Hi everyone,
Last week I promised you a number. How many of my nine trial agents passed the Monday review. How many went back a step.
I did not do the review. Monday came and went. The row is still open on my board.
So no split this week. What I have instead is the reason it slipped, and it is a better story.
A hard drive that sat for two years
I run a small music event series called Elektronik Serpme. Four gatherings in 2024. Every one was filmed.
All of it went onto one hard drive. About 1.5 terabytes of raw video, plus the sound recordings of every DJ set. Then nothing happened. For two years.
Turning that into anything needed an editor for weeks. Someone to watch every clip and sort it. Cut it, match it to the beat, cut it again. I never hired that person. So the drive sat.
GPT-6 Astra, OpenAI's new model, came out on September 3. On Friday night I plugged the drive into it.
What I asked for, and what came back
My first ask was small. Label the files. Rename them. Put them in clean folders.
It did that. Then I asked one lazy question: what can we do with this?
Here is what came back, by my count. Twenty-four vertical clips for Instagram. Four horizontal films, one per 2024 event, about a minute each. One 90-second film that mixes all four events. A page with every file ready to download. Captions written. An Instagram post drafted, with a how-to for posting it.
Then I asked if it could go on the website. It rebuilt the site around the films. You can watch them here. One main film, four gatherings, three short moments.
Start to finish, about 45 minutes to an hour. I mostly watched.
Other people are seeing the same thing
Two public examples from the same week line up with mine.
Nate Herk gave it one prompt: take me from idea to finished YouTube video. It researched, scripted, voiced, edited and rendered the whole thing. About 50 minutes, an estimated $60 in API cost. The edit ran inside HyperFrames, the same tool I build our own videos with.
Then the other side. One developer spends around $330,000 a month on AI inference. He reported two wins on the t3.gg podcast. It built walkable 3D scenes. It cut his platform's latency from 800 milliseconds to under 30. The same model made YouTube Shorts he called "cringe and bad." It filed a pull request, then said yes to the review fixes and did nothing. He had to keep prompting it to finish.
Same model. Same week. One person gets a finished video, another gets a half-done fix.
Same model, same week. The difference was one line in the brief: what to do when it stops.
Why this makes the brief matter more
Here is my take. None of this surprised me. The direction was obvious. What changed is the length of the leash.
A year ago an agent ran for a minute and handed back a draft. Now it runs for most of an hour on its own. Mine did. Nate Herk's did. That is the new normal, and it changes what you need to write down first.
Last week I gave you the five parts of a one-page brief. Result. Sources. Limits. Evidence. Exceptions.
Run my Friday night through them. Result was vague. "What can we do with this?" It still worked, and here is why. The limits were built in. Private footage, a drive I own, nothing sent anywhere, nothing deleted. When the downside is zero, a vague result is fine. That was play.
Now run the pull request through them. The model said yes to the fix, then did nothing. It never said so. That is a missing exception rule. Nobody wrote down what "stop and tell me" looks like, so it just stopped.
The longer the leash, the more that one line matters. A model that works for an hour unattended will fail for an hour unattended too.
One thing to try this week
Find the drive. Every business has one. A folder of recordings, a backlog of screenshots, a year of support tickets nobody tagged. Work that never got done because it needed a specialist for weeks.
Give it to an agent with two limits written down. What it may touch, and what it may never send or delete. Then ask the lazy question. What can we do with this?
Watch how far it gets before it needs you. That gap is the brief you have not written yet.
Save this one for the next folder nobody has opened in a year.
About that review
The nine-agent pass is still owed. I am not going to promise a date again, since I just missed one. When it is done, the split goes here, agent by agent.
What is sitting on a drive in your business right now? Something that waited for a specialist who never got hired. Hit reply and tell me. I read every reply, and our conversations help shape the newsletter.
Thanks for reading.
Tim