What struck me here is not that Claude can help with a tiny firmware project. That part is almost boring now. The interesting bit is how quickly the constraints of the device turn the whole exercise into a very human workflow: the model can get you a rough result fast, but once you hit text placement, rendering limits, and hardware latency, you’re back to being the person who has to notice what looks wrong.
I actually like that honesty in the post. It doesn’t pretend the watch face was “fully agentic” or magically polished. The author says the first pass was rough and that text overlapped because the model was basically guessing at sizing and positioning. That sounds right to me. Small-screen UI is one of those areas where LLMs can get you 70 percent of the way there and then stall on the annoying last mile. You need eyes on the screen, or at least screenshots in a tight loop. Without that, a model can produce something that is technically plausible and visually unusable.
The more interesting move was the one where he stopped trying to be elegant and just shoved the static parts into a background image. That feels like the sort of hack a real engineer makes when they’re trying to get a thing done on weird hardware. It’s also a nice reminder that “best architecture” and “works on a watch” are often different questions. If transferring a 240x240 image over Bluetooth takes about 10 minutes and refreshing the whole screen takes 1-2 seconds, then yes, you’ve probably hit the edge of what this device wants to do. But for a personal prototype, maybe that’s fine. I think that tradeoff is the whole story.
I’m a little less persuaded by the “Claude did it” framing, though. The author quietly says he actually did most of the work in OpenCode with open-weight models like Kimi K3, K2.6, and DeepSeek v4 Pro and Flash. That doesn’t weaken the post; it just makes the headline a bit squishier than it first appears. Which, to be fair, is normal in this ecosystem. People say “Claude” when they mean “an AI coding setup with Claude somewhere in the loop.” Still, if you’re reading this looking for a clean Claude-only case study, that isn’t really what it is.
What I do take from it is that the combination of documentation, simulator, and a small surface area is what makes these projects viable. The watch wasn’t special because it was smart. It was special because it was hackable, cheap, and boring enough that a model could make progress without needing a giant app architecture in its head. That’s the sweet spot for coding agents right now, I think: not glamorous product work, but tiny systems with good docs and a fast feedback loop.
There’s also something refreshing about the tone. No grand theory, no “the future of software” sermon. Just a drawer, a cheap device, some iteration, and a watch face that went from forgotten junk to something usable in a few hours. That’s the kind of AI story I believe.
Reference: Hacking with Claude on a $27 Smart Watch
