CD
Search esc to close

projects archived 2026

Sunrise Sunset Tap Co.

A booking site, a texting booking agent and an owners' portal for a tap truck some friends and I talked about running in Perth. A proof of concept we didn't take further.

#ai#perth
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.
The demo at 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'd 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. The name comes from the hill the run club runs over on Wednesdays.

Before anyone spent a dollar, I wanted to see what running it would look like. So I built the whole business as a demo: 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. There's also an explainer of how the truck runs itself and a business plan.

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 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.

What follows is the write-up from the demo's own How this was made page, lightly edited. It covers three days in October, from the first message to a hardened build.

First message to write-up59 h
Pull requests, 19 merged20
Tests, from none on day one360
Decisions, 17 on a card29

The demo, its explainer and its business plan were made in 59 hours, from the first message on the morning of 22 October to the write-up. Nobody wrote code by hand. I asked for things in plain words and made the calls. Claude wrote the code, the copy and the plan, and other models made the pictures and reviewed the work.

Timeline

Each piece of work ran in its own thread, and many ran side by side. Every bar is one thread, from its first message to its last.

Timeline

Every thread, first message to last

22 to 24 October, Perth time · the lines are midnight

Wed 22Thu 23Fri 24
Tap truck business plan
Brand imagery and video
Brand video
Design language
Design and build review (Fable)
Design and build review (Opus)
Site copy editor#3 #4 #5
Review fixes#1
Agent setup adoption#2
Site polish pass#6
Hand-made feel#7
Owner SMS agent#8 #11
Brief business plan
Owner admin#9
Inbox page review#10
Hero and chat button width#12
Hero video upscale#13
Domain availability
Vercel and Cloudflare access
OG image logo#14
Vercel Web Analytics#15
Payback chart styling#16
Owner admin restyle#17
Admin restyle review (Fable)
Inbox queue and SMS review
How it works site#18
Claude Managed Agents fit
Answer engine optimisation
Agent adversarial review#19 #20
The write-up#21
  • build
  • review
  • pictures and video
  • plan and research
From the project's threads · the numbers on the right are the pull requests each thread opened
The full log, hour by hour

Thursday 22 October

Friday 23 October

Saturday 24 October

How the work is set up

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.

Everything happens in one Claude project. Requests go to a coordinator in the project chat. The coordinator starts a thread for each job, and each thread is a Claude Code session in its own cloud container with the repo cloned. Threads run in parallel, often several at once, and each one only knows what its brief and project memory tell it.

Each thread works the same way. It posts a checklist, does the work, runs npm run typecheck && npm test && npm run build, takes screenshots at 375, 768 and 1440 px for any UI change, and opens a pull request. I open the preview and merge. Nothing reaches main any other way.

When a choice changes the goal, I make it. The thread posts a decision card with a recommended option and keeps working on that option while it waits. Picks go into project memory, so the next thread starts with them. The repo's own rules are in AGENTS.md, added in #2 on the first day.

Tools and models

Tools

Tools and models, and what each did here

Tool or modelWhat it did
Claude projectsThe coordinator, the threads, project memory and the shared files folder. Decision cards and checklists in the Claude app.
Claude Code on the webEvery thread: reading the repo, writing code, running tests, shooting screenshots with Playwright, opening PRs.
Claude OpusThe builder in every thread, and the second of the first two reviews. Extra High effort for the hard hardening units, Medium for small ones.
Claude Fable 5.1The reviewer: the first design and build review (20 findings), the inbox review (19), the white admin review (26) and the migration review before #20 merged.
Remote ControlThreads that need a local machine run there: the red team and the auth review.
/counselorsOne prompt to three models from three labs: Fable 5.1, Codex Astra 6 and Gemini 4 Argon (through the Antigravity CLI). They review on their own, then the findings are merged and checked.
ArtifactsLong results as pages: the reviews, the 428-string copy deck, the business plan with in-place editing, the answer engine research.
Figma (GPT Image)The first photoreal stills.
Fal, by MCPEvery other image and video. GPT Image 2 and Nano Banana 2.1 for stills and edits. Kling v3 Pro, Gemini Omni Flash 1.1 and Seedance 2.5 for video. Topaz Starlight Precise 2.6 to restore the soft loops.
ffmpegLoops: H.264 for browsers, seamless ends, posters.
Next.js 15, React 19The site, the admin and the story pages. Drizzle on Postgres: PGlite in WASM locally and in tests, Neon (Sydney) in production.
Anthropic SDKThe two agents, booking and owner, on the Messages API with tools. A rules engine answers when there is no key.
GitHub, VercelPRs, previews, production, Cron for the daily jobs, Web Analytics.
FontsBig Shoulders, Josefin Sans and Instrument Sans on the site. Geist in the admin. Shantell Sans for the diagrams.

Two tools were looked at and left out. tldraw lost to a small hand-drawn engine written for the explainer, because its SDK needs a paid licence in production; the diagrams here use that engine too. Claude Managed Agents was more than this needs: a booking agent doesn't need a managed sandbox per session, so it stays a Messages API loop with a daily Vercel Cron.

Fal cost real money. The account ran to −$36 on the first afternoon, so the rule after that was a cost estimate before every batch. The rear-tap loops cost about $15, the truck wrap about $2 and the Starlight restore about $1.70.

Claude cost more, in tokens. Across 33 sessions the bill comes to $329.87 at API list prices. Almost all of it is reading: each turn reads the whole conversation again, and the prompt cache serves 98% of that.

