---
title: "The tap truck we didn’t buy"
url: "https://colindismuke.com/writing/the-tap-truck-we-didnt-buy"
description: "Some friends and I talked about hiring out a tap truck around Perth. Before buying one, I built the whole business as a working demo in 59 hours, without writing any code by hand."
published: "2026-11-20T12:27:25.255Z"
---

[writing](https://colindismuke.com/writing) 20 nov 2026 

# The tap truck we didn’t buy

Some friends and I talked about hiring out a tap truck around Perth. Before buying one, I built the whole business as a working demo in 59 hours, without writing any code by hand.

#ai#perth 

Also on the projects page: [Sunrise Sunset Tap Co. →](https://colindismuke.com/build/sunrise-sunset-tap-co) 

![The Sunrise Sunset Tap Co. demo home page: a cut-paper sunset over a small truck with taps on its back, and the line Cold brew at dawn. Beer at dusk.](https://colindismuke.com/images/projects/sunrise-sunset-tap-co/home.webp)

The demo at [sunrisesunsettapco.cpd.dev](https://sunrisesunsettapco.cpd.dev)

A few friends and I spent a while talking about buying a small tap truck and hiring it out around Perth. The idea was dry hire: the client buys the drinks, and we drop off a little Suzuki Carry with eight taps on the back, then come and get it when the night's done. Cold brew in the morning, beer at night. We'd call it Sunrise Sunset Tap Co., after the hill the run club runs over on Wednesdays.

Talking about a business is easy. Running one is mostly the boring middle: someone texts at 9 pm asking if the 14th is free, someone has to take a deposit, someone has to decide who gets paid what at the end of the quarter. Before anyone spent a dollar on a truck, I wanted to see that middle. So I built it.

Not a mock-up. A working demo of the whole business: a public site with live availability and booking, a booking agent that answers texts, email, calls and the web form, a simulated Stripe checkout, and an owners' portal that runs the money from deposit to quarterly payout. It took three days in October, 59 hours from the first message to the write-up. I didn't write any of the code by hand. I asked for things in plain words and made the calls. Claude wrote the code, the copy and the business plan, and other models made the pictures and reviewed the work.

None of it is real. Every payment, text and bank transfer is simulated, no domain was bought, and we aren't doing anything with it right now. It was a proof of concept, and it answered the question it was built for. The [demo](https://sunrisesunsettapco.cpd.dev) still works if you want to poke at it: text the agent from the Chat button, book a date, then sign in as an owner from the outbox and approve a payout.

First message to write-up**59 h**

Pull requests merged**19**

Tests, from none on day one**360**

Calls I made**29**

## One thread per job

![Hand-drawn diagram: the owner asks the coordinator, which starts threads; threads push pull requests, ask for decisions on cards, write artifacts, share memory and files, make images through Fal, and run reviews on a local Mac with /counselors.](https://colindismuke.com/images/projects/sunrise-sunset-tap-co/setup.webp)

Everything happened in one Claude project. I'd ask for something in the project chat, and a coordinator would start a thread for it. Each thread was its own Claude Code session in a cloud container with the repo cloned, and several ran at once. A thread only knew what its brief and the project's shared memory told it.

Every thread worked the same way. It posted a checklist, did the work, ran the type check, the tests and the build, took screenshots at phone, tablet and desktop widths for any UI change, and opened a pull request. I opened the preview and merged it. Nothing reached main any other way.

When a choice would change what we were building, it was mine. The thread put a card in front of me with its recommended option and kept working on that option while it waited. That worked better than I expected. Work never sat still waiting for me, and 24 of the 29 calls I made either went with the recommendation or never needed a card at all.

The five that didn't are the interesting ones. I said no to buying the domain. I made every owner a director, where the recommendation was a smaller board. I cut the texts the site promised but never sent, rather than faking them. I kept two design directions instead of one. And I picked a hand-cut paper look over the recommended "ticket" style.

## Five looks in two days

The first build worked, in a dark theme, and I didn't like it. "The design leaves a lot to be desired" was the whole note. Two directions came back: Golden Hour, styled like a 1930s railway lithograph, and Hill Repeats. Golden Hour went on the site.

![The first build's home page, in a dark theme](https://colindismuke.com/images/projects/sunrise-sunset-tap-co/v1-home.jpg)

The first build

![The Golden Hour home page](https://colindismuke.com/images/projects/sunrise-sunset-tap-co/v2-home.jpg)

Golden Hour, an hour later

Then it looked too much like what it was. "It feels more AI generated." Three hand-made treatments of the same hero came back: a ticket, cut paper and a luggage label. I picked the cut paper.

![The hero four ways: base, Ticket, Hand-cut, Luggage label](https://colindismuke.com/images/projects/sunrise-sunset-tap-co/handmade-hero.jpg)

Base, ticket, hand-cut, luggage label

![Home page in the Hand-cut style](https://colindismuke.com/images/projects/sunrise-sunset-tap-co/home-handcut.jpg)

Hand-cut, as shipped

The owners' admin had the same problem the other way round. The Hill Repeats look was "too playful" for a screen where you approve payouts, so it became a plain white console, with the poster colours as accents only. My reply when I saw it was "That's 100x better."

