Twenty dollars
I was bored, I had twenty dollars in my bank account, and I spent all of it on a Claude Code subscription. That is the whole of the origin. There was no plan attached to it and no project I was trying to finish. I wanted to find out what the thing could do.
In June I moved up to the hundred dollar plan, before any of it had made a single cent back. I have not missed a month since.
Learning to code, in about six weeks
I had taken AP Computer Science to prepare for a computer engineering degree, and I was not learning much in it. So I taught myself instead. Thirty days of Python and roughly three hundred exercises, all kept in one enormous document. Then two weeks on Java, eight hours a day, through March and April, before finals. I am not going to claim I learned Java in two weeks. I learned enough of it to read it, write it, and finish the exercises I set myself, which is a much smaller claim and the true one.
The detail that mattered was not the language. It was that nothing ever told me I was wrong. The Java exercises were typed into a document and checked by eye. Nothing compiled. Nothing ran. When I got something wrong I stayed wrong until I worked it out myself, and that sometimes took days.
How long you stay wrong is set by how fast something tells you.
That is the only idea I took out of those six weeks, and it is the one every system on this site is built around. The scheduler texts me at 7 AM instead of waiting for me to open a dashboard. The simulation harness runs a batch of scenarios with the expected answers written down first, so a wrong result is visible in seconds instead of at the end of a week. The teaching assistant loop takes assignments off a syllabus, puts them on my iPad, takes the finished work back and marks it, and the only question I was asking while designing it was how few steps that round trip could take.
I did not learn that from a book about software design. I learned it from writing Java for two weeks with nothing to check it against.
Three attempts at the same system
By May I wanted a second brain: somewhere my notes, my sources and my schedule could live together and stay connected. I built it three times.
A week, then nothing
It lasted about a week. It was too convoluted to keep in my head and there was no system underneath it, so every new piece made the previous pieces harder to change. I could not have told you what was done and what was not.
I handed the whole thing to the model
The second time I started slower, and then made the opposite mistake. I decided the model was the intelligent one and got out of its way. It proposed, I approved, every single time, for as long as that lasted.
Then I read what I had. It was not good, and worse, I could not say why any particular piece of it existed. Handing over the decisions does not save you the work. It moves the work to the point where you have to understand a system you did not design.
The third attempt is the one that is still running. Its first commit is dated May 13, three days after I abandoned the second. The difference was not a better idea. It was that I stopped approving and started directing: reading how people actually structure this kind of project, deciding the architecture myself, and making the model build against decisions I had already made.
What it was supposed to be
Small. A notes database, because I thought the linked-notes idea looked interesting. A page holding my goals with their dates on it, and a second agent that could read that page and write a summary of it. That was the entire ambition.
I wanted calendar integration from the beginning and assumed it would take a week, maybe a month. It took multiple months, because I kept refusing to build it on a foundation I did not trust yet. The first eleven phases, about two weeks of work, produced no features at all. They decided how the system would work.
That is the trade I would defend to anyone: the calendar was late by months and it has never once written an event I did not want.
What it turned into
Fifty-five phase documents, each with an acceptance test written as one sentence before any code existed. A hundred and twenty-eight architectural decisions recorded with their reasoning and the alternatives rejected. A separate log of every correction I have been given, which I read before starting anything non-trivial. Three hundred and sixty-one wiki pages where every claim carries the source it came from. Fifteen scenario batches kept as regression suites.
And three things a stranger can look at: a 185-page workbook that is on sale, fifty-two courses that replaced a subscription I could not afford, and a system that plans my day and texts me about it while I am asleep.
What it costs
One recent twelve-hour session metered at $138.40. That session was about a quarter of a week's usage, and I run roughly twenty of them a month. Two different ways of counting agree: the tokens I use in a month would cost somewhere between $2,500 and $2,800 at published API prices.
I pay twenty dollars a month.
I am not claiming that is what the work is worth. It is a price comparison, and the only reason it is on this page is that people ask what the setup costs, and the honest answer is less than a textbook.
Why any of this
Two reasons, and the second one is the real one.
The first is that I wanted the things. I wanted a course that would get me through Calculus I, something to open instead of scrolling, and a system that would stop me quietly dropping work. I built all of them for myself, so it would be hypocritical not to use them. I use all three.
The second is that I think this is where the field is going and I want to be fluent in it now rather than in four years. Every project here was also an excuse to learn how to direct these tools properly, which is a skill with a short head start attached to it. I would rather spend that head start on real systems than on tutorials.
In March I could barely code. Five months later I am building systems that run without me.
I am not going to pretend that is a large amount of time or that any of this is finished. It is the beginning of the thing, not the end of it. But the distance between March and now is the only evidence I have that I can learn something hard quickly, and it is the reason I would rather be handed a real problem than a curriculum.
The three things this produced are written up in depth: Cogito, Cockpit and The Language of Calculus. If you run a lab, a compute center, or a team where an undergraduate can be useful early, email me.