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 · Version 2

This site isn't finished — it's alive.

The first version was up after two days. This is the sequel: what this site became in the week that followed — a blog with several posts, a moderated comment system, a newsletter, a full English edition. A build turned into a system you operate. This case study tells the whole arc — from the first self-recorded note to the bilingual site you're reading now.

7 days
9–16 June 2026, one continuous thread
150 inputs
typed by hand · 100% interactively guided
~1.55B
tokens processed (V1: ~560M)
64 commits
blog, comments, newsletter, 2 languages (V1: 25)

All figures come from this repo's git history and the analysis of the session transcripts. By “inputs” I mean the messages typed by hand to the machine; it responded around 3,000 times, totalling about three hours of pure response time.

With RAMP the story is autonomy. With the first version of this site it was the opposite: closely guided collaboration, to a live site in two days. This second version is about what came after — and about something a one-off build can't show: duration.

The heartbeat of the project

Each bar is an hour, its height the number of inputs in that hour. You don't see steady assembly-line work but bursts: on the left the dense block in which the first version was built, on the right the targeted growth spurts in which blog, comments, newsletter and the English edition were added. In between, pauses — the 13th stayed entirely empty.

V1 · the build V2 · the growth 0 5 10 Tue 9 Wed 10 Thu 11 Fri 12 Sat 13 Sun 14 Mon 15 Tue 16 4 inputs1 inputs1 inputs3 inputs2 inputs3 inputs2 inputs1 inputs1 inputs1 inputs7 inputs7 inputs2 inputs10 inputs5 inputs7 inputs3 inputs1 inputs1 inputs4 inputs2 inputs4 inputs2 inputs10 inputs1 inputs1 inputs2 inputs1 inputs1 inputs4 inputs2 inputs2 inputs4 inputs1 inputs7 inputs11 inputs3 inputs3 inputs8 inputs3 inputs9 inputs3 inputs
Inputs per hour (local time, CEST). Source: session transcripts. Machine responses and tool steps are not counted. Peak: 11 inputs in a single hour (15 June).

From building to growing

The first version was a build with a clear end: spoken notes became a site. The second has no such end. A site with a blog, newsletter, comments and two languages is never “finished” — it gets a new post, a correction, another language. So the proof is no longer the fast first shot, but that the same close guidance also holds over days and over many small changes, without the site fraying.

From a single page to a system

At first it was static pages. Then, piece by piece, came: a blog built from text files; a standalone service for moderated comments, running separately from the site; a newsletter with double opt-in; and a full English edition — machine-translated, but as its own, maintained branch with English addresses. Each building block is small on its own. Together they're the difference between a business card and a system you operate. How these parts work together under the hood I'd rather show on a concrete project than here — the point of this case study is a different one: that closely guided AI work scales from a single page to several interlocking building blocks, with the same discipline.

What becomes visible at the breaking points

You learn the most where something doesn't work on the first try. An example from this week: the newsletter sign-up suddenly stopped working — and not in my code, but deep inside a provider's third-party JavaScript. The machine could pin down and name the error. The actual decision, though — patch the foreign building block or replace it with a lean solution of my own — was a human one. That's the pattern of the whole site: the machine supplies speed and diligence; at the forks, the human decides. The exact course of that repair I'll save for a post of its own; what counts here is the lesson — the value shows not when everything runs smoothly, but at the fork.

What this shows

RAMP shows what's possible autonomously. The first version of this site showed the other end of the scale: closely guided pair engineering on a small, fast project. This second version shows what matters most in between — duration. An AI-built site doesn't just have to come into being once; it has to grow, correct itself, live in two languages. That this works with the same discipline as the first build, only over more days and more breaking points, is the real message of Version 2.