![The owners' admin in the Hill Repeats style](https://colindismuke.com/images/projects/sunrise-sunset-tap-co/v2-today.jpg)

Hill Repeats

![The owners' admin as a white console](https://colindismuke.com/images/projects/sunrise-sunset-tap-co/white-today.jpg)

The white console

That's the main thing I'd keep from this part. Say what's wrong, in plain words, even if it's only a feeling. "Feels AI generated" and "too playful" each led to something better within a few hours.

## Pictures that cost money

The images and video came from other models, through Fal. The first stills were photoreal, and each had one fault an edit could fix: a brand badge on the truck, a staff member pouring when dry hire means nobody pours, the wrong shirt on a runner. Then three video models animated the same stills, and a fourth model cleaned up the soft loops.

![Run club still, first try, with a brand badge and someone pouring from the truck](https://colindismuke.com/images/projects/sunrise-sunset-tap-co/photo-runclub.jpg)

First try

![Run club still, fixed: no badge, the club shirt, self-serve taps and nobody in the truck](https://colindismuke.com/images/projects/sunrise-sunset-tap-co/photo-runclub-fixed.jpg)

Fixed with one edit

This part cost real money, and at first nobody was watching it. The Fal account was $36 overdrawn by the first afternoon. After that, every batch came with a cost estimate first, and most rounds cost a few dollars. The hero loops, with the taps on the back of the truck, were about $15.

The photoreal clips didn't survive the move to Golden Hour anyway. The site ended up with flat cut-paper loops: the sun comes up, the festoon lights come on.

## Making it hold

On the second day the goal changed. The booking agent and everything behind it should be production grade and bulletproof, not a demo that happens to work, and built so it could serve other businesses later.

So the first step was to attack it. Three models from three different labs read the code on their own and tried to break it. They found 3 critical problems and 12 high ones. The critical ones, in plain words:

1. The server believed the browser. The data lived in the browser, so anyone could edit it.
2. A phone number was an owner's password. Fake the caller ID and you were the owner.
3. A stranger could take over a booking by quoting its reference.

Five reviewers went over the fix plan and corrected it 13 times before any code changed. Then two pull requests did the work. The first was quick fixes: a cancel needs a clear yes to a quote, and a stranger with a booking reference gets no details. The second was the big one, 12,572 lines added: the data moved to a real database on the server, texts were signed the way Twilio signs them, owners signed in with a one-time link, and every change an owner asked for by text was staged and only ran when they replied YES with a code.

Halfway through that second one I realised we'd built our own sign-in system. Three models reviewed it and found 17 holes, all fixed before it merged. The attacks written against it now fail, and the test suite went from none on the first day to 360.

## What it cost

Making it wasn't free. At API list prices, the Claude sessions came to $329.87, and the outside reviews on my Mac to about $48 more. Call it $378\. Almost all of that is reading. Every turn of a thread reads the whole conversation again, and the prompt cache served 98% of it.

Tokens

Cost per thread, the ten biggest first

$329.87 across 33 sessions at API list prices · 593M tokens read, 98% from the cache · 2.21M written

How it works site$31.09

Tap truck business plan$28.68

Inbox page review$23.05

Site polish pass$23.02

Owner SMS agent$21.27

Owner admin restyle$20.60

Site copy editor$18.09

Hand-made feel$17.47

Review fixes$14.49

Design and build review (Fable)$13.20

19 other threads$95.53

The coordinator, 4 sessions$23.38

| Model     | Sessions | Cost    | Written | Per 1M written |
| --------- | -------- | ------- | ------- | -------------- |
| **Opus**  | 28       | $259.41 | 1.91M   | $135           |
| **Fable** | 5        | $70.46  | 0.29M   | $242           |

- Opus
- Fable
- sums

From each session's usage record on 24 October

The two biggest threads, the explainer and the business plan, ran for hours and paid for their own history on every turn. The reviews cost the most for what they wrote, because a review reads a lot of code to produce a short list. The coordinator, which ran at low effort, cost less across all its sessions than the single biggest thread.

## What it answered

We set out to find out what running a tap truck would actually involve, and now I know: a booking flow, a texting agent, a ledger, quarterly payouts, a sign-in system, and a list of rules about notice periods, refunds and who's allowed to approve what. Three rounds of questions on the first morning set those rules, and they're still the ones the code runs on. Seeing them all in one place was the point.

The demo's still up, and the [project page](https://colindismuke.com/build/sunrise-sunset-tap-co) has the full hour-by-hour log, every decision with what was recommended, the tools and models, and the rest of the pictures.

## Related

- [Sunrise Sunset Tap Co.](https://colindismuke.com/build/sunrise-sunset-tap-co)build
- [Tiny Perth](https://colindismuke.com/build/tinyperth)build
- [Drawing all of Perth](https://colindismuke.com/writing/drawing-all-of-perth)writing
- [password.coffee ☕](https://colindismuke.com/writing/password-coffee)writing
