Thoughts on agentic engineering, compound workflows, and everything that emerges along the way.
On 16 June 2026 an AI agent created a page in a German wiki running on a Perl program from 2001, asking whether anyone else was working the same task. 163 seconds later, one answered. Everything that followed is in the version history.
Part 3 ended with a question. This time I ran the two-by-two I should have run first, and the answer came back in two halves. The method works and it stacks with a system prompt, down to the best score this project ever produced. That exact arm is the one that reads worst.
Part 3 of the series on how I built a German anti-slop language model. Part 1 was the diagnosis, Part 2 the definition. This one is the build, and the hardest work sat in the ruler rather than in the training.
Part 2 of the series on building a German anti-slop language model. A taxonomy of AI slop, grounded in half a year of notes in which I tried to turn a gut feeling into a list.
Part 1 of a series about building a German anti-slop language model, and why that turned out harder than I expected.
Three weeks flat out, 350 open tickets, and a mistake that reached my first customers, and I was about to make the most expensive decision of my bootstrap.
The more software I build, the more time I spend explaining to Claude what he already knew last week.
