Three weeks ago, I decided to go all-in on using Claude Code as part of my design workflow. Not "let me try this once and see." Fully committed. Prototyping, refining, exporting, breaking things, trying again.
68 sessions. 374 messages. 88 screenshots. Some genuine breakthroughs. Some spectacular frustrations.
Here's what actually happened.
The high: designing at the speed of conversation
The best moments felt like having a design partner who never gets tired and builds things instantly. I'd describe a UI direction, a Miller's Column navigation, an onboarding flow, a policy compliance dashboard, and within minutes I had something live in the browser to react to.
My favorite discovery was how naturally the "show me, then fix it" loop works. I don't spec things upfront. I never have. I need to see something before I know what's wrong with it. And with Claude, I could see something almost immediately. Then I'd fire off corrections: "tighter spacing," "softer shadow," "the animation needs to ease out, not ease in."
I built interactive prototypes, refined component animations down to the millisecond, designed responsive layouts across breakpoints, polished hover states. All through conversation. For a designer who doesn't code, it genuinely felt new. Like discovering a material I didn't know existed.
At one point, I had custom commands set up for my most common tasks. One for normalizing spacing. One for polishing the overall visual feel. One for handling animations. I was directing design work the way a film director works: setting the vision, reviewing takes, calling adjustments. The "engineering" was invisible to me.
The low: the day my design disappeared
Then there was the day everything went sideways.
I'd been refining a website redesign for about an hour. Adjusting typography, tweaking colors to match our design tokens, getting the component spacing just right. I was in the zone.
And then Claude's conversation memory compacted. It basically forgot everything we'd been building together. The design, the decisions, the context. Gone.
When I asked it to rebuild, the result was too safe. Too generic. It didn't match what I'd lost. I was staring at a conservative approximation of an hour's worth of design thinking, and there was no way to get back to where I was.
That one hurt.
The ugly: setup tax
Something nobody mentions in the polished AI workflow demos: the setup tax.
About 20% of my sessions, roughly one in five, were just getting things connected. Expired authentication tokens. Plugin connections dropping. Tools not loading. I'd sit down to design and spend the first fifteen minutes debugging infrastructure instead.
Across three weeks, at least eight sessions had zero actual design work. Just configuration. Eight times I showed up with creative energy and got stuck in a parking lot.
I also ran into Claude taking the wrong approach on the first try more than I expected. Once it spent ages exploring and planning when I just wanted it to start building. Another time I asked it to export something to Figma and it prepared a Jira ticket instead. Small things, but they add up when you're trying to stay in flow.
What I'd tell another designer
If you're a product designer thinking about trying AI tools in your workflow, here's what I'd want you to know.
Start with prototyping, not production. I didn't try to replace my engineering team. I used Claude to build throwaway prototypes, quick visual explorations I could react to and refine. No git, no deployment, no production concerns. Just: show me something, and let me shape it.
Your design eye is the actual skill here. The technology builds things fast. But it doesn't know if something looks good. The ability to look at a preview and say "the weight is off" or "this needs more breathing room," that's the irreplaceable part. AI gave me speed. My taste gave me direction.
Save your work somewhere external. I learned this the hard way. Conversation context can disappear. If you're making design decisions you care about, write them down somewhere the AI can't lose them. A simple document with your design tokens, key decisions, and current state will save you from the heartbreak of a lost session.
Budget for friction. This isn't a polished consumer product yet. Things break. Connections drop. You'll have sessions where nothing works. If you go in expecting that, you won't quit after the first bad day. And the good days make up for it.
Where I'm headed
I'm not done exploring. Three weeks in and I've already built a design-to-Figma pipeline that would've seemed impossible to me a year ago. I'm a product designer who can't really code, working inside a code tool, producing better design work than I was making in Figma alone.
Next I want to make Claude check its own work. I want to set up a loop where it builds a UI change, screenshots it, evaluates the result against criteria I set, and fixes problems before I even look at it. A design assistant that does its own QA rounds before presenting work.
I also want to try running parallel design variants. Ask for four completely different directions at once and pick the best one, instead of iterating one direction at a time. That could turn a forty-five minute exploration into a ten minute decision.
The tools are rough. But I'm having more fun designing than I've had in years. And that's the part I didn't expect.