Eric Migicovsky had an incredible 2025: he announced the rebirth of Pebble, launched 2 new Pebble watches, shipped one of them, and launched a new weird wearable ring.
And then the SOB starts off 2026 with a bang by announcing a third new Pebble just on the second day of the year: Pebble Round 2! Oh man not sure if he can keep this up, but I sure hope so!
One of his rationals behind an early announcement of the Pebble Round 2 was to give people who have an open order for the Pebble Time 2 the option to switch their order to the Pebble Round 2, which after some internal debate I did. And the reason I did is because I really miss my Hamilton Khaki Automatic and thought: I bet I could program a watch face like it for my Pebble Round 2. The image above is the twelve iterations I did over a day, not programming but asking different AI models to do the heavy lifting for me. So read on for my experience!
ChatGPT
I started out using ChatGPT and at first asked it to modifying an existing round watch face example provided by Pebble but I quickly realized that was way too confusing and just asked it to start from scratch from a picture I uploaded. It immediately identified it as a Hamilton Khaki-style and listed the traits to copy:
I told it to go ahead and it provided me the header and c files that I copied to my working folder, crossed my fingers, and compiled it. And it even worked! It did not quite get everything right on the fist iteration, but it compiled and even ran in the emulator, something that could have taken me a couple days to have achieved on my own:
Gemini
After three iterations with ChatGPT we had gotten as far as we could and the code it was providing began giving errors when compiling, so I decided to make the jump to Gemini. Gemini got on board with a simple prompt and providing the current code:
Gemini took me through 5 more iterations and was even getting a bit snarkey:
Claude Code
After multiple more iterations with Gemini I was having problems trying to get the right font for the numbers and at some point in the evening I was chatting with my friend John who said he had gone down an AI rabbit hole and commented “I use up my subscription time on Claude and it needs “a break” from me for 6 hours or sometimes 2 days. I am a scorned lover and set an alarm to run back.” And that reminded me of this post on X from Boris CHERNY (creator of Claude Code) that had taken the development world by storm. When I read the post I really could not make heads or tails of it, but it did peak my interest and now my friend John was confirming Claud was like crack I knew I had to give it a try.
Claude Code requires a PRO account so I had to subscribe for my first time ever to an AI service but it was painless as was installing Claude Console in my Windows Linux Subsystem. Once installed I went into the directory where the source code for the watch face was, and began telling Claude Code what I wanted. It analyzed the code, came back and proposed the changes. I told it ok and IT made the changes, compiled it, and loaded it into the emulator! No longer copy and pasting. Just tell it what I want to it begins working. It got even crazier when I kept insisting that the font was not quite right and so ClaudCode recommended installing an additional font. I told it to go for it and it handled it all installing the font in linux and then compiled and launching the emulator. IT GOT REALLY CRAZY when the font still was not quite right and ClaudCode recommended creating a new font just the way I wanted. I gave it the ok and it installed some program to edit fonts, made a new custom font, included it in the project, compiled it, and loaded it in the emulator and it was spot on!
I finally after 12 iterations I had a result I can’t wait to test on my actual Pebble Round:
Post vibe coding reflections
I really cannot express what a fundamental change Claud Code represents. It is a true “AI Agent” there doing the work directly. I now understand the paradigm shift that Boris CHERNY’s post represents: every morning he fires up 5 AI agents and puts them to work on different tasks while he does the guidance and oversight. Instead of a one man show, he now is as productive as 5 or as many agents he is able to manage at a time. My only word of caution would be not to drink and manage AI agents because as what happened to me in the wee hours of my vide coding session: my oversight began to get a bit spotty because as Claude Code was doing all the work, I did not need my hands for typing so I just sat back and drank beers hitting “yes” every once and awhile and ended up having way too many beers.
Links, References and things that helped with this
Thanks for reading and feel free to give feedback or comments via email (andrew@jupiterstation.net).