modul 5 — quality and anti-slop
design law: a system the agent can't drift from
Agents produce statistically average design, and average reads as template. A written design law of tokens and anti-patterns turns your taste into something enforceable.
~3 min read · ~10 min with the exercise · v1.1 · verified: 2026-07-15
why agent UI regresses to the mean
A model predicts the likeliest next token, so unguided design requests return the likeliest design: the purple gradient, the three-column feature grid, the rounded card with a drop shadow. None of it is wrong, and that is the problem. Likeliest means most seen, and most seen reads as template. Your taste can override it in a session; taste does not survive handoffs between sessions. A written law does.
The site you are reading keeps its law in a DESIGN.md: every color, spacing step and easing curve defined as a token, plus a numbered list of anti-patterns the agent must refuse to produce. The file ends design debates before they start, because "if it's not in the tokens, it doesn't exist."