It is evident that AI tools are here to stay, at least in programming and dev environments. It is nonetheless very frustrating to find fully AI generated slop websites at the top of search engines instead of trustworthy expert written articles. Although I’m not an expert in anything (except for wire ECM), the visible content of this website is fully written by me, although the code behind it might not be.

What I write myself

Unless explicitly stated, AI is not used, was never used, nor will it ever be used to generate main content (i.e. information presented to the reader) for this website. The content of this website is hence guaranteed to be free of (AI-)slop.

What I let AI build

For writing code, AI is an insanely powerful tool and I use it a lot. Large parts of the back-end of this website, the algorithms and interactive figures in the articles (such as the cartographic projections tool) and several of my projects are built with Claude Code as an AI agent.

I use it to implement, not to think. It is strongly steered: the idea, the decisions and the judgement of the result are mine, the typing often is not. Some tools would never have existed otherwise, because I would not have found the time to write them.

In practice I describe what I want, try what comes back and send it back until it does what I had in mind. How closely I read the code differs per project, which is what the badge below is for.

The reasoning stays with me: which question to ask, which method to use and what to conclude from the result.

The badge on a project

Every project carries a badge that says how much AI went into it.

• Human-written. Architecture and code are typed by hand. AI was at most autocomplete, or helped with the documentation.
• AI-assisted. Written by me with an AI as pair programmer. I wrote or reviewed every part.
• AI-built, human-directed. I set the architecture and made the decisions. An AI agent wrote the code.
• Vibecoded. An AI agent wrote the code and I directed it as the architect. I did not need to read the code to get there.

Where it is known, the project page splits this up into architecture, code, tests and documentation.

What I mean by vibecoding

In a vibecoded project the AI writes the code and I do not need to read it to end up with a working tool. I act as the director, the architect and the visionary: I decide what the tool is, how it is built up and where it goes next. The little code I add myself is typed by hand.

I still think it is important to know what is written. An LLM writes code that works in every case. It does not simplify and it does not adapt the implementation to the goal, so what it produces is often suboptimal. Knowing what is there is the only way to notice that and to steer it.

Most of these disagreements about the implementation are caught before any code exists, in an extensive planning phase. My first prompt almost always ends with something along these lines:

Research, discuss, research more, plan, discuss. If we agree on the complete plan, start the implementation.

Models

I use the latest model of Anthropic that my subscription gives me, through Claude Code. At the time of writing (October 2026) that is Claude Opus 5.5. When I run out of Claude tokens, I continue with Codex if I really cannot wait.

The agent does not get the run of my computer. On Linux it runs inside a bubblewrap sandbox: it sees the one project it is working on and nothing else of my home folder, has no SSH keys or tokens, and cannot become root.