Opus. A blind Claude Opus 4.5 judge compared full-page screenshots for prompt fit, aesthetics, polish, usability, and distinction.
What the number means
| It tests | It does not test |
|---|---|
| One proposed instruction revision against an older original, under the author's screenshot-judging setup. | Current pinned instructions against no skill, behavior on a live route, keyboard access, or a user's preference. |
The PR is not merged. We did not inspect the generated pages or rerun its evaluation. The result is useful evidence that instruction wording may matter in that setup, not an install recommendation for every model and project.
For a project choice, record the exact skill bytes that loaded and the same no-skill brief. Check components, tokens, required states, and mobile layout, then test named interactions and error recovery. Keep visual and functional results separate; include time and token cost. See the two-check method.
Evidence status: contributor-reported A/B evaluation; no pool-run skill outcome test.