I barely ride the metro

I barely ride the metro. Two weeks ago I built it anyway - every station and tunnel beneath Prague, modeled in 3D with real-world depths wherever the data allowed. More than 35,000 people opened it in 10 days. But this isn’t really about the metro. It’s about what’s left for a designer to do when making things stops being the hard part.

The wrong person for the job

I'm not a transit person. Ask me about rolling stock and I'll change the subject to cars or watches. I love to drive cars, which in Prague is an adventure of its own.

What I do like are cross-sections, side views and 3D maps. Anything that shows how something is built, not just how it looks. I also care about navigation, HUDs and type more than is probably healthy, and I've never been fully convinced by where the Prague metro's new wayfinding was heading.

So one evening, in between work on the new Countly, I opened Claude Code and started playing with metro data. A few evenings later, the first version was live.

For me, not instead of me

For a long time I was deeply skeptical of AI in software design. Everything I saw was about speed and savings. Look what I built in a day. Look how much we saved.

What changed my mind was something else: how much easier it had become to turn something I could imagine into something I could actually use.

AI wrote the code. It didn't design Metroskop. That difference is the whole point.

I love designing interfaces. I grew up on Sketch, and I've never believed in a perfectly systematic process. Real design is messy. Exploring. Dozens of paths that lead nowhere, and somewhere in that mess, something that works. Working with AI feels a lot like the years I spent standing behind designers. A bit to the left. No, there. Perfect. Yes, exactly like the old meme where someone keeps moving the toilet around the bathroom.

Left alone, AI has a habit of drifting to the same place. Beige backgrounds. A sans paired with a serif italic. Purple gradients, emoji for icons, cards with a vertical colored stripe, big numbers that say nothing. And it keeps adding. Another feature, another effect, another paragraph. Onur Alp Soner has a fitting name for where that ends up: a big pile of ****.

So most of my job was saying no. A diorama style. A glowing heatmap. Thicker tunnels. Extra stories. A lot of it could be good, and none of it helped anyone understand what's under the street.

What survived, I tuned by feel. Over and over, until it felt right.

Many sources, no truth

There's no complete public 3D model of the Prague metro, and detailed architectural drawings are hard to come by. What exists are fragments. OpenStreetMap shows where things are, but not how deep. Wikipedia "knows" depths, but not where they’re measured from. Transit data knows the floors, but not their shape. Then lidar terrain, the city’s digital technical map with the real track positions and heights, the operator’s pages on lifts, photos, and the occasional cross-section from the operator.

Each source is precise about one thing and almost silent about the rest, and they often disagree. So the rule became simple: take from each only what it’s good at, derive the rest with explicit rules, and label every estimate as an estimate. The data is the skeleton. The model is mine.

Two moments made that rule real. At first none of my sources had the track height between stations, so the tunnels were interpolated, and until we capped them at 8 meters below ground and 12 below the river, the tunnel under the Vltava popped out above the riverbank. Later the city’s digital technical map gave us the real track heights. And Flora station sat 54 meters off its own tracks until stations started snapping to the real rails. Neither showed up in the numbers. Both were obvious the moment I looked.

That’s the part AI can’t do for you. It will always produce something plausible. Someone still has to look at it and say: that doesn’t feel right.

Just because I wanted to

After all that cutting, I made room for two things I simply wanted to build.

The first is station interiors. They were a whole new kind of hell. I'm not a 3D artist, and I had to figure things out as I went. They're still unfinished and labeled beta, and for good reason. Ultimate realism was never the goal. It's a map, not a ship in a bottle.

The second is Ride mode. Pick a line, cover a train in graffiti, then hop on and glide through the tunnels with music playing. It answers a question nobody asked. It's also the part I open most.

Every product needs one thing that exists simply because someone wanted it to. That's feel, too.

Better data, better map

Of course people wrote in to tell me that something isn’t accurate. A lift at the wrong end of a platform. A hall that looks different in real life. They’re right. It isn’t accurate, and in places it never fully will be.

Metroskop was never meant to be technical documentation. It’s a popular science project, something to explore and enjoy, and it’s only as good as the data behind it. I've asked the transit operator for better data. Until then, I'm working with what's publicly available, marking estimates as estimates, and fixing what people point out.

So I work with what’s public, mark every estimate as an estimate, and fix what people point out. Every report makes it a little better. That’s the deal, and I’m happy with it.

Still my job

I've never been particularly interested in drawing hard lines between design and development. I've known developers who could build beautiful products end to end, and designers who could code their own. AI is blurring those lines even further. I don't particularly care what we call the job. I care about making things worth making.

The tools made Metroskop possible in about a week. They didn't decide what was worth building. That's still my job, and it's the best part of it.

If you have ideas, and you're never quite satisfied with anything, this is an incredible time to be a designer. The distance between what you can picture and what you can actually build has never been shorter. Just let AI work for you. Not instead of you.

I still drive almost everywhere. I just know what's under the road now.