Auto-translated. This page was machine-translated from German with Claude Code and has not been editorially reviewed — the German version is authoritative. Read the German original →
// Case study — promptarchitect.de

This site is itself an AI project.

The page you're reading is the smallest of my examples — and perhaps the most revealing. It came to be in about two days, in a single connected session with Claude Code. Not a swarm of autonomous agents like RAMP, but the opposite: closely guided collaboration.

2 days
9–11 June 2026, note → live
1 session
connected · 100% interactively guided
~560M
tokens processed (mostly context)
25 commits
11 pages + blog, analysis, deployment

All figures come from this repo's git history and the analysis of the session transcripts.

This case study describes the first published version (June 2026). The site has grown considerably since — blog, a moderated comment system, a newsletter, an English edition and more. The sequel tells “This site – Version 2”.

With RAMP the story is autonomy — planning and implementing agents, a test bot, a closed fix loop. Here it's the exact opposite: a single, connected session over two days, guided by hand. I describe, the machine builds, I review and steer. No loop running without me — a close dialogue in which I'm there at every fork.

Voice in, structure out

At the start there were no texts but speech: hours of self-recorded notes on what I've experienced over months with Claude Code. Transcribed locally — without the cloud —, those spoken notes became structured, readable text. The AI didn't invent the topic; it shaped my skeleton. How I do that I'll happily explain in a direct conversation — what counts here is the result.

Plan first, then build

Before a line of code: information architecture, content map, the stack decision — a static site, no CMS, rebuilt and shipped on every change. Nothing exploitable at runtime. Only then the build, then the content.

An independent second read

Every public text goes through a deliberately unbiased AI second read before publication — fresh context, no conversation bias. This text went through it too.

What the numbers tell

The roughly 560 million tokens processed are largely not new text but the same context that rides along at every step — exactly the phenomenon I write about on the blog. The pure waiting time on the machine was about two hours; the rest was two days in which human and machine took turns.

What this shows

RAMP shows what's possible autonomously. This site shows the other end of the scale: what closely guided pair engineering achieves on a small, fast project — and that the substance comes from the human, the speed from the machine. Both are the same discipline, just dosed differently.