Four days to build. Three weeks to fix.

At two in the morning on a Monday, my writing app locked up on me.

Not a crash. Worse than a crash. It just stopped, mid-sentence, cursor blinking like everything was fine. I had written the thing. I had tested the thing. The thing was, at that moment, sitting in Apple's review queue waiting to be sold to strangers.

Hold that thought. I am going somewhere with it.

You have probably seen it: the new Mac Mini "killed AI subscriptions forever." Local models, always-on hardware, agents humming at 3 a.m. for pennies of electricity. Cancel everything. Own your stack.

My first reaction was an eye roll. My second, after a few days with it, was that the eye roll was lazy. There is something real under the hype, and it has quietly changed how I work.

Local models are not replacing frontier models for the hard things. Real reasoning, architecture calls, code that has to survive contact with users: a small local model is not in that conversation. Anyone telling you to cancel everything and go fully local has not shipped anything difficult lately.

Here is what is true, and it surprised me more than the headline did. The price of persistence just collapsed.

Not the hardware. That is still about two grand. What collapsed is the price of leaving it on. A Mac Mini sips power like a lightbulb, and a local model parked on it can watch, sort, summarize and draft around the clock while your cost per hour rounds to nothing.

It is not smarter than a frontier model. It is just cheaper to leave awake than anything else you can own.

That is the shift. For years the only expensive question was "how smart is it?" Now there is a second one next to it: "how long can it stay on?" Different problems, and the moment you solve both with the same tool, you overpay for one and underbuild the other.

Which brings me to something I have been chewing on for months.

I call it the L0 to L5 AI Delegation Ladder. It is really just a way of asking what a task is made of.

L0 and L1 are pure watching. Something that flags the three emails that need you and buries the forty that do not. No judgment, just attention that never tires.

L2 is first-draft triage. Summarizing the meeting, tagging the notes, turning a rambling voice memo into an outline before you finish your coffee.

L3 is shaping, not deciding. Tightening a paragraph, finding the hole in your logic you were too close to see. Frontier territory, but you are still holding the pen.

L4 and L5 are the real calls. What to ship. What to write. What to bet on. This never leaves your hands, and if you are honest, you do not want it to.

The bottom rungs are about persistence, which is exactly where a cheap always-on machine now wins outright. The top rungs are about judgment, where you reach for real intelligence, when you need it, on your terms.

Not local versus cloud. Local for persistence, frontier for judgment, and you holding both sets of keys.

I did not get that from a diagram. I got it from the bug.

Cloister Ink took about four days to build and three weeks to make correct.

The building was almost pleasant. First commit went in on the 8th of July. By the 10th there was a real app: the vault, the sidebar, the quiet serif editor. If I had stopped there I could tell you a wonderful story about how fast this goes now.

Then I started using it, which is different from testing it.

That two in the morning session? I sat down to actually write something in my own app, and in about an hour I found ten bugs. Ten. In a build I had already submitted. I shipped four replacement builds before the day was over, then turned off automatic release on every app I own, because if Apple had said yes while I slept, all ten would have gone to real people.

There were rejections on the 15th and the 16th. There was a keyboard that covered the chat bar and survived three rounds of fixes before I admitted the problem was not the keyboard, it was the layout.

And then there was the freeze. The one from the top of this email.

Here is what it was, and the answer is almost funny. Work that grew with the length of your document was running inside a piece of the screen that redrew itself on every single word the AI streamed back. Every word. On a long manuscript, that is thousands of full redraws, each one redoing the whole job while you watch a cursor blink.

Once you see it, it is obvious. Getting to the point where you can see it took three weeks.

Machines did most of that work. They ran the builds, installed to my iPad, wrote the regression tests and ran them again at four in the morning while I slept. Tireless, cheap, uncomplaining. That is persistence, and it was most of the hours.

But nothing ever walked up and told me the freeze was hiding in the redraw. Those three weeks came down to a handful of calls only I could make: which symptom is the actual bug, is this fixed or just quiet, do we ship or hold.

Those calls took minutes. They were also the entire job.

The hours are persistence. The job is judgment. Rent the first cheaply. Never hand off the second.

Cloister Ink is my writing app, on iPhone, iPad and Mac.

It is built bring-your-own-key: you plug in your own Anthropic key and the AI bills you directly. No middleman subscription marking up somebody else's model. It works for you because you are the one paying for it.

It lives at L3 on purpose. A writing tool should be an editor, not a ghostwriter. It sharpens what you wrote. It does not put words in your mouth.

That freeze fix ships in the next update, so if you are already writing in Cloister Ink, it is on the way.

It is $29.99 once. No subscription, not now, not later.

One honest note, since it is already on my site and I would rather you hear it from me: $29.99 is a founding price and it does go up. I am not going to invent a countdown, because I have not set a date and will not pretend otherwise. What I will promise is that whatever you pay now is what you paid, forever, and every update this app ever ships is yours.

If the ladder is the part that stuck, stay. Sorting persistence work from judgment work is most of what I write about here.

Own your words. Own your stack.

Christopher

P.S. The same rule runs through everything under Seven Swords Productions: you own the stack. MedKeep, my health journal, takes it furthest. Your data never touches a server, because there is no server. What is the last thing you delegated that you wish you had kept? Hit reply. I read everything.