An experiment, not a case study
What happened when I asked AI tools to design beyond the phone
A heads-up before you scroll: this is not a case study. There's no user problem, no research phase, no gap I heroically found in the fitness market. Nobody needs another workout app, including this one. It's just a vehicle. This is an experiment about tools and craft: how good are AI design tools, really, when you push them off the screens they know? Everyone's generating landing pages and mobile apps. I asked for a TV interface and a watch face instead, devices with their own rules. Short answer: the tools fail, in genuinely interesting ways. What follows is a working two-device prototype and a field report.
Try it: arrows move focus, Enter selects, Esc goes back. Tap and touch work too.
The Prototype
This is a working TV interface, and your keyboard is the remote. Arrow keys move focus and the background follows; Enter commits. Mid-workout, press anything and the controls surface; ignore them for five seconds and they slip away. That appear-when-summoned rhythm is the heart of 10-foot design, and the thing no static mockup can show. The session keeps honest score too: skipped exercises are marked skipped, not done, and ending early gives you a checkpoint with stats, not a cliff.
The watch isn't a second mockup. It's a second client of the same session: same heart rate on both screens, coaching cues landing on the wrist as a haptic-style flash, and it works backwards, so pausing from your wrist pauses the TV. It gets exactly one loud moment. Cross the workout's halfway point and the dial erupts in sparks for three seconds, then back to quiet. Everything else whispers so that can shout.
Image: milestone celebration, sparks mid-burst
Built to spec
Nothing here is eyeballed: the TV is calibrated to 10-foot guidelines at 1920px (24px+ body text, 5% safe margins), the watch to a 450px Wear OS dial with every tap target above 48dp. What's embedded here is a true scale model of both devices.
The Process
01
The pretty, wrong pictures
I started by prompting Figma Make and Stitch. The outputs were beautiful and confidently wrong. Every TV screen came back wearing a desktop sidebar, and real TV apps have no chrome during playback. The tools have seen a million dashboards and maybe twelve TV interfaces, and it shows.
Image: early generation with the sidebar
Image: best static attempt
02
The realization
Around the fifth "no, that's a mouse pattern," the actual problem surfaced. I was asking Stitch to show behavior (focus, summoned controls, attention) that can't exist in a still frame. The medium was wrong, not just the mockups. Figma Make came closer, and its output was genuinely decent, but the dimensions never held up: it kept designing to phone and desktop proportions, never a 10-foot screen.
03
Building instead
So I switched to writing code with AI, iterating in conversation: "it shouldn't start instantly" became the overview and countdown; "skipped isn't completed" became honest stats; "can I resume from the watch?" became bidirectional sync. Every feature is the residue of a correction. The watch came last and proved the point: instead of designing its screens, I gave it a radio to the TV session, and designing the relationship beat designing either device.
Image: Iterations
