Hi there,

Not even three weeks since the last edition. A warm welcome to all new readers! Shorter email this time — mostly because this edition is less "here are five new guides" and more "here's what actually changed in how I work." A few of you have told me the real struggle stories are the part you read first, so I'm making that a regular section. It starts now.

🧰 What I'm actually doing with AI right now

I cancelled my ChatGPT subscription a few months back and said so in my review. Claude was dominating. I've now re-subscribed, twice (privately and professionally). GPT-5.6 is a really good release, and it has become my workhorse for exactly the tasks that fill most of a workday — researching, translating, coding, designing, grinding through a pile of well-defined work. OpenAI's desktop app has become a pleasure to use, too. What surprised me most was the generous usage quota on the subscription; it’s very different from Claude’s. There’s not even a 5h limit anymore.

However, none of that changed who's king. When I need really good thinking, really good specification, or design work, I still go to Claude's Fable. The value I get out of a Premium subscription (via my startup) is absolutely mindboggling. Those are the tasks where a wrong answer costs me several hours, and the right answer saves me days. The way I'd put it: Fable is who I ask what to build, GPT-5.6 is who I hand the build list to.

And there's still a third seat. Google updated Antigravity, and the new desktop app looks almost exactly like the ChatGPT and Claude apps — same clean chat-first surface, agents underneath. Its model, Gemini 3.7 Flash, lags the frontier by a fair margin on hard problems. But it's fast, it performs well on simple-to-medium tasks, the usage quotas are very generous, and I enjoy using it to experiment with websites. More on what else it did for me below.

So: three desktop apps, three roles. A year ago I would have called that subscription hoarding. Now each one earns its seat, and the skill is mostly knowing which task goes where.

📝 Don't let your AI grade its own homework

Running multiple AI subscriptions raises an obvious question: can they check each other's work? That's the guide I published right after the last edition went out: Don't Let Your AI Grade Its Own Homework.

The short version: a model reviewing its own output tends to flatter itself. There's a good paper behind this (it's cited in the guide), and it matches what you've probably seen: ask ChatGPT to critique the text ChatGPT just wrote and it finds two adjectives to switch and declares the structure excellent. A second model is a cold reader with no ego in the game. But the guide is also honest about the limits: the two models often make correlated errors, so the second opinion is worth less than it feels, and there are setups where the whole exercise is just theatre. There's a version in there that works with fresh contexts on a single sanctioned tool, for everyone whose company only allows one.

Given the stack I just described, this loop stopped being theoretical for me. Fable writes the spec, GPT-5.6 executes, and each one reads the other's work cold. It catches real mistakes. And for me it’s becoming really valuable to know the characteristics of each model well.

🔧 The AI on my desktop fixed my Wi-Fi

Something happened this month that I keep telling people about. My own computer had a flaky network connection and a power-management problem I'd been ignoring for weeks - it would awake from its sleep in the middle of the night. It's the kind of thing where you'd normally research, struggle with Windows settings and give up. I let Antigravity work directly on the computer instead. Controlled computer access. It inspected the network configuration, found the problem, walked me through what it wanted to change, and fixed it. Then it fixed the power issue too.

The leading AI desktop apps have all become good at this kind of computer use; Antigravity is just the one that happened to fix mine. They can install and configure software, run system maintenance, analyze what's actually happening on your machine, and explain it in plain language while they go. For a lot of everyday cases that covers the usually required IT support.

Two things before you try this at home. First: do it on your own computer, like I did. Your company laptop belongs to your IT department, and an AI changing its settings is exactly the kind of surprise that gets tools banned. This is private-track learning, like everything else I recommend doing on your own machine. Second: make the AI explain what it wants to change before you let it. The point is a competent helper you supervise; the skill of reviewing before approving is the same one you use everywhere else with AI.

I'm considering writing a proper guide on this — using the AI on your desktop as your personal IT department, safely. If you'd read that, reply and tell me what breaks on your machine. Some of the best material on this site came out of replies to these emails.

📖 The thing I've been quietly writing

One more piece of news: I've started writing a book. Nobody outside this list knows yet.

It grew out of this site, but it's a different animal. The site is a reference: you come with a question, you leave with an answer. The book is one connected path instead of forty separate guides — the mental models, the practice, the way AI changes how a company works. A lot of it is new material that isn't on the site and won't be. If you've read my guides, you can guess the central character: the brilliant new hire who knows everything except your business gets the book-length treatment.

There's no launch date and I'm not selling anything today. I'm telling you because this list is small and friendly, and I'd like to write it together with you. If the idea speaks to you, reply and tell me: would you want to be an early reader? Is there a chapter you'd hope is in there? Those replies will shape what I write.

📇 Meanwhile, on the site

The model and tool catalog caught up with the month: new entries for GPT-5.6 (the Sol/Terra/Luna family explained in plain language) and Gemini 3.7 Flash, plus refreshed reviews of ChatGPT and Google Antigravity that say everything above in more detail, including why ChatGPT's rating went up. All of it in English, German, and French, same as always.

That's it. Shorter, like I promised. If any of this edition made you want to hit reply (the stack, the IT-support experiment, the book), please do. I read everything.

Best,
Fabian