Tinkering with code using AI

There's no need to introduce AI or how it's taking over software development. Opinions differ on how much it will replace humans, but it's cemented itself as a programmer's primary source of information, the way Stack Overflow did in the pre-AI era.

As a seasoned programmer and homelabber, one thing I love about AI is how easy it makes getting a small personal project going, and how it keeps things fun. A recent example: I wanted to hook up Apple Health to Claude for yet another attempt at getting a solid workout plan out of AI. I'd tried various "agents," "projects," "prompts," and so on before, with disappointing results. This time I had hope. Combining access to Apple Health, Calendars, and Claude Cowork, a Medium article proposed a tempting setup for creating, following up on, and personalizing workouts for me. The joy was short-lived, though.

The Apple Health connection isn't available in the EU (or something like that; I didn't dig into why, I just couldn't get it working). That made the whole setup from the Medium article useless, since it needed access to daily key metrics like sleep, HRV, steps, and more. But this wasn't the end of the road. Long before AI, my first instinct would have been to find some API or other way to scrape the data I needed. In my case, I use a Garmin watch that feeds all the relevant metrics and events into Apple Health anyway, so I could just cut out the middleman.

Garmin has no public API, but there's a great Python package called garminconnect that uses the same API the Garmin apps do. You log in with your own account, grab a refresh token, and you're good to go, with 100+ endpoints exposed. More than enough to get the data I needed.

Using Claude Code, I went from prototype script to a Docker-ready server deployment in under 30 minutes. Pre-AI, I'd have probably spent at least an hour figuring out how the Python package works, setting up a venv, wrestling with auth, and so on, before even attempting to fetch the data. I have a short attention span, so that would likely have been the end of that side quest. But since the loops with AI are so short, I get to my goals wayyyy faster. Claude also helped me figure out an iPhone Shortcut that triggers on a schedule to fetch the data and paste it into the workout Claude project I'd set up from the Medium article.

The cherry on top was how fast I could deploy it on my own private infrastructure. The core tools here were Coolify and Tailscale: Coolify to manage deployments (from a self-hosted Gitea instance) on my homelab, and Tailscale for private, secure remote access to my local infra at all times. Short attention span strikes again: if deployment had taken ages to figure out, or if I'd had to connect to a VPN every time I wanted to use the service, I'd simply have given up. AI helped me debug the small deployment issues, and within an hour I had the whole thing set up, secure, and working on my phone.

This story is just one example of how AI has improved most of my computer-related workflows. By cutting down the amount of knowledge I need before I can ship something working, it's let me be more creative and productive. I'll leave the question of how this affects programmers, the workforce, or humanity for another time. For now, I'm happily shipping small lifehacks faster than ever.