I use AI to do real work,
and write down what happens.

Twenty years building digital products, ten of them on productivity software, now spent finding out what AI does for the work you already have. I ship with it every week and publish what I learn, including the parts that fail.

I was a lead engineer on Wunderlist before Microsoft acquired it in 2015, then a Principal Product Manager there. I co-founded informed, a news company backed by HV Capital and 468 Capital. Today I am CTO of Umane, an athletic health club in Berlin where I build the platform and the digital twin behind it, and co-founder of Supercharged, where we run AI workshops for teams.

Most of my week is spent writing code with these tools. I have put several apps and platforms into production that way. Not prompting an assistant for snippets: building the context, the tools and the guardrails a model needs, then running it against a real codebase until it holds up under load.

I keep a harness of my own for this and rebuild parts of it most weeks as the models change. Every release gets tried on something real rather than a benchmark, which is how you find out that the thing everyone is excited about breaks on your data. That is where the writing comes from.

What Actually Works

Every week I take one real task, hand as much of it as I can to AI, and write down what happened, with the prompts, the cost, and where it broke.

Recent issues:

A month of invoices in one command, and the trap month I built to break it
Two piles, then drafts: an agent on my inbox for a week, and the two things I had to undo
The routine that lied to me, and why I keep it running anyway

It goes out on Fridays. Everything in it is tested before it reaches you.

Writing

The routine that lied to me — A scheduled agent researches my newsletter every week. Twice it has been confidently wrong, and both failures were worth more than the research.
Two piles, then drafts — Handing an agent my inbox for a week. The instruction that made it work was the one telling it to stop.
A month of invoices in one command — I built Claude a bookkeeping skill, then built a month designed to break it. It scored 100%. The same model without the skill scored 90%.

Everything else is in the archive.

Working together

I co-founded Supercharged in Berlin with Ole Tillmann. We run AI workshops and hands-on sprints for teams: Ole works with leadership on strategy and adoption, I build the technical side, the agents and automations that keep running after the workshop ends.

Most of the teams I work with are past the demo stage. They have tried the tools, seen something promising, and now need it to work when they are not watching. If that is where you are, send me an email.

Elsewhere

LinkedIn · GitHub · X · Email · RSS