Design engineering when AI writes the code
I design and build the same product, and AI writes most of my code. What my job is now, with three small misses from this site that show why.
I don’t write code anymore. I still ship it every day.
At my day job I design and build a product for sports and leisure venues: online ticket sales and the admin panel venues use every day. Most of the code there is now written by AI. This note is about what’s left for me to do, and why that part matters more than it did a year ago.
Design engineer, in my words
To me, design engineering is an evolution. Designers always wanted to ship their own work. Developers always wanted nobody to tell them how to build things. A design engineer sits in between: one person from the first sketch to the working screen, so nothing gets lost on the way.
I’ve done that job for years. AI didn’t change what it is. It changed where my time goes.
AI builds, I look after the system
Does AI make mistakes? Yes. Do I fix them? Yes. Would it ship a good feature on its own? No.
Broken code is the easy case, because you see it. Harder is code that works and doesn’t fit. Someone without design experience types “add dark mode to the whole site”, and AI does it. Black turns white, white turns black. A real dark mode needs new tokens, contrast checked again for every pair of colors, brand colors that still hold up on a dark background, the smallest icons, every hover and focus state. You have to know all of that exists to ask for it.
Same with something small. Ask AI to make a button smaller and you’ll get a new button: a new component, new code, one more thing to maintain. Ask it to add size="small" to the button you already have, and nothing new appears.
AI solves the task in front of it. It doesn’t ask how the answer fits the design system, or where the product is going in a year. That’s my job now.
Three small misses from this site
I build this site the same way, with Claude Code. These three come from the last two weeks. None of them broke anything, and all of them would have gone live if I hadn’t looked.
The line height that didn’t change. When Claude built the notes page, it gave the text a line height of 1.6 for easier reading. On the page it stayed at 1.4, because another style on the same element won. The code looked right, so nothing flagged it. I noticed because the text felt cramped. The fix was a separate text style for long reading, used only in posts.
The right style in the wrong place. Post titles used Heading 1, the style from my home page. That’s what a design system is for: reuse what exists. But Heading 1 is set tight, for one short line in the hero. A post title runs two or three lines, and they nearly touched. Post titles now get more leading and slightly looser tracking. A design system tells you what to use. It doesn’t tell you where it stops working.
White that was too white. In dark mode, body text had a contrast of 16:1. On paper that’s great. In a long post, near-white on near-black glares. I added a softer text color for posts, about 11:1. Still far above what accessibility asks for, and easier on the eyes.
AI made each of these changes in about a minute. Noticing them was the job.
How a screen goes from Figma to code
For simpler features I mock up the screen in Figma, roughly. I skip the slow part, setting every token and fixing every spacing, and send AI a screenshot with something like “this is messy, clean it up using our design system”. In the product I work on, the design system lives in SCSS variables, so AI can see which tokens and components exist. The first version is usually good. Other times I connect Figma through MCP.
Then comes review, and that’s the part I can’t hand over. AI misses something specific to that screen, gets mobile wrong, adds an animation nobody asked for or a label nobody needs. Now and then it breaks something else along the way.
So every screen goes through the same list:
- Values come from tokens, not hard-coded colors or pixels.
- It uses an existing component, or a new variant of one, not a lookalike copy.
- Every state is there: hover, focus, disabled, loading, empty, error.
- It works on a phone.
- Contrast holds in light and dark.
- Nothing extra: no animation, label or wrapper that doesn’t earn its place.
- Nothing else broke.
Five ideas instead of one
The biggest change for me is prototyping. Building a prototype in code is now faster than drawing it in Figma. With deadlines, I used to stick with the first idea that worked, even when I knew it wasn’t the best. Now I can try five in the same time and keep one.
AI also helps before anything is built. I take the requirements and ask what good practice looks like for that kind of screen, and how other products solve it. It points out where the request makes sense and what’s worth questioning. Sometimes the better idea comes from AI. Usually it’s a mix.
The time I save goes into questions AI is bad at: how to make a feature simpler, what to remove, how to check that it works for people, and how to present it so they use it.
Where AI still falls short
Design from zero. Even with a detailed prompt, without a reference it makes something generic and adds things nobody needs. We all know the little label above every heading in every section, whether it says anything or not. It drifts from the design system and invents elements that already exist. And business logic gets harder for it with every variable you add.
Designer or developer?
My bet is that design is further from what AI can do than code. AI writes code well. It still can’t tell what a screen should be. The three misses above are small, but each one needed someone to look at the page and feel that something was off. That’s taste, and I don’t know how to prompt it.
So I think it’s easier today for a designer to grow into code than for a developer to grow into design. That doesn’t make developers useless. We’ve all seen the memes: the .env file pushed to a public repo, the app launched on localhost:3000. That knowledge still matters, and designers who start building should learn it instead of skipping the boring parts.
Who this matters for
If you’re a designer, learn to read your design system in code: tokens, components, variants. That’s where AI looks first, and it’s where you’ll catch most of its mistakes.
If you run a product, one person from the first sketch to the working screen means fewer handoffs and less lost between design and code. With AI, that person can also try more ideas before picking one.
AI raised the bar for designers and developers alike. I still design, and I still build. I just type a lot less.