Can you teach an AI taste?

Hypothesis

AI-generated design converges on a recognisable slop: gradient heroes, glassy cards, purple everything. Taste — a coherent, personal aesthetic — lives in choices, and choices can be written down. A bundle of anti-slop design skills plus an explicit personal style profile should make an AI design in your taste, not the internet's average.

What we built

A design-taste system for our agents: a set of anti-slop skills (each one encoding a specific discipline — typography restraint, motion rules, layout judgment, what never to do) routed by a coordinator, plus a written style profile of the aesthetic we actually want — earth tones, Japanese avant-garde restraint, texture over gloss. Agents doing UI work consult the router; the profile overrides model defaults.

Learnings

• Negative space is teachable: the most effective skills are lists of what NOT to do — forbidden gradients, banned shadows, dead phrases.
• A style profile written from real references beats adjectives. "Like these three examples" outperforms "minimal and elegant" every time.
• Taste routing matters more than taste volume — most design tasks need one relevant discipline loaded, not thirteen.

Log

• 2026-07-03 — Router + style profile in use for new UI work; unused skills parked to keep the loadout lean.
• 2026-06-22 — Anti-slop bundle assembled and adapted to our stack.