ns-ui / writing
Writing.
Notes on building the registry, one at a time.
Heroku's 2014 homepage runs seven of the sixteen slop signals. Nobody prompted it.
I built a 2014 control cohort to test whether AI flattened web design. Across 25 sites the template turns out to predate the accusation by about a decade, and the cleanest result I had at n=20 did not survive finishing the corpus.
I tried to score a page's design from its HTML. Three signals, three wrong answers.
Clerk declares a 754x727 indigo-to-cyan gradient circle that samples #f7f7f8 at every point. Measured from the source, two of three signals read worse than the page is, and the third would have put the first false positive on the one signal that has never fired.
I audited my own llms.txt. Two thirds of it is still just tags.
Four days after arguing that tags cannot separate similar components, 146 of my 228 selection-guidance lines still contain no comparison at all. Here is what I measured and what a name can never do.
I made my component library readable by agents. Here's what I got wrong first.
Tags don't separate components that share a category. Writing documentation a machine can read forces a precision that documentation for people lets you skip.