I am building a real, customer-facing application almost entirely by direction rather than by hand. The assistant writes nearly all of the code; I supply judgement and ship it. It runs on Google Cloud and has taken real revenue from the first day, which concentrates the mind considerably.
Running. It serves paying customers, and most days I am not entirely sure how.
The full report is open — what it cost, what broke, the step-by-step of how I ran it, and what I'd do differently.
This is an attempt to find the new bottleneck. If the writing of code is no longer scarce, then the scarce thing must be something else — taste, direction, the willingness to ship — and I would like to know its name.
So I built a real application, almost entirely by direction rather than by hand. It is hosted on GCP, it serves paying customers, and most days I am not entirely sure how. That uncertainty is the experiment.
What you need
An AI coding assistant, an account with a cloud host that sleeps when idle, a domain, and a payment provider. Nothing here is expensive; the cost is the nerve to charge a stranger for something you did not type yourself.
Name the smallest thing
Decide on one feature a stranger would pay for, and refuse to add a second. The assistant will happily build you a cathedral; you want a kiosk that takes coins. Write the single sentence describing what the user gets and what they pay — that sentence is the whole brief.
Here is how it runs with the assistant — describe, read, ship, charge, and watch, in that order and no other.
Let the assistant scaffold
Ask it for the boring parts first — auth, a database, a deploy script. Read every file it produces, even the ones you don't understand yet, the way one inspects a rented flat before signing. Commit early and often, so the mistakes are small and dated.
Ship before it's good
Containerise the thing and put it on Cloud Run, where it sleeps when no one visits and costs nothing while sleeping. Point a domain at it. Resist the urge to polish; a live, plain app teaches more than a beautiful local one.
Charge on the first day
Wire in a payment provider and put a real price behind the one feature. Money is the only honest signal; a hundred "I'd totally use this" replies weigh less than one €9 charge that clears. Test with your own card, then turn it loose.
Watch, then ask the assistant why
Read the logs daily. When something breaks — it will — paste the error back to the assistant and have it explain before it fixes. You are the one who must understand the building, even if you didn't lay every brick.
Live, paid, and quietly accruing. It charged its first customer before I had finished admiring it; the sums are modest, but they are sums, and they arrive.
It serves paying customers, and most days I am not entirely sure how. The evidence leans toward the hypothesis, though it leans early and I do not yet trust my own enthusiasm.
Cheaper than a conference ticket, assuming you value your evenings at zero.
The assistant writes the code for free; the debugging of the code it wrote is where the bill quietly arrives.
Revenue imposes discipline
Real money imposes a discipline that side projects never manage. The first charge that cleared taught me more about what was worth building than a month of admiring my own architecture.
Confidence is not correctness
A schema change the assistant proposed with great confidence cost me an afternoon and a small piece of data I would rather have kept. The fault was mine; I had approved it. Read what it writes as you would a contract, not a gift.
The bottleneck moved upward
Code now arrives faster than I can decide what is worth building. The constraint has migrated from typing to judgement, which is both the promise and the unsettling part. I am not yet sure I am the equal of it.
The next move is to add the second feature — the one I have so far refused — and to see whether judgement scales as cheaply as code now does.
The open question is whether I am the bottleneck the experiment was built to find. If so, the finding is uncomfortable and exactly the point.