A day building Vidi.

The prompt isn't the difference.

— MSW/Noor and Machine · 10 October 2026

Vidi is my video agent. It makes the videos for my history channel.

I've paused the first one. The script is fine. Not final.

So instead of fixing it line by line, I spent the day building the thing that makes it.

Here's what I did, in order.

Morning

I asked what it could do.

I asked Vidi to tell me what it had learned so far. Short. No fluff.

Voice rules. A script playbook. A study of the first two minutes. A study of titles and thumbnails. It listed them and suggested the next three things to build.

One of them was a loop that re-records any line the voice gets wrong. I asked the dumb question: if you change the seed, doesn't the voice change?

It doesn't. The seed is the take. The voice is the voice. Good to know.

Then I asked what the loop would cost. About three extra renders a video. A few dollars. Fine.

Midday

I made it clean up after itself.

The video project had taught us things that never made it back into Vidi. Twenty-nine of them. And nine places where Vidi now said the wrong thing.

It said use quiet voice tags. We'd found those tags make the narrator whisper. It said one speaking rate. The strategy said another.

I asked it to explain, in plain words. Then I made two calls. Stability 0.5. 150 to 185 words a minute.

It moved everything into the right place. Nineteen files.

Then I had a thought. These rules fit this channel. A dramatic channel would need quiet. So I asked for rules I can switch on and off per project. It wrote the idea down as a decision for later. Didn't build it. Right call.

Afternoon

I asked for the angle study.

Same method as the others. Research first. Then the niche. Then what the YouTube people say. Then our landscape: 10,740 titles. Then the 48 videos we already had transcripts for, coded blind by six agents who never saw the view counts.

A table of angle features against over-performance: the title carrying the angle fully has the biggest effect — One result stood out. When the title and the video tell the same story, the video wins.

Then a seventh agent coded my own script on the same sheet.

My script coded on the same spec, next to what the 48 videos do — Mine has the title right. And stacks four of the weak choices.

Now I know what the rewrite is. I didn't know that this morning.

The point

Capability, not prompts.

None of this was a clever prompt. I said what I wanted, plainly, the way I'd say it to a person.

The difference between a good result and a poor one is whether the agent has the capability to do the job. Not the wording.

So I build the capability. Find the best of the best. See what they actually do. Take what fits me. Build it so it does that better than they do.

Then the agent only has a few jobs, in a workflow, and it does them well.

Tomorrow: the voice loop.

← blog