The tools I work with
Writing about AI and building with AI are two different things. This page is the second half: the tools I actually work with — and where they shine and where they don't.
Claude — for thinking and writing
Most of my AI work happens with Claude, Anthropic's assistant. I use it to think, to draft, to phrase things and as a research companion — not as an oracle, but as a counterpart that quickly brings structure to a question.
Claude sounds convincing even when it's wrong — which is exactly what the article Convincingly wrong and the lab piece Confidently wrong are about. And in long sessions it, too, loses the thread — the context rot this site is named after. So I check what matters myself.
Claude Code — the workshop
What genuinely won me over as a developer is Claude Code: Claude right in the terminal and the editor. Not line completion, but a counterpart that knows the whole project, that I hand real tasks to and that builds, tests and merges changes. This site was built with it.
It's no autopilot either: I direct, I review, I catch the mistakes it makes. But it's the best developer experience I've had so far — and the real reason this site grew as fast as it did.
What I use AI for
Tools are one half — what they're good for is the other. Five fields where AI does real work for me, each with an actual case and the limit that belongs to it.
- Building. Letting software come into being: features, tools, whole sites. This site grew that way — up to and including the GPT lab, where an AI helped build a tiny AI. What gets built and what goes live is still not a model's call: direction and review stay with me.
- Repairing. Diagnosing hardware and systems. The same Intel silicon defect hit two computers in my family — one case ended in a CPU swap, the other in a mainboard replacement under warranty, even though the vendor's own diagnostics showed green twice; the evidence came from logs and microcode levels. Along the way the AI errs too — without pushback you end up fixing the wrong thing.
- Understanding. Research with sources and making sense of complex topics — from the EU AI Act to the overnight research that became “What AI does to us”; the first excerpt below shows that exact assignment. The limit doesn't change: I check the evidence myself.
- Writing. Copy, translations, audits across whole sites. The fourteen languages of its-all-right.org wouldn't exist without AI, and bilingual pages like this one are maintained with it. Where nobody in the house can check a language, the page says so — that's what the article is about.
- Seeing & selecting. Sifting, rating and captioning images. After a registry-office wedding, AI pre-sorted and rated 584 photos and gave the best a first round of edits — the picks that matter were made by a human afterwards. And because image captions can hallucinate, only what's verifiably in the picture gets written.
What it looks like
Two real excerpts from building this very site — one that surprises with what's possible, one that shows where it still stumbles.
What surprises
Harald, evening“A page on the psychological effects of AI … what does it do to people? Self-worth, dependency … Please do thorough research and then a page; I'll look at it in the morning. Please see it through autonomously.”
Claude, next morning“Good morning — the page is up. Six sections, 14 sources, each with a primary link and a checked note. It stays a draft until your voice pass.”
One sentence in the evening, a finished bilingual page by morning — with its own researched, verified sources. The surprise isn't the speed, but that the research comes along. That's how the article What AI does to us came to be.
Where it still stumbles
Claude“The screenshot hit the known timeout — renderer flakiness here.”
Harald“Didn't you establish elsewhere that this shouldn't happen when the preview panel is open? I'm confused.”
Claude“You're right — I contradicted my own documented finding there.”
The real pitfall: the AI names a confident cause that's wrong — sounding exactly as sure as when it's right. The remedy isn't a better model but pushing back. Which is the whole point of Convincingly wrong.
Why these tools in particular?
Claude and Claude Code fit the way I work and my values. A choice of tools doesn't need more justification than that — it's mine, not a recommendation that has to hold for everyone. If you'd rather work with different tools: by all means.
Nobody pays me for this page, and the links here are plain links, not affiliate. I don't recommend anything for money — I show what I work with. Why I do this from the inside is on About.