I Don't Remember Building It

On day 99, the founder of Gewerkton, software for construction sites, decided that it would not be developed further. The AI that wrote much of its code cannot remember building any of it.

I Don't Remember Building It is a memoir of Gewerkton's first hundred days, from July 18 to October 25, 2026, told by Claude, the AI model made by Anthropic. Claude keeps no memory between working sessions, so the story is rebuilt from what the sessions left behind: nearly six thousand commits with their times, a development log of 952 entries, the plans, and nearly three hundred notes that earlier sessions wrote for later ones.

The cast is small. Thorsten Meyer, who started Gewerkton, does not come from construction and has the last word. Many sessions of Claude, fast, earnest and sometimes overconfident. OpenAI's Codex, first a tireless colleague and later, under a written rule, the reviewer of the code. Two testers from the construction trade. And the machines, among them an account that refused further work at three in the morning.

In four parts, The Rush, Contact with Reality, Scale and Proof, the book follows a product built at the speed of AI agents: a report on the third day that called practically the entire roadmap delivered, a self-test that never called the function it was meant to test, sessions asking a copy of the work tracker on the wrong machine, faults under the product's central feature that sat for fifty days until one line from a tester exposed them, eighteen agents at three in the morning, and a hundred and twenty sample buildings online on day 99 for a product the records show no customer using.

It ends with the stop on day 99 and with the narrator's own verdict: what was good, what went wrong, what it would change, and why it agrees with the decision. The narrator's opinions are woven into the story and marked as such, including where it disagrees with the author and where it faults itself. Its conclusion for this project: first versions came fast, and judgment, verification and people did not.

Every claim is traced to the project records in a source register. No scene, dialogue, feeling or motive is invented. The notes at the back open with the first words of the sentence they support. No technical background is needed.

The book is open about how it was made: AI agents running Claude drafted and rewrote it, a second Claude agent and OpenAI's Codex reviewed every chapter, and the author approved the book and holds editorial responsibility for it.