Taste is the job

I don’t ask one agent to build and judge the same change. The first pass gets it working. Separate skills handle PR review, simplification, copy, UI, and visual QA.

Default shadcn, Lucide, Inter, a glowing border. That’s the tell. I use the same stack. The difference is the edit after generation: delete decoration, pick a typeface, fix the icons, keep motion between 0.2 and 0.3 seconds.

I steal from calm products: Zed, OpenAI, Mintlify, Raycast, Linear, Things 3, Family, ElevenLabs. Then I point the same eye at my own screen. An agent gives me options. I still have to pick.

Taste Training is the course version of this. It has 48 lessons, with the first unit free.

Read the long version: You don't have to make AI slop