July 12, 2026

What happens when you just let AI do its thing

Translated from the German original.

At six in the evening I was standing in a fitting room trying on shirts. I had a client meeting the next morning and still needed material for it. All I really wanted was an agenda. I dictated to Claude for two or three minutes what it was about, then went back to the shirts.

What came out of it was not an agenda. It was a complete training course with full documentation, built from a workshop I had developed months earlier for an entirely different client. I barely had to touch it.

The honest part of this story is that it was neither magic nor a particularly clever prompt. The actual trick happened months earlier, on vacation.

I had time there and a hunch. If I build myself a structure once, I can work in it effortlessly afterwards and never have to tell the AI everything three and ten times over. So I created a folder that looks like my business. One area per client, plus areas for the internal topics. And what I put in there was not just notes but the real work. Presentations, workshops, decisions, rules. Told once, saved as a file, and from then on it simply applies.

Since then the system compounds. Every design rule I write down once shows up in every following presentation, even when I have long forgotten that I ever noted it. Every style, every phrasing, every decision pays in. And at some point exactly what happened in the fitting room happens. Knowledge from one project becomes, almost without friction, the finished solution for another.

The craziest part is that this training would never have been built by hand. Not because it was impossible, but because the effort would never have paid off. Rebuilding a workshop for a different context simply cannot be priced in any sensible way, so it would just never have happened. To me that is the most underrated effect. A system that compounds does not just make existing work faster. It creates work that otherwise would not exist at all.

The most common concern I hear is whether I really give the AI access to everything. No. I give it targeted access. A lot in development, where it can read, build, test and throw things away. Deliberately little in production. My systems are sealed off, and access controls decide what is possible. Creating a task, logging hours, adding a new client to the system … that works. Everything else does not. Those limits are exactly why I can be this generous with access in the first place.

Today practically my entire day to day business runs through this one system. I reference my codebases in it, issue my invoices through it, answer emails, send out surveys and have them analyzed. And the strongest piece by now is something I never planned in the beginning. I have decisions presented to me visually. Before anything is built, I see what the result will look like, not as a wall of text but as a real, designed proposal. Essentially I get a small pitch deck for every decision. I decide, the AI does the work.

If I had to start from zero today, I would do it exactly like back then on vacation, just earlier. Create a folder. Put in the work that exists anyway. And then not ask what the AI can do, but tell it what I want done, and let myself be surprised by how often it knows the way there. “I want you to read my emails and give me a report every morning. I use Outlook and have no idea how to connect it. Help me.” That is how simple the start is allowed to sound. Whoever lists the limitations up front gives the system nothing to grow in.

The lever is not the prompt. The lever is the structure the AI works in, and it becomes more valuable with every day you feed it. My most valuable working time this year was a vacation with an empty folder.