English Cymraeg
Async Digital is a one-person studio testing, in the open, what really changes when AI can write the code.
The claim everywhere right now is that AI replaces the engineer. What I keep finding is that it moves the engineer instead: from writing the code, to building the system that writes and checks it, to deciding whether to trust what comes out. That last decision is the part that doesn’t automate, and it’s still mine. It changes what the work is, day to day: I groom the tickets the moment a discovery lands, not because grooming is the done thing, but because the English is the code now and the Swift is what comes out of it. The system that does the rest is Ordova.
The system behind the work
The system the studio runs on: software agents that review the code, enforce the conventions, check the copy, and watch for drift. The last call, whether to trust what they produce, stays human.
ordova.async-digital.comI can run this experiment because the risk is mine to absorb. A business can’t gamble its revenue to find out whether hands-off AI works. I can, on my own products, and report what actually happens. If one fails, the lesson is the point.
The products below are real. Ordova is how they get built. The findings are in the notes and case studies. All of it is public, because proving it is the whole exercise.
Rob Nash Engineer
I want to verify for myself that AI is actually useful. And I don’t want to wait for the market to adapt and inform me of where I belong. I started out worried it would replace me. Having actually used it, I think it will make me better instead, but I’d rather find that out first-hand than be told.
Published · book
Principles for running an AI coworker in a one-person studio.
coworker.async-digital.comI wrote the notes for myself, because I kept re-learning the same lessons about running an AI coworker. The book is those notes, ordered and argued for.
What follows is ordered by time where the time is honest, and by state where it is not. The writing carries a publication date, so it comes first. A shipped app does not, so it sits further down under what it actually is.
Seven working notes, newest first. Four of them in August.
August 2026
30 Aug
Note
The polite way to coordinate two AI sessions sharing one save file, asking whether it has finished, is the message that strips its protection. I caused the bug live while researching it.
30 Aug
Note
A file my agents keep said the opposite of the truth about my own machine for weeks. A June paper measures why corrections get thrown away by the check meant to stop duplicates.
20 Aug
Note
An agent filed nine paragraphs to change two words in a comment. It had followed its instructions exactly.
5 Aug
Note
A session hands itself off to the next one, and won’t trample work another session is already doing. Neither behaviour is a prompt.
May 2026
18 May
Note
The rules you build with an AI assistant are professional capital. They belong to the studio, not the vendor.
2 May
Note
The primitives missing from agent-swarm runtimes, and the disciplines that fill the gap.
April 2026
24 Apr
Note
I counted what thirty days with an AI coworker actually produced, and about a third of it never reached git.
Six studies and an internals track, added to as the work goes.
A case-study series on Iris, the deep-linking layer I build here. Six studies: one states the principle, five measure it against a demo messaging app built as its consumer, and an internals track goes inside the library. No single publication date, because it is not finished.
The principle
One resolver for taps and URLs
Five measurements
When URLs arrive faster than the UI race safety
Where URL handling lives containment
Same URL, right meaning, from anywhere state
When navigate means more than push composition
A primed draft from a single URL the agentic bar
Internals · four short pieces
How it works, from the inside the URL-to-intent seam, the existence gate, the sheet-versus-push resolver, the one-shot flow baton
Moving now. None of it has a finish date.
The system the studio runs on: software agents that review the code, enforce the conventions, check the copy, and watch for drift. The last call, whether to trust what they produce, stays human.
The lessons have not stopped, and that is the finding. Everything else here taught me something and then finished; this one keeps changing, so it keeps teaching.
A Swift package that routes incoming URLs into typed SwiftUI navigation, so an app doesn’t have to hand-wire deep links. MIT, tagged from 1.0.0, listed on the Swift Package Index. The series is the evidence for it.
Deep links are plumbing right up until something automated starts firing them. Tests, Shortcuts and agents send URLs in bursts, mid-animation, into whatever state the app was left in, and none of the polite assumptions survive that.
Releasing a Swift package and cascading the version bump through every internal package that depends on it, in one command.
I pin to the next minor, so only patch releases float in. Untested, but a risk I accept at that size. The tool assumed a caret, called the plan clean, then failed mid-cascade with the tags public. It blocks that now.
Teaching a complete beginner Texas Hold’em on the big screen. Simulated chips, no real money.
The Apple TV remote has four directions and a click. Every teaching idea I had needed a keyboard, so the remote is doing the design work now.
Shipped, paused, or stopped. Older work, and I am not presenting it as new.
Now on the Mac App Store
A macOS app that uses on-device AI to transcribe and search the audio and video you give it.
audient.async-digital.com
Don’t let the agent run the project. Archive the decisions, keep the epics focused and clear, layer the code and stay critical about the architecture, and install hard gates.
On-device transcription from the command line. Point it at a file and it writes out the words, with no app and nothing uploaded. It runs here every day; the repository is private, so there is nothing to link to yet.
Local AI works. Transcription runs entirely on this machine, and I trust it enough to build on. I should still have built the command line first: doing it after Audient meant proving the same thing twice, in the harder order.
Principles for running an AI coworker in a one-person studio. I wrote the notes for myself, because I kept re-learning the same lessons. The book is those notes, ordered and argued for.
A glanceable dual-timezone clock, with a Home Screen widget and a watch complication. It is paused, not dropped. Ordova grew into a job of months, and the books are being written alongside it.
Shelving it was right and it still nags at me. A widget and a complication turned out to be two products, not two features.
Turning work notes into four-blocker reports with on-device AI. I could not put a boundary around what the model would say.
I could not honestly promise a user what the model would never say about their colleagues. That was the entire product, so it stopped.
Abandoned. Never released.
Counting the sleeps until the next birthday, on one screen. A finished v1, built for my daughter, who did not take to it.
Built for one user, aged seven, who wanted something else. Knowing what holds a child on a first screen is a designer’s specialism, and a brief cannot carry it.
Read the post mortem →