How I Design with AI
AI made making cheap. It didn't make judgment cheap.
The question design teams ask has changed.
It used to be "have you designed AI features?"
Now it's two questions:
how does AI fit into your daily work — and how do you judge what it gives you?
Here's my short answer to both:
AI touches almost everything I make.
It decides almost nothing.
It writes my code, tags my research and judges my icon sketches.
But it never gets the final say on a font, a number, or what ships.
A font is taste, so that's mine.
A number is a fact, so code computes it.
Shipping is accountability, so nothing goes out until I've seen it.
AI made making cheap. It didn't make judgment cheap.
So most of my work with AI isn't prompting. It's deciding, before I start, which part of the output can be checked, and by what.
Every day: Claude Code
I build in code most days now. At work, on Optaire, I create UI, workflows and flows end to end, mostly in Claude Code.
I've also worked in Cursor, GitHub Copilot, Antigravity and other AI-native IDEs.
The tool matters less than the rules I put around it.
When a team needs to react to an agent, a slide doesn't help. So I built a full-stack prototype on LangGraph and LangChain, where PMs and domain experts test real agent behaviour instead of imagining it. Earlier, at Innodeed, when I'd just started exploring AI-native IDEs, I built the CKYC automation frontend in Cursor and shipped it in under ten days.
This portfolio is built the same way, with Claude Code, under a rules file for the project. The rules matter more than the prompts:
- I do all the visual design. Layout, type, color, spacing, motion. Unless I ask, the model stops at plain, temporary markup. If it catches itself choosing a font, a spacing value or an easing curve, it's been told to stop.
- "Done" is checked, not claimed. Nothing is done until typecheck, lint and the build pass — and the build fails if a page shows words the site's AI assistant can't read.
- Both screens, every time. Every UI change is checked at phone width and desktop width before it counts.
- Motion is drafted first. New motion gets a prototype, videos and a storyboard, and waits for my go before it touches the site.
- Nothing ships until I've seen it. Design work stays uncommitted until I've reviewed it locally.
Why so strict? Because the model is fast and agreeable.
It will happily make a hundred decisions I didn't ask it to make.
The rules keep the decisions that are mine with me, and hand over everything that can be verified.
Friction first: a Figma plugin
The habit underneath all of this started small.
In 2023, a 30-second annoyance in Figma — making pages one at a time — became Add Multiple Pages, a plugin I built with AI that adds and names as many pages as you ask for. More than 1,000 designers use it now.
It taught me that good tools start from a real annoyance, not from a technology looking for a use.
AI made building tools cheap.
So choosing the friction worth a tool is now the hard part.
Icon Foundry: code draws, the model judges
Icon Foundry is my side project: a foundry where a team's icons are made, checked and shipped. 1,855 icons, nine house rules, and packages for React, Vue and Angular.
Its first rule comes from what models are bad at. Drawing new geometry is what models do worst. Placing a library's own parts and cleaning up a designer's sketch is what code does best.
So the work is split:
- A Librarian agent names the concept, or finds the icon that already is it
- A sketch is drawn several times, and a Judge keeps the one that reads
- The foundry's own code draws the icon, checks it against the house rules, and writes the files
Making and checking an icon costs no tokens. The model only proposes, plans and judges.
Then I gave it an MCP server — MCP is the open standard AI apps use to connect to other tools. Claude or Cursor can now search the library, make an icon, check it, export it for Figma and save it.
That changes what a design system is for.
An agent that can look up the real icon doesn't invent one.
The foundry also says what it can't do yet: seven logistics icons — cross-dock, milk run, emissions and four more — still need a designer drawing them by hand.
Review Insights: a number a model writes can't be checked
Review Insights is a free, open-source tool that turns Google Play reviews into design decisions.
Most AI review tools ask a model to summarise the reviews. You get confident text you can't check.
So I split it into three layers:
- Evidence. A model tags every review. Risky tags — "they're leaving", a competitor's name — are kept only when the review says so in its own words.
- Analysis, computed in code. Counts, trends and scores are calculated, never written by the model.
- Decisions, written by AI and checked. Every recommendation cites the reviews behind it, and any citation to a review that doesn't exist is removed.
The AI writes; the numbers are computed.
A finding with a review under it is evidence. A finding without one is an opinion.
It has an MCP server too, so the analysis can be read from Claude, ChatGPT or Cursor.
One output I rejected
When I moved my case study on search suggestions in an AI chatbot from my old portfolio to this one, I asked Claude Code to migrate it.
What came back was clean. Consistent. On template.
And it was a summary of a case study.
It had squeezed a detailed study into the same generic set of sections every study on the site uses. Each sentence was fine. The explanation was gone. My old site explained the project better, and in more detail.
I rejected it because a case study's value is the reasoning. A summary keeps the conclusions and throws away the thinking — the one thing a reader is there to judge.
So we rebuilt it in full: the coded stakeholder interview, the empathy maps, the journey map, the live sales call replayed, and a search box you can type into to feel the debounce, the cache and the fuzzy "Did you mean?".
AI's usual failure isn't being wrong. It's being average.
Left alone, it falls back to the template.
The rule I kept from it: only a case study's cover follows the template. Every body is explained its own way.
How I judge AI output now
Before I accept anything a model makes, I ask:
- What here can be checked — and what checks it?
- Is this the problem talking, or the template?
- Which decisions did it make that I didn't ask it to?
- What's left that only taste can decide?
Code checks what must be true. Evidence checks what's claimed.
I check what's left: the craft, the voice, and whether it's honest.
AI made making cheap. It didn't make judgment cheap — that part is still mine.