18 October 2026 · Frontend Design evidence check
A design skill can leave a model default intact.
Decision: If you dislike a specific pattern, name it in the brief and inspect the rendered page. Do not assume that “avoid generic design” or a skill will remove it.
A contributor reported that 10 of 21 Claude Opus 5.5 pages generated with the current Anthropic Frontend Design text had pill-shaped buttons. The test used seven briefs without visual direction, three runs each, medium effort, and a computed-style rule: button radius at least half its height. We did not rerun it or inspect the linked output files.
This is a with-skill result, not a skill-versus-no-skill result. The issue does not name a source commit, so it cannot be assigned to our pinned listing. A pill button can also be right for a product. The count identifies a repeated shape under those conditions, not a quality failure rate.
What to do on your project
| Before the run | After the run |
|---|---|
| Give the agent your route, content, design tokens or visual reference, and any shape you want to avoid. | Inspect the actual desktop and mobile render, including focus and error states relevant to the task. |
| Keep the model, task brief, project state, and viewport fixed for a no-skill baseline and a skill trial. | Compare both outputs for system fit, useful visual difference, functional checks, time, and token cost. |
Anthropic's Opus 5.5 guide says broad anti-generic wording can replace one default style with another and recommends naming concrete patterns, then checking the first result. Its skill-creator method uses a paired baseline and treats subjective design quality as something to review, not an automatic pass/fail assertion. Neither source proves a listed skill improves your screen.
Source status: one contributor report, two Anthropic documents, no pool-run visual outcome comparison. The method explains how the catalogue separates source inspection from results.