Hi everyone,
Last week I pulled 775 comments off six videos about the tool I teach every day. I wanted to see what people ask for before I record my own.
The most upvoted comment of all 775 has 430 votes. It sits under a video made by the company that builds the tool. "show instead of talking - I want to see editor, code, prompts etc."
Second, at 269 votes: "I like that this channel is a bit calmer, lots of ai channels sound too hyped up."
Further down the list, at 158: "I clicked on this video because it was literally the only one without 'This changes everything!!!'"
People are tired. Every week something changes everything, and then Thursday comes and the work is still sitting there.
So here is my week with three new things, and what each one turned out to be good for.
The model that cannot write
On September 15 a company called TypeSafe released a model called Jev. They call it a System One model, after the fast, instinctive thinking in Daniel Kahneman’s book.
It does one job. You hand it some text and a question with a fixed set of answers. It hands back a number or a label. It cannot write you a sentence, and the docs say so plainly: "System One models do not write replies, produce code, or generate explanations of their reasoning."
Their founder, Diogo Almeida, worked on the research behind ChatGPT. His reason for building this one stuck with me. "Models have been superhuman at chat for years, so where is all the automation?"
I gave it three transcripts from my own training sessions. It answered 256 questions about them. Who asked real questions. Who went quiet. What I promised out loud while teaching and then forgot about.
All three transcripts together cost under a cent, and each answer came back in less than half a second.
So here is the rule I wrote down after a day of this:
If the answer is a number or a label, send it to the judgment model. If the answer is prose, send it to a chat model. Counting stays in ordinary code, because neither model counts reliably.
The trap is that it cannot say no. Ask it to summarise something and it still answers, inside the shape you gave it, with something tidy and meaningless. I turned that rule into a check in my own code that blocks the request before it ever leaves.
It failed at plenty, too. I asked it to guess which subject lines on this newsletter would land, across ten issues. No signal at all, so I dropped that idea the same day.
The best thing it found was in my own writing. I have a habit I keep trying to kill, where I set an idea up just to knock it down. It sounds sharp and it says very little.
This week I ran the model across 162 paragraphs of my recent drafts. My plain text search for the obvious phrasing found nothing at all. The model found 27.
The top hit, at 98%, was a line from last week's newsletter. "A schedule is not a job."
Their docs also carry a page listing nine things the model is bad at, written by the people who sell it. One line reads "Jev is not a calculator". I would like a lot more of that in this industry.
Twenty minutes in the car
For a long time ChatGPT was maybe a tenth of my week, and only for personal things. Now it is about half. My hardest work goes to Codex with GPT-6 Astra, because the quality is hard to argue with.
On Monday I was driving, one AirPod in, talking to it.
Codex was running on the Mac Mini at home. OpenAI has a feature called Remote that connects the phone to that machine. Their own line for it is "start, guide, and review coding tasks from your phone", and the work runs on your own computer. Voice sits on top.
In a 20-minute drive it drafted a client agreement. A real one, from my own templates and instruction files.
I turned to my wife and said: this is almost there.
The honest part is that the drive only worked because of two years of setup. The templates, the skills, the layered instruction files, the sources it is allowed to read. Take those away and talking to a model in the car is just talking to yourself.
Anthropic shipped the same idea from its own side last week. Claude projects, in beta. You give it a goal, and it "scopes the request, delegates the work, coordinates parallel threads, reviews the outputs".
It has real limits today, though. Those threads run in the cloud, and the docs are clear that local sessions cannot join a project. It needs a Pro or Max plan, and it caps at 200 new threads a day.
The direction is the same on both sides. The phone becomes the place you steer from, and the work happens somewhere else.
One file that every tool reads
Small thing, real payoff.
I keep my company's instructions in plain text files next to the work. Every area has one. It tells whichever AI I am working with how that part of the business runs.
The annoyance was that each tool wanted its own filename. So I kept one file for one tool, and a second file whose only job was to say "go read the first file". It worked, though it was silly.
Last week Claude Code started reading the same filename the other tools already use. On Monday I renamed 42 of those files across the company and deleted the pointers.
The lesson is older than the filename. Keep your context in files you own, in plain text, next to the work. Tools change their minds every few months, and you do not want to rewrite your company each time one does.
One thing to try this week
Find one prompt where you ask AI to review, score, rate or sort something.
Read what you asked for, then read what came back.
If you wanted a number or a label and it sent you three paragraphs, you are paying for writing nobody reads. Ask for the answer on its own. Then put that answer into a rule you can act on.
If you wanted the thinking, keep the paragraphs. That one is the right job for a chat model.
Before you go
My wife and I started a music project, and our first single, In a New Life, is out on September 28. You can pre-save it here.
When it comes to our music, the creative process is sacred. No AI anywhere in it.
What did you try this month that turned out to be genuinely useful, and what did you quietly stop using?
Hit reply and tell me. I read every reply, and our conversations help shape the newsletter.
Thanks for reading.
Tim