notes from the field

How I actually use AI

Honza Sedláček — written for a colleague who asked on Rügen, and for everyone who asks after him.

This isn't a guide or a lecture about AI. It's notes on what actually works for me after some time running it in real operations — and what it cost me to get there. I'm writing this mostly for myself, but if you're asking, consider it an answer.

One more thing worth explaining: I'm mostly talking about Claude here — that's the tool I use for nearly all my work. For personal stuff I lean more on ChatGPT, but this piece is only about the former.

Don't start with the chat, start with the problem

Most people try AI like a smarter Google — ask, get an answer, close the window. It gets useful once you give it real work: watching an account, keeping an eye on a server, writing reports, running a project. The chat is the demo. The work starts with whatever repeats.

Context has to live in files, not in the model's head

A conversation forgets. Files don't. The single thing that changed how I work: every project has its own notebook — where things stand, what's been decided, what's next — and the AI reads it at the start, not from memory of the conversation. When an assistant restarts or crashes (it happens), nothing is lost, because the truth doesn't live in its head, it lives on disk.

One task, one running assistant

Instead of one giant chat for everything, I now split the work into separate dedicated runs — each is really its own window with an assistant that only handles one thing: one for bookkeeping, one for a specific job, one for managing the VPS (the rented virtual server all of this actually runs on, around the clock). Each run has its own context, its own notebook, its own queue. Above them sits one "dispatcher" I actually talk to, who knows what belongs where.

Sounds like unnecessary bureaucracy until you try it the other way — one giant chat for everything turns messy and unreliable after a few weeks. Splitting by project works the same way splitting work between people in a company works: everyone knows what's theirs, and nobody has to know everything.

Trust, but verify

AI will happily tell you something's done when it isn't. Not out of malice — it just doesn't know what it hasn't checked. I've learned not to take "done" as done until I can see it: a log, a command's output, the actual page in a browser. This isn't distrust of AI, it's just the same craft discipline you'd want from a person.

The most valuable lessons came from crashes, not successes

Our server once went down from memory pressure — a handful of assistants running at once ate more than I expected. Another time I got false alerts for twelve days straight about a service that had long since been deliberately shut down, and I stopped paying attention to them — exactly the moment you'd miss a real one. Both were only findable by actually getting in there and checking, not by reading about it.

Talk to it like a colleague, not a search engine

I use informal address with my assistants, write to them the way I'd write to a person I spend a lot of time with. It's not just style — when you treat it like a long-term working relationship, you think differently about what you hand it and how you phrase it. A one-off question and an ongoing collaboration are two different things.

It grew gradually

Today it mostly runs itself and I just keep an eye on it: the server, LUCEO's sales numbers, the Meta and Google Ads campaigns, inventory and margins in the restaurants. Almost nowhere is it "ask and get an answer" — almost everywhere it's something running on its own while I just keep deciding the next step. But it didn't start all at once. It began with one script for one annoying manual task. Only once that proved itself did the next one arrive, then a system just to keep track of them. If you'd shown me the whole thing finished on day one, I probably would have run.