How I design with AI
I use AI, mostly Claude and Claude Code, across most of what I do as a designer. Not as a novelty, and not to make prettier pictures faster. I use it to remove the cost of the parts that were never the point, so I can spend more time on the part that is: deciding what to build, and knowing when something is off even though it technically works. Here is what that looks like in practice.
1I prototype in the real stack
The best way to know if a flow works is to use it, not to look at it. On Episode I built the harder flows in code: the generation loader, the brand-confirm step, the Studio shell. When the prototype is the real React app, timing, empty states and error paths get designed against real data instead of faked in a static file and lost in handoff.
Claude Code makes that cheap enough to do early, so I can put something real in front of a teammate or a user before the idea has hardened. Most of it gets thrown away. That is the point.
Below is Yada, a Bible reader I'm building. Claude Code works through the build plan while the app runs in the simulator beside it, and I check each screen against the design.
Tap a verse, or Listen.
2I make the call, then update Figma
Judgment is still the job. AI is good at laying the options out and arguing both sides. It is not the one who decides. A typical loop: Claude reads the Jira ticket and inspects the Figma frame, tells me exactly what is off, and lays out the change. I make the call, and it carries the decision into the file.
| Property | Current | Recommended |
|---|---|---|
| Card padding | 16px | 24px |
| Item gap | 8px | 12px |
| Subtitle | Manage your settings | Workspace settings |
3Research at the speed of the question
The slow part of research was never the interview. It was turning forty transcripts and a survey export into something you can act on before the moment passes. I point Claude at the raw material and my questions, and it helps me find the pattern, which I then verify against the data myself.
On Rivva this is how I pressure-tested the number the whole team trusted. I asked whether the event we called activation actually predicted whether anyone stayed. It barely did: it fired on setup, before a user had scheduled or finished any real work. Cohorting on behaviour instead surfaced a much stronger correlation, which I then verified in the data myself. The synthesis was fast. The judgment about what to do with it, redefining activation around experienced value, was the work.
| Measure | Result |
|---|---|
| Activation as finished setup | 56.8% → 50.2% while users grew |
| 3 tasks scheduled by Nia, 1 completed, in 48h | Far more likely to return |
| Cohort size | About 100 people a week |
4First drafts, never blank
A blank page costs more than it should. PRDs, research briefs, test scripts, screeners: all of them take real time to draft from nothing, and none of them are better for having been started from scratch. I hand Claude the product brief and a template, and it fills in the structure. Then I review, cut and sharpen.
5Checking my work for accessibility
AI is good at the checks people skip. I ask Claude to go through what I've built for keyboard focus, what a screen reader hears, and contrast. Then I check each finding live before anything changes.
Two fixes from Episode in September. Keyboard focus rings never drew: a pair of styles cancelled each other out, so 117 controls across 83 files showed no focus at all. And the whole chat was a live region, so a screen reader read out every word as it streamed. Now it says "Reply ready" once, when the reply is finished.
| Problem | Fix |
|---|---|
| No focus ring on 117 controls | Drop the style that cancels the outline |
| Screen readers read every streamed word | Announce "Reply ready" once |
6Designing the AI, not just using it
Using AI to work is half of it. The other half is designing the AI that ships to users, and there the hard part is restraint. On Rivva I designed Nia as a controlled action layer: it could break a vague task into steps, replan a day, draft a message or suggest the next move, but nothing was scheduled, sent or committed until the user said so.
The same rule runs through Episode's automation, which shows what it is doing and hands control back on demand. When the machine can act, the design job is deciding what it may do on its own, and making the moment of control impossible to miss.
7How I keep AI from shipping slop
AI is quick to make something that works and is slightly wrong, so I never ask it for finished work. I give it real context first: screens from apps that solve the same problem, and the design tokens it is allowed to use. Then I ask for a few directions, pick one, and sculpt it in small passes, one change at a time, saving after each so any pass can be undone.
Before anything goes up for review, I run a cleanup pass for what AI tends to add: comments that narrate the code, checks for things that cannot happen, colours that are not in the system. And a check in the code fails on any colour or text size outside the design system, so drift is caught before a person has to spot it.
The part that does not get cheaper
AI removed the cost of execution across all of this. It did not remove the need for taste. Knowing what to build, what to leave out, and when something feels wrong even though it works: that instinct still has to be earned the slow way, by shipping things and watching people use them. The tool gives me more reps at the decisions that matter, and fewer at the ones that never did.