For the most part of the past years, ever since Chat GPT first got released, AI tools weren't much more than a gimmick. Good at killing time, but not much more. 

This changed in 2026. Using Claude Code for several months now, I see productivity increases when prototyping or developing software in general by multiples. Of course, it's still AI slop and relying on the output of a AI for anythng more important still seems hazardous. Yet, the productivity gains accelerate and the output is becoming better with every passing day.

The first larger setup I tried was to create a CMS for Steffis yoga site (see https://yoga.drewing.de) on a raspberry pi. I started out with telling Claude to have a look at the static site and create a CMS that would be capable of generating this site from a Posgres DB running on the PI and deploying it automatically via SSH to our Strato web space. The most effective part of using Claude is imho creating the technical structures. Where it is really lacking is in terms of visual design: If you try anything like this you'll end up writing a lot of the CSS yourself since the tool still has no idea what typography is. But that's okay. If I had written the CMS manually it would have taken months instead of a couple of weekends.

During the current week I revived an old XCode-project of mine of a more geeky nature. A text adventure engine for the iPhone written in Swift. I already hat it running with the manual approach, but I needed an editor app to go with it in order to create the contents - otherwise I would have been working in XML directly, making the whole endeavour pretty much an editing nightmare. Bringing the old code up to speed and adding the editor tool took some evenings this week. Earlier this would have been several weeks of intense work. Using the 20 € pro plan for claude this really is a bargain.

But it doesn't stop with these rather simple apps – today I told Claude to create an area measuring app as an XCode-Project, since the regular measurement app from apple doesn't allow for measuring areas. Real work hours input from my side was about 30 minutes.

After that I wondered how effective Claude could be when prompted with the task of creating a photogrammetry app using all sensors and the camera of the iPhone to create a 3D model of a real world object and then offering the data as a textured Collada file (.dae) to download directly from the app. Well ... the mesh certainly isn't perfect, but for an hour of work, the results from the app are very much okay.

One caveat though: It's worth adding a skill for Claude to regularly create git commits whenever he creates or changes anything in a repo, for he does mess up sometimes. And that's when you are really happy you've got the repo under git control.

Prompting claude to user Uncle Bob's clean code principles and the 12-factor app development ideas as well as asking him to check for memory leaks helps also to keep the code readable and functional.

Btw.: I had Claude take my old Go repository of the static site CMS that I wrote years ago, modernise the code and add a locally running admin web backend to it. If you can read this, that's one project that works fine with AI-enhancements ;)