Florian Thompson/blog

I always wanted a blog. The keyboard was the problem.

This is for me.

I’ve been talking to friends who just started dabbling with AI. I keep noticing I know a few things they don’t. Some of that I learned by doing it wrong. Some from people on X who are actually good at this.

I don’t want to type a blog. Typing was always too slow. I talk it with Wispr Flow, then format after. An exercise. Put the thoughts down so they actually sink in. Maybe something here is useful. Maybe someone smarter already wrote it better. Don’t criticize me too much. This is just a write-up.

I’ve always been a builder. Early Android apps, then iOS. What I loved was the whole chain: wouldn’t it be nice if this existed, then you build it, then someone is actually using it.

The pinnacle of my career was not the acquisition. I didn’t even have a say in that. It was seeing someone on the subway using a feature I built. That was the wow. You can come up with something, you can build it, and people use it.

AI lets you follow those thoughts much faster. It used to take months. Now it’s days or weeks. You can let the flow drain instead of sitting on it.

Same with a blog. I always wanted one. I used to do Instagram and the keyboard was the constraint. I couldn’t type as fast as my thoughts wandered. Now I just talk the post.

I see it with customers. I see it with myself. I see it with my family, who aren’t very technical, doing things that are a real help.

Nobody really knows what the future of AI is. Hard to predict. The trend I’m living, with Grok Bot, with OpenClaw, with Hermes: the UI gets less important because agents run most of it.

Yesterday, for the first time, my agent booked a doctor’s appointment. The agents are doing business paperwork. The busy work that always went to the back of the list.

We used to talk about coding faster. That’s real. But there were always a hundred other things we should have been doing. Now we can just do them. It feels like I have ten personal assistants.

The problem is keeping them performant.

I switched to Grok Bot. I like it a lot. It’s also very token-hungry. I’m pretty much maxed out on every subscription: Claude Code Max 20x, a Codex 5x plan I lean toward because the output is clearer than Opus or Sonnet, and Cursor so I can actually use Grok Bot. Before that I ran Hermes on my Hazil VPS.

I was doing everything the Twitter people talk about. Still burned through about 30% of a max token allowance per day. I upgraded. Two days later I was out of usage. I did great things with it. Then I just couldn’t use it.

Matt Pocock’s training on context is what unlocked a real system. What is a model, what’s an agent, what belongs in context and what doesn’t. A chat always gets resent. It might be cached, but when you come back it loads again. You either compact, you clear, or you pay.

Just because a model says it can take a million tokens doesn’t mean you should. Quality drops because the relations per token scale badly. I’ve heard the smart zone is about 125k to 150k. After that the output slowly gets worse. Keep context small.

So I told Grok: we can’t run everything in one chat. Here’s what changed.

Those two switches, the isolation and the doer, took me from about 30% of daily tokens to about 5%. Grok organizes the chats, the agents, the thinking. Claude Code or Codex does the actual work.

This used to be a whole project for my wife too. She wanted to throw a couple of prompts a day. She got her own Claude subscription. Then we had to sync repos. Her agent didn’t pull my changes. Skills drifted. A lot of setup for very little use.

Now she just talks to my bots. The ones I already optimized do the work. If something needs doing, it runs on my infrastructure.

I put how I have it set up here: florianthompson/grok-bot-setup. If you run agents, feed that to your bot and read it critically. Not saying it’s perfect. I’ll change it the next time I learn something. If it applies, hopefully it saves you tokens too.

I’m in the fortunate position of running my own business. I eat, sleep, and breathe this. The reason I love it goes back to the first point. This should exist. It doesn’t. Okay, let’s build it.

My old problem was ten ideas and a rule that I had to pick one. I always assumed the one I didn’t pick might have worked if I’d given it more time. I like getting an idea to life. I was never a fan of committing the next ten years to a single bet.

We live in a different time. I can do this much more freely. I love it.

Once the mass market actually feels what engineers are doing right now, I think that’s the breakthrough. I’m not trying to reinvent the wheel. I’m trying to plug into these systems. That’s where this is going.

If you have questions, hit me up on X.

← Florian Thompson