Hi there,

Two weeks since the last one. I know — who is this person and what has he done with Fabian. Don't get used to it; something shipped that I couldn't sit on for a month.

I built a game. An actual browser game, free, no signup, and it exists because of a problem I kept running into: I can explain agentic AI to someone for an hour, they nod along, and a week later it's gone. Reading about delegation doesn't stick. Losing to a corporate giant because you skipped onboarding your AI hire — that sticks.

🎮 Welcome to Brilliant & Co.

If you've read anything on the site, you know my favorite mental model: AI is a brilliant new hire who knows everything in general and nothing about your business. The Company Simulator makes you live it. You found a small company, you hire coworkers who are exactly that kind of brilliant-but-clueless, and you find out the hard way what it takes to turn raw talent into work you'd actually ship.

Your rival is Monolith Group, a corporate giant that starts miles ahead of you and responds to the AI wave the way most real companies do: by buying thousands of licenses and changing nothing else. They win contracts you're not ready for. It's genuinely annoying. It's supposed to be.

The moment I most wanted to get right is the dip. When you onboard your first hire, your capacity curve goes down. Every hour you spend explaining the job, clarifying, reviewing, and fixing weak drafts is an hour of your own work gone, and the game counts all of it. For a while you're doing worse than if you'd just done everything yourself — which is exactly where most people quit with AI in real life, convinced it doesn't work. Push through, build the systems, and the curve crosses break-even and doesn't look back. Skip the onboarding and you get fast, confident, wrong output forever. I've watched both versions happen in real companies. Now you can speedrun them in an afternoon.

A few things worth knowing: there's no AI behind it — it's a deterministic simulation, so your losses are entirely your own fault. It runs in your browser and saves locally. And the lessons that pop up along the way form a six-chapter playbook covering the same delegation concepts I use at work, minus the slides. Fair warning: it's English-only for now, and beating Monolith takes longer than you think.

My ask: play it, then reply and tell me where you gave up or how you won. I'm still balancing it, and the fastest way to get it right is hearing where real players hit a wall.

🌍 The site now speaks three languages (I still speak two and a half)

Good AI Guide is now fully available in German — every guide, every tool review, every model entry, the quiz, even the certificates (you can have your KI-Praktiker document in German now). And a French edition just launched with the first guides translated and the rest following.

I didn't translate a word of it, and I didn't hire translators either. I wrote a one-page rulebook (Swiss High German, no ß, keep my analogies, don't soften my opinions) and managed a swarm of AI sub-agents that did the actual translating. Twenty-five guides, roughly 54,000 words of German, produced in parallel while I reviewed samples and corrected course. My job wasn't translation. It was quality control and taste.

That experience taught me more about managing AI at scale than anything else this year, including the unglamorous discovery that translation is cheap and upkeep is forever: every time I edit an English guide, I now owe that edit to two other languages. I wrote the whole story up, including the numbers, the model comparison, and what the AI got wrong.

🧠 The guides I promised you last time

Last edition I said the next one would be about getting more out of the AI you already pay for. The game jumped the queue, but I keep my promises. Three guides, one sentence of honesty each:

What the context window is, and what to do about it — if your AI gets dumber the longer you talk to it, that's not your imagination, and five small habits fix most of it.

Think expensive, execute cheap — model choice is really a money decision, and the trick is using the expensive model to plan and the cheap one to grind.

The setup that makes AI stop being generic — an hour of setting up a permanent profile beats a year of re-explaining who you are in every chat.

And two newer ones for readers who are past the basics: why I stopped talking to one Claude and started spawning a team (the same trick behind the translation project above), and how to give an AI a finish line instead of prompting every step — including where that bites you.

📇 Small but worth knowing

The models database now carries Claude Sonnet 5, which has closed most of the gap to the top-tier models on everyday work — for most people it's the sensible default now (my take: effort to medium. Higher and Opus is the better choice). I’ve also written a love letter of sorts) for Fable as it’s just such a joy to work with this model. And a fun one from the follow list: Andrej Karpathy, whose "learn by building" philosophy I quote constantly, joined Anthropic. I take that as validation of my Claude bias and nobody can stop me.

That's it for this one. Shorter gap, shorter email — I suspect that's the sustainable shape of this thing.

Go play the game. And when Monolith Group is still miles ahead of you a few days in, remember: you could have onboarded that hire properly. Reply and tell me how it went — I read everything.

Best,
Fabian