How it started, how it’s going

For a long time, the best tool I had for automating things was Google Apps Script. I could write it — but not quickly, not cleanly, and only as far as my understanding reached.

So when AI tools arrived, I started using them to write the scripts for me. I described what I needed — pull data from this sheet, format it, send a summary email — and ChatGPT wrote the script. I copied it, pasted it into the editor, ran it, hit an error, went back, fixed it, ran it again.

It worked. I felt clever.

What I did not notice at the time: I was still the one doing things. The AI wrote the code. I executed it. The loop ran through me.

That was the shape of AI assistance for a while: a shortcut that still ran through me. Every output from the AI was an input for me.

I kept using it this way for months. I got faster at generating the scripts I wanted. The execution model stayed the same.

The shift happened when I moved to Claude Code.

Not immediately. At first I used it the same way — describe what I need, get the script, run it. But Claude Code could also run the script itself.

When the AI can run what it writes, you are no longer in the chain. You describe the task. The work happens. You get the outcome.

I did not fully understand this until I started asking for more — and watching things just get done.

A message sent. A report filed. A booking briefed. A form filled. A follow-up sent.

Not given. Done.

That distinction — given vs done — is where everything changed.

Before: AI as a better way to write things I still had to do.

Now: AI as a system that does them.

Where it is now is harder to describe concisely because it has become infrastructure.

Automations run every morning. Notifications arrive on my phone before I’ve opened a laptop. Workflows I set up weeks ago have not needed my attention since.

My EA gets a morning brief every day — what she needs, formatted, ready to act on. I wrote that automation once. It has run every day since. I have not thought about it.

I am not doing less work. I am doing the work that requires judgment, not the work that requires execution. The line between those two used to blur. Now it is sharp.

Between the early days and now, AI has changed. But what really needed to change was the way I was using it.

I was treating AI as a tool — give it a task, take back the output, go do something with it. The way you use a calculator. Useful, but passive.

Claude Code is not passive. Or more precisely: it is as passive as you make it. If you ask it to write something and do nothing else, you have a calculator. If you ask it to write something, run it, handle what happens, and report back — you have an agent.

The difference is what you ask for.

I started by asking AI to help me write things I was still going to run.

Now I ask it to run them.

It works. I feel powerful.