Redwood Pizzeria · design R&D · run-019

Steering AI image generation toward expressive, custom design

First-attempt brand toolkits and page mocks tend to come out safe and AI-generic. This run worked out why — and the prompt-language levers that reliably produce bold, custom, buildable brand design instead. Two pieces:

Part 1 · the system

The Prompt Taxonomy

38 axes and 204 values: the language elements of an image prompt that steer the aesthetic (which world) separately from the boldness (how far from safe) — with measured findings on which levers actually work.

Open the taxonomy →
Part 2 · seeing it

The Style Survey

A visual canvas: one brand rendered across 18 mediums × 43 worlds × 5 postures, sweeping each axis while holding the others fixed. Click any kit for its full generation prompt.

Open the survey →

Also in this run: the results write-up — why first attempts come out safe, and the validated fix (an A/B test beat the current process 3–0).