The designer's AI loop

Ten sessions, one running product: Cadence, an AI meeting-notes app - the same product the Branding, Marketing and Content tracks use, so its brand already exists and you are designing inside a real constraint instead of a blank canvas. Your job across the track: design the "shared summary" feature end to end. Tonight the loop. b2 is art-direction prompting - the brief that produces usable work. b3 is research and moodboards. b4 drafts the UI. b5 is the critique lab, where the scoring you meet tonight becomes a working standard. b6 puts it into a design system with real tokens. b7 is illustration and imagery. b8 is content design and microcopy. b9 hands off to code. b10 assembles the whole thing and scores it. Every session ships something you can put in a file, not a feeling.

Every AI design task runs the same five steps, and they are not equally hard. Brief: say what you want, in the language of constraints rather than adjectives. Generate: let the machine produce more options than you would have had patience for. Critique: judge what came back against a standard you can state out loud. Direct: change the brief, not the pixels - a bad output is usually a vague constraint. Ship: make it real, in a system, with rights you can defend. AI is extraordinary at step two. It is thin at step one and step three, and those are the two steps that decide whether the work is any good.

Designers describe their judgement as intuition, and some of it genuinely is - composition, wit, restraint, whether an idea has any tension in it. But a surprising share of what we call taste is measurable, and therefore teachable to a machine and checkable by anyone: does text pass contrast, do sizes come from one scale, does spacing sit on a rhythm, is there exactly one loudest thing, how many hue families are fighting on screen. AI output usually fails on this measurable half - which is good news, because you can catch it without arguing about taste at all.

Below is Cadence's shared-summary screen as an unbriefed generator produced it: pleasant, plausible, and 17 out of 100. Nothing about that number is decoration - your browser renders the screen, then measures it. Real WCAG contrast ratios. Real computed pixel values for every padding and gap. Real count of hue families. Start with every lever off, read the failure rows, then bring them on one at a time and watch which row moves.

The same screen, scored at every rung of the ladder, computed in front of you: 17 → 32 → 46 → 66 → 86 → 100. Each rung is one lever, and each lever is a session. Then the last row: with the screen at 100, turning on Add colour costs 19 points - it introduces a fourth hue family, breaks two spacing values, and puts yellow text on a yellow highlight. The most common instinct in a design review, priced.

• Run the full ladder until you can predict which row moves before you click. Then press Add colour on a 100 and write one sentence explaining the loss to a non-designer.
• Take one screen you shipped recently and score it by hand on the five checks. Be specific: worst contrast ratio, number of distinct font sizes, number of hue families, how many elements are shouting.
• Rewrite one adjective brief you have actually sent into a constraint brief - one primary action, one type scale, one spacing unit, one accent, and a contrast floor.
• Generate the same screen twice, once from each brief, in whatever tool you use. Keep both. b2 opens with your pair.

This track teaches the working core of the current AI design toolchain and the craft standards you judge it against - W3C accessibility criteria, platform design guidance, the vendors' own documented behaviour and limits - with a live critique lab on the key pages. Certificates, videos and tool subscriptions stay with the official providers, linked each session. This page covers: