Hi there,

Look at me, back in your inbox after about a month instead of two. Last time I admitted the gap had embarrassed me a little, and I made a quiet promise to myself to keep this thing closer to monthly. One edition isn't a streak, but it's a start. Steter Tropfen höhlt den Stein — constant drops wear through stone.

This one is the edition I teased at the end of the last: the hands-on, built-it-myself stuff I'd been sitting on. The headline is that part of this website is now run by an AI assistant I built. So let me start there.

🤖 Meet Clawdia

Everyone keeps saying "AI agents." For months it felt like a word that was everywhere and meant nothing, so I did the only thing that ever helps me really understand something: I built one (and broke it a few times).

An agent is the step past a chatbot. A chatbot answers; an agent acts. It doesn't tell you how to send the email, it sends it. Mine is called Clawdia (I have never been good at naming things), she lives in WhatsApp, and she runs on an open-source project called OpenClaw. I can send her a voice note in Swiss German — the dialect, not even proper German — and she transcribes it, works out what I meant, and does it. Most relevant for you: I can describe a change to this site, she makes it, builds a preview, and sends me a link. I reply "looks good," and it's live. Usually from the tram.

I'll be honest about the rough parts, because the demos never are. This is not a polished product you install and enjoy. I needed a small rented server, I configured parts of it by hand, and at one point I broke it so badly the only way back was restoring the whole machine from a backup. I came close to giving up. But the friction was the point — you can't form an honest opinion about agents from the sidelines, same as you can't learn to cook from videos. Four months later she's still running, now more of a general assistant than a website robot, for under $60 a month all in.

If you build things or run a business, this is the most tangible glimpse of where AI is heading that I've found. The full story is here — including the part about what you're really signing up for when you hand an autonomous system access to your files and your inbox, and how I keep that safe.

🗺️ The other thing I built: a weekend with its own app

The smaller build I'd promised you. I gifted my wife a weekend in Hamburg, planned the whole thing in Perplexity, and ended up with a pile of notes that would have been miserable to use on the trip. So I had an AI build me a little mobile map app instead — our hotel as home base, every reservation pinned with its time, the hop-on-hop-off bus route traced on top. Twenty minutes to the first working version.

The interesting half is where it stopped being magic. Getting to "works for me on this trip" was an afternoon. Pins were off by 200 metres, the mobile view was broken, links opened the wrong way — each fix another round with the agent, each round burning credits. And turning my weekend toy into something a real bus operator could ship would have meant translations, payments, security, maintenance forever. That's the line: vibe-code it when the user count is small and the lifetime is short; everything past that is still the full discipline of building software. The whole story is here.

🎨 A fresh coat of paint — and a word on what's actually mine

You'll notice the site looks different again. I regenerated almost all the artwork — nearly every guide and category image is new, most of it made with GPT-Image-2. It's a much more coherent look, and yes, the irony isn't lost on me: an AI assistant edits the site, AI writes a lot of the words, AI paints the pictures too.

So let me say the quiet part out loud, because it matters to me and it should matter to you. The machine is the fast hands. It is not the head. Every guide here comes from something I actually did — a system we built at work, an agent I accidentally broke on a weekend, a model I tested on real tasks instead of toy ones. The strategy, the opinions, the analogies, the calls about what's worth your time and what's hype: those are mine, and I sweat them. I'm pouring my own experience, taste, and judgment into this precisely so you get the filtered signal and not the marketing exhaust. AI lets me ship it faster. It doesn't get to decide what's true. And in my humble opinion, that’s exactly how AI should be used.

🎓 The fluency path is a real thing now — and I need your eyes on it

In the first email I shared the AI Fluency Levels. In the second I told you I'd rethought them. Since then I've actually built the rest of the system out, and it now lives in one place: the AI Fluency Path. Take the quiz to find your level, read what each level means, then work a hands-on checklist (Levels 1 through 3 are live) that turns "get better at AI" into specific things you can tick off.

The part I'm most curious about is the certificates. There are two you can earn. The AI Explorer covers the first two levels — ten questions, seven to pass. The AI Practitioner is properly harder: fourteen scenario questions that test judgment, not recall — which model for which job, what to ask before you build, how to spot an AI that's just flattering you. It's open-book, it runs entirely in your browser (your name never leaves your device), and you can retake it as often as you like. Pass and you can download a real certificate or add the credential straight to your LinkedIn profile.

My actual ask, and it's a genuine one: go earn a certificate, then hit reply and tell me what you think. Were the questions fair or annoying? Did the level you landed on match how you'd describe yourself? Did anything feel too easy or unfairly hard? This is the newest, least-tested part of the site, and the only way I'll get it right is hearing from people who actually took it. I read and answer every reply.

🧰 Also new in the directory

A few tools earned their place since we last spoke. There's a new AI Agents category now (OpenClaw, the thing Clawdia runs on, lives there). I also split AI models out into their own structured Models database — 24 of them so far, with context windows, pricing, speed, and whether they're open or closed, so you can compare like for like instead of guessing.

One subtraction worth flagging: OpenAI discontinued Sora in April, so both the image and video entries are marked dead. Things move fast, and a directory that pretends otherwise isn't worth much.

Next time I want to get into the stuff that makes the AI you already pay for work harder — how the context window quietly wrecks your results, why model choice is really a cost decision, and different ways to give AI a permanent memory so it stops starting from zero every session. In a month, ideally.

As always, if any of this is useful, that's the whole reason I do it. And if you take a certificate, I really do want to hear how it went — just reply.

Best,
Fabian