My secretary doesn't sleep, but it does get yelled at

I run a one-person company. For a long time that meant one person did everything, and the AI was a very smart stranger I met in a browser tab.

That stranger had a problem. Every morning it was a new hire.

Stage one: I wanted someone who knew me

Chatting with Claude on the web is like hiring a brilliant temp every day. Brilliant, and completely new. So every day I explained who I am, which apps are mine, where the ad revenue report lives, and that I hate long dashes in my writing.

At some point I stopped wanting a temp and started wanting a secretary. Someone who already knows.

So I gave it a desk. One long-running session inside my portfolio project, with its own notes about me. It knows the AdMob report is a local file I refresh by asking. It knows to talk to me like a friend, not a bank. It knows that when I need a config file on another machine, it should send me the file, not a paragraph telling me to write one.

It learned that last one the way secretaries learn things: I told it off.

My favorite moment from this era is when I asked it to move my whole dev folder to an external drive. It moved everything, fixed the paths, then moved its own memory into the new place and told me to close the terminal and reopen it over there so it could keep working. My secretary packed the office, then packed itself, then handed me the new address.

Stage two: I wanted to reach it from anywhere

A secretary who only works when you are sitting at the desk is really a desk phone.

So I connected the session to Telegram. My Mac stays on, the session stays open, and I talk to it from my phone.

The first thing I did was not useful. I locked the Mac, turned the screen off, walked away, and texted: can you still hear me? It could. I felt like a manager calling an employee at midnight just to see if they pick up.

Since then the texts have been less of a test and more of a lifestyle.

I texted "fetch AdMob" from the couch and got the numbers back with a note about a drop in match rate.

After midnight I asked it to replace the paid app I use for Windows games with a free one. It installed the free one, got Steam running, and when my game refused to start it found the graphics setting that fixed it. Then I told it to delete the paid app to free up space, so it did.

I cut the power to my own computer by accident. When it came back, I did not check my files first. I texted the secretary to ask if it was still alive.

It also gets yelled at.

Early on it sent me a set of shell commands glued together in one paragraph. I could not tell where one ended and the next began, and I told it so, with a word I will not print. It now has a rule, written in its own notes, that every command goes in its own box.

py -3 -c "import sys; print(sys.executable)"

npm.cmd install -g @anthropic-ai/claude-code

claude --version

Another night I asked it to try making a 3D model of a small aquarium catfish. It generated models, rigged them, rendered swimming videos, and filled a folder with scripts. I looked at the result and replied with exactly two words: "gross, clean up." It cleaned up. Every secretary has had a boss like me.

Stage three: I want it to manage everyone else

Every project I have now comes with its own agent. The game has one. The cashback app has one. The prototype I abandoned a couple of weeks ago has one, still remembering exactly why I abandoned it.

The problem is me. I do not want to open five terminals. I want to text one sentence, "do this in that project", and have my secretary walk over, hand out the work, and come back with the answer.

So we tried it. My secretary sent a message to the agent working in another terminal: reply with "pong". Shortly after, it got back "pong".

That was the first time two of my employees talked to each other without me in the room. I felt proud, then slightly left out.

It can also open an old session from a project I have not touched in almost two weeks and ask it what we were last doing. It answered correctly: an experiment that failed, and me telling it to stop everything.

Here is where I am now. I used to be one person doing all the work. Now I am one person who manages a secretary, who manages a team of agents, who do the work. Still a one-person company. I just added a layer of middle management.

The part that is not a joke

The further the work gets from me, the less I see it happen. When my secretary says "done", it is reporting what another agent told it, which is based on what that agent believes it did.

I have not handed it a real multi-hour job yet. I will, and I will check halfway through, because "done" and "correct" are different words.

Last night

Last night it told me my match rate dropped because of a bug in my code. I said I did not think so: if it were code, it would have been broken all along, not suddenly. It went back, measured again, and came back with: you are right, I was wrong, here is what actually changed.

That is the whole job description. Not a secretary who never gets it wrong. A secretary who, when yelled at, goes and checks.