I was writing book reports to AI

At some point, I realized I was spending more time explaining edits than spotting them.

AI would hand me a landing page, a campaign, an app feature, or the beginning of a system. I could tell what felt wrong almost immediately. Then I would sit there for another 15 minutes turning that reaction into something neat enough to send back.

Section one needs a better opening. Section two is saying the same thing twice. The button works, but the promise around it does not. Actually, scratch that last note - I just found the explanation farther down the page.

By the end, I had written a little book report to my AI.

That was backwards.

I started keeping the artifact in front of me and talking through it instead.

Then the next section. Then the interaction I had missed. Then the correction to my own correction.

The feedback was not tidy. It was much closer to how I actually think while I am reviewing something. The next draft was usually better because the system got the reaction while I was still inside the work, not after I had cleaned the life out of it.

That is the first place voice became genuinely useful for me. Not as a faster keyboard. As a way to keep my eyes on the thing I was judging.

The next shift was weirder.

The voice conversation stopped being one place where I dictated a prompt. It started becoming the place where I could direct the work around me.

I could state the outcome, let a task run, look at what came back, and talk through the next move. The system could reach coding tasks, research, browsers, local files, and other tools without making me act as the copy-and-paste department between all of them.

In the raw recording I called it my Jarvis. That is generous. It still gets confused, interrupts at the wrong time, and occasionally needs the technological equivalent of being pointed toward the door it is already standing beside.

But "voice assistant" now feels too small.

The useful part is not that I can talk to a computer. It is that the conversation can coordinate real work while I stay focused on the decisions only I can make.

There is one part I do not want voice to blur.

Talking through an edit is direction. It is not permission to send an email, spend money, schedule a post, or publish something with my name on it.

My working version is simple:

That last step can feel slower than the rest of the loop. Good. It is supposed to.

If the copy changes, the image changes, the destination changes, or the timing changes after approval, it is a different package. The old yes should not float over to the new thing just because both versions have the same title.

Speed is helpful. Ambient permission is not.

This has also changed how I think about documents.

For a long time, the document had to do two jobs. It held the full context, and it was the thing a person had to read in order to understand that context.

AI can separate those jobs.

The long record can stay detailed because the system can work with it. The person reviewing the work may need a one-page summary, a diagram, a short video, or a conversation where they can ask questions in their own order.

The source does not disappear. The interface gets closer to the way the person actually understands the work.

I am still figuring out where that idea goes. The immediate change is more practical.

When I can see what needs to change, I no longer assume the responsible thing is to sit quietly and type a polished memo about it.

Sometimes I leave the work on the screen and start talking.

That has been a much better use of both of our time.