In 2021 a client paid me to figure out how to kill her own business.
That is not how she put it. She asked for a competitive threat scan. She ran a service business — the kind where revenue is a person’s hours times a rate — and she had started hearing the word “AI” in rooms where it had never come up before. The assignment was to find the places where a machine would undercut her billable rate, and to find them before somebody else did.
I took the work because it was interesting. I wrote the memo. The memo was right…but it was useless, because a memo tells you the water is rising, but it does not teach you to swim.
So I jumped in and learned to swim. Fast, and mostly by diving in over and over.
Five years, compressed:
I stopped writing about the tools and started cataloging or using them. Then I started wiring them together. Then I started building things I knew no one was likely to buy, but that I wanted to see in the world. Voice agents that acted as thought partners to operators at two in the morning. Inbox triage automations that respond like me. A knowledge base that studies the field while I sleep.
Automations that survive the week I forget they exist.
The receipts are public and I will not belabor them: 780 hours freed per year for one operator. Fifteen-plus hours a week for another. A thousand dollars a month of tooling deleted outright. One week in July where I logged 101.6 agent-hours against 40.7 hours of my own — a 2.49x multiplier, method published, including where the count is generous.
None of that was planned. It was a result of the refusal to stay on the analysis side of the glass.
Which brings me to the least impressive request I have ever made of a machine.
I record video. Course modules, build walkthroughs, the occasional talk. Every raw take has the same four problems: dead air where I gather my thoughts, volume that drifts from clip to clip, no logo, no intro, no outro.
Four problems. That is the entire list.
Adobe would like about thirty-five dollars a month to solve it. Premiere Pro is a magnificent instrument and it is aimed at a person whose job is editing. My job is not editing, and I don’t want it to be. I only needed four operations applied identically to a folder of files, forever, and I was being asked to rent a four-hundred-function cockpit to do it — and to keep renting it for as long as I intend to keep making videos.
That is the part worth sitting with. The subscription is not priced against the work you are doing. It is priced against the work you might do. I was paying for the ceiling but living on the floor.
I did not want a better editor. I wanted the four things, on a schedule, with my name on the output.
I opened Claude Code, put it on Fable 5, and asked for something I had no business asking for: a real Mac application. Not a script. Not a folder of terminal commands I would forget the syntax of by Thursday. A window you drag files into.
Then I put on my Tyrone Biggums beanie and did not come up for air.
I did not know how to build a Mac application. That was the point. I have gotten good at pointing an AI at a stretch goal and then refusing to accept an answer I do not understand — because the moment you accept the part you cannot read, you have stopped building and started praying. Dumb mistakes and smart questions, over and over. And Claude is a good and patient teacher — it will not mock you for asking it to break something down “like I was a third grader.”
So I learned it in public, one argument at a time.
The app landed audio at −17.3 LUFS against a −16 target. That one number turned into a crash course in sound engineering, studio setup, and 4K cameras — and once I understood what it was actually measuring, the fix was obvious. Real footage lands at −16.3 now. Issue after issue went like that: the right question, then the fix.
And then one problem truly stopped me. I ran a batch and the app reported a defect I had not asked it to look for: the right channel of a take was dead. Sixty-three decibels down against a live left channel. It had been dead the whole time. The editor had been faithfully preserving a broken recording because that is what editors do.
I had built a tool to fix my videos and it found a fault in my room. The microphone was miswired. The fix belonged upstream, in the studio, not in the software — so I wrote a guard that rescues the take, and a brief that fixes the cause. If that guard ever fires again during a fresh recording, it is a red flag, not a green one.
That is what a tool you built yourself gets you that a tool you rented never will: it tells you the truth about your own setup.
Somewhere in there I learned more about sound than in the thirty-odd years I have been listening to it, and finally understood what a friend of mine used to do for a living. That turned into its own field note.
Sixteen hundred and sixty-nine lines. Six tests that run in eight seconds.
Then I got greedy, and decided a thing I built should install like a thing you buy.
That means Apple has to sign it. Which led to a dormant developer account being reactivated, a certificate that defaults to the wrong keychain and fails silently, an intermediate authority certificate nobody mentions until nothing is trusted, and a config flag that reads true where it wants a team ID.
Every one of those cost me an afternoon. Every one of them is now written down, because the second app will take an hour.
This morning my machine told me this about a file I made:
Accepted. Notarized. My name.
Anybody can install it now — no right-click workaround, no “unidentified developer,” no apology. The same gate that stands in front of every application on this platform opened for a piece of software I wrote to trim silence out of my own videos.
Five years ago I was writing a memo about what the machine was going to do to somebody’s business.
I want to be careful here, because the wrong lesson is learn to code.
The lesson is smaller and stranger than that. There is a category of problem — probably several sitting on your desk right now — where you are paying a monthly fee to route around an annoyance, and the annoyance is four operations long. You have been trained to believe that the distance between that annoyance and a real, signed, installable piece of software is a career. For a long time it was. The tools were behind a guild, and distribution was behind a signature you had to be somebody to get.
It is eight days now.
Eight days if you are willing to not understand something in the morning and understand it by dinner.
So: what is the small, dumb, expensive thing you have been working around?
Here is an idea of what I am doing. I have my thought partner going back through every idea I ever parked — the half-built things, the notes-to-self, the folders named “later” — and ranking them. Not brainstorming new ones. I have plenty. Refining the ones I already had and judging which deserve the next eight days.
Turns out the backlog was never the problem. The problem was that nothing in it looked buildable.
But everything in it looks buildable now.
Stay tuned.