I swapped Claude for ChatGPT, in a single evening

On a Wednesday evening I set my regular AI assistant aside and let its biggest competitor take over the work. Claude out, ChatGPT in. Same tasks, same records, same rules. The switch took one evening. After that, the stand-in ran my system for more than a day while I mostly watched.

The result: it fully completed two open tasks, paperwork and all, and built an entire measurement setup for a third. It stuck to the house rules, looked things up before claiming them, and on its way out wrote a handover a new colleague could actually work from.

Why you want to test this once

Everything I do with AI hung on one vendor until that evening. That's fine as long as the vendor exists, stays affordable and doesn't squeeze you. But my monthly allowance regularly runs out before the month does. Prices change. Services shut down. And the more work you hang on such an assistant, the more painful the question: what if it's gone tomorrow?

That's not a question you want to ask for the first time on the day it happens. So I rehearsed it, the way you run a fire drill. First a test run on a copy, then the real handover that same evening.

Why one evening was enough

The swap went that fast not because the replacement model was so good. It went fast because there was little to move. Two years of AI conversations sit as ordinary files on my own drive. My tasks live on a board any assistant can read. And the working method, from "how do you close a task" to "what must you never throw away", is written out in plain language, like an induction manual for a first day on the job.

The AI isn't the system here. The AI is the operator who comes in to run the system. Swap the operator, and the new one simply reads the same manual.

What was honestly different

It wasn't a photo finish. The stand-in worked noticeably slower. And its language was more businesslike, more distant, less pleasant to work with. Maybe that's the model, maybe it's the instructions I handed over; I still have to find that out. But the work was sound, and that's what the test was about.

One thing didn't change: the last button stayed mine. Never more than one assistant works on the same records at a time, and nothing goes live without me having seen it. Precisely when you swap operators, you don't want to lose that human gate.

The next dependency I've now tested too

Am I independent now? Not yet. My task board runs on a cloud service, and that's just the next version of the same risk. When I wrote this piece, the exit plan for it was still a task on that very board. By now I've carried it out.

I now make a full copy of that board to my own machine every day: all the tasks, the complete history of changes, and the structure of my projects. And just like with the swap test, I didn't stop at a copy. I built a test that pretends the cloud service is gone and tries to rebuild my entire administration from that copy alone.

That test failed the first time. And that was exactly the point. It flagged a gap I'd never have spotted otherwise: one layer of my projects was in the app, but not in the daily copy. A test that always passes proves nothing. This one was allowed to fail, and it did. I closed the gap and ran it again. Now everything comes back.

That keeps it honest: every time I solve one dependency, I point at the next, and I test whether the fix actually holds.