Same prompt. Same model. One had the skill loaded.

Seven pairs, each built in a single pass from one casual line of prompt. The only difference between the two versions of a pair is whether the agent read the skill before writing code. Both are live pages: tab through them, hover things, and turn on reduced motion if you want to check that path too.

Pairs
7
Findings, without
94
Findings, with
0

Counts come from ui-craft-detect. Regenerate them withpnpm scene:scores, so they are never typed by hand.

How the pairs were made

One casual prompt each, the kind a real user types, answered in a single pass as a self-contained HTML file with no CDN and nothing loading off the network. Every prompt ends in “light theme” so both sides play on the same field; without it the baseline hides its defaults behind a dark coat of paint.

The full methodology, including what each pass changed and why, is inPROMPTS.md.