
AI was to produce on-brand designs from a design system, but the system was written for people, not models.
I designed the logic layer: a traceable path from brief to design, slot rules, component specs an AI can read, and a tool that turns feedback into rules.
Layouts governed by rules with reasons, not fixed templates, and 22 landing page templates approved.
The goal was in-product messages, like small notifications and landing pages, created with AI. They had to be on-brand, compliant and consistent, across several brands.
A design system organised for people isn't enough for that. The AI needed a logic layer: rules for which components are valid where, metadata it could read, and a way to keep both up to date.
Fixed templates break as soon as a message needs something slightly different. Then the AI guesses, and guesses look plausible: a font that never loaded, an icon redrawn rather than fetched, a warning bar on a message with nothing to warn about.
I already knew the design system, and the people behind it, from an earlier review. This time the brief was to build, not just recommend.
An AI can only be as consistent as the rules it's given. So I wrote the rules down, and made it show its working.
The signals found, the intent chosen, which template and why, how each slot was filled, and which rules it passed.
One switch highlights everything the AI approximated, like a placeholder logo or a guessed spacing value, each with how to fix it.
Working HTML and React from the same trace, so the decisions and the design never drift apart.
In one round the AI lost the typeface and drew its own icons. It had declared the font but never loaded it, and assumed the icon couldn't be downloaded without trying. I sent it back: “you've lost the typeface and also not tried on the icons”.
I had the same brief built three ways, each scored on 7 criteria: rule compliance, template fit, reading the brief, component choice, copy, how many assumptions it made, and the code. Switch between them below.
Includes only what the rules require. A minimal, conservative reading of the brief.
Rule compliance and safety. Nothing appears that shouldn't.
Can look sparse. It left out a bar the template showed.
Copies the template's structure as closely as possible.
Looks complete, and matches what designers expect to see.
Copied a warning bar onto a message with nothing to warn about.
Chooses each component by its documented purpose, and notes any doubt.
Purposeful choices, with uncertainty written down, not hidden.
Only as good as the component guidance it reads. Which is why I wrote better guidance.
Each spec covers anatomy, variants, states, tokens and usage, in plain Markdown that people can review and models can read.
Each spec starts with metadata the tool can act on: its category, its intent, which parents it's allowed in, its sizes, and how many can appear in one section.
I broke in-product landing pages into 6 slots. The canvas, background, app bar and padding never change. Inside them, each slot has valid options, and rules about what it can sit beside. Every rule says why, so people can challenge it. Build a layout below, and watch the rules judge it.
Valid layout
Simplified from the real landing page slot rules.
I designed and built Slots Studio with AI. It reads the slot spec and component specs, generates every valid layout, checks each against the rules, and scores them.
A keyboard-driven review queue (Y, N or M) makes it quick to get through hundreds. Every layout traces back to the exact prompt, specs and rules that produced it.
Saying no asks for a reason. Reasons become proposed rules, written in plain English, and the next run applies them.
For small messages, a reviewer's “no illustrations here” became a rule after one review, applied to every run after it.
Landing page layouts that passed the rules, run by run, as review comments, like not leaving a checklist on its own or doubling up buttons in the footer, became rules.
In one session, six rules from feedback, including a ban on calls to action, together ruled out all 240 layouts. Rules interact. So rules get the same review as the designs they shape, with a history of every change.
The templates became examples of the system, not the system itself.
Handover becomes a checklist instead of a surprise.
No more inferring that a page shouldn't have five primary buttons.
So the same mistake isn't made twice.
The AI can follow: from brief to intent, template, slots, copy and rules.
14 core components, with where each can and can't go.
6 landing page slots, and 22 templates approved.
Reviewers' comments become rules.
The slot rules and specs handed over to the engineers building it.
This is the version I can share openly. The full case study, with the client named and the detail left in, is in my private case studies. Ask me for the password.
Available from January 2027
I take on contract roles as a design consultant, head of design or fractional design lead, across product, service and design systems. Remote anywhere, or hybrid across the UK.