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
ModelSessionsCostWrittenPer 1M written
Opus28$259.411.91M$135
Fable5$70.460.29M$242
  • Opus
  • Fable
  • sums
From each session's usage record on 24 October · this write-up's own thread stops where the data was taken

Fable costs more per token, and a review reads a lot of code to write short findings, so the review threads cost the most for what they wrote. The coordinator ran at Low effort, so all its sessions together cost less than the biggest thread. The biggest threads, the explainer and the business plan, ran for hours and read their history again on every turn.

The two reviews on my Mac, the red team and the auth review, used /counselors. Each counselor CLI keeps its own logs, so their tokens are counted separately. Three of the six counselor runs failed the first time (an empty file and two timeouts) and ran again. The two reviews came to about $47.98, which makes the total about $378.

Reviews

The two /counselors reviews on the Mac

About $47.98 · failed first tries included

ModelRoleReadWrittenCost
Codex Astra 6counselor8.45M0.06M$15.90
Fable 5.1counselor1.63M0.15M$14.66
Gemini 4 Argoncounselor2.84M0.15M$7.20
Opusran the reviews35M0.07M$10.22
Estimated at list prices from each CLI's token counts, since none reports a bill · Gemini at its launch price, cache reads priced as input

How the look changed

The look changed five times in two days, each time after a plain note on what was wrong with the last one.

Night Run. 22 October, 10:37. The first build, in a dark look. It worked, but "the design leaves a lot to be desired".

Two directions. From three references (golden age travel posters, contemporary flat, Shelf Edge charts). A was Golden Hour, a 1930s railway lithograph. B was Hill Repeats. I picked A for the site and B for the admin.

Golden Hour. 22 October, 11:33. The restyle lands on the site.

Rear taps. Dry hire means nobody pours from inside the truck, and the taps are on the back. So: no runners in the hero, and all eight taps on the rear.

Hand-cut. "It feels more AI generated." Three hand-made treatments of the same hero: A Ticket, B Hand-cut, C Luggage label. B won, against the recommendation.

White console. 23 October. The Hill Repeats admin was "too playful". The white console in #17 is in Geist, with the poster colours as accents only. My verdict: "That's 100x better."

Inbox. The inbox went from a list of chats to a decision queue with drafted replies (#10, option C), then into the white console.

Some rules came out of this and still hold: no left accent bars on rounded boxes, Emil Kowalski's motion guidance, Golden Hour tokens only, and no play and pause button on the loops.

Pictures and video

The first stills were photoreal. Each had a fault a model could fix with an edit: a brand badge, a staff member pouring, the wrong shirt. Nano Banana 2.1 fixed the wedding at 2K with an unbranded fridge and a chalkboard. GPT Image 2 fixed the run club: the Suzuki badge, the club shirt, self-serve taps and nobody in the truck.

Then three video models animated the same two stills. Veo was retired on Fal by then, so Google's entry was Gemini Omni.

The photoreal clips were retired with the move to Golden Hour. For flat, cut-paper motion Seedance beat Kling: the sun rises cleanly and the festoon lights come on. Video then went on hold until the stills were right. The third round of stills went to A2 for the hero and B1 for the truck.

Third round of hero stills: the old hero, option A1 and option A2
Stills, round 3: the old hero, A1 and A2. A2 is the hero now.

The Seedance loops came back soft, at 0.24 Mbps. Topaz Starlight Precise 2.6 at 1x restored them without changing the approved motion (#13).

Loop frame before and after the Starlight restore, zoomed
Zoomed: Seedance as delivered (left), after Starlight (right)

Two models also tried the Golden Hour poster, and the business plan got the seller's Marketplace photo with the wrap edited on.

How the system grew

The first build already did a lot. Each later step added a part, or took one away. On the left, the first build: everything lives in the browser, and the browser runs the jobs. On the right, after the last pull request: data in Postgres, signed texts, owner sign-in and approvals, with the retired parts struck out.

The tests grew with it. There were none on the first build.

Tests

The test suite as each pull request merged

None on the first build

#128
#840
#1049
#1157
#19173
#20360
Diffs

The biggest pull requests

Lines added and removed, from GitHub

#1+3,306 −693
#8+1,215
#17+1,227 −839
#19+2,442 −371
#20+12,572 −1,904
#18+2,186

Hardening

On 23 October the goal changed: the agent and booking system should be production grade and bulletproof, and built so it could serve other businesses later. A demo for other businesses stayed out of scope.

So the system was attacked first. A red team ran on my Mac through /counselors, with three models from three labs reading the code on their own. They found 3 critical and 12 high problems. The three 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. A faked caller ID got the owner agent.
  3. A stranger could take over a booking by quoting its reference.

The fix plan went through 5 reviewers, who raised 33 findings and 13 corrections to the plan itself. Then two pull requests:

  • #19, phase 1: six quick units in an afternoon. A cancel needs a clear yes to a quote, a stranger's reference gets no details, and money settings move to the admin. Each attack it covers was run 3 times before and after: it now fails every time.
  • #20, phase 2: nine units, 12,572 lines added across 99 files. Data on the server in Postgres, texts signed the way Twilio signs them, owners signed in by a one-time link, every owner change staged and run only on YES with a code, and the jobs on Vercel Cron. Midway came the point that "we built our own auth". Three models reviewed it and found 17 holes, all fixed before merge. 8 attacks fail, and the 16 owner flows pass.

Before #20 merged, Fable reviewed the database migrations and the fixes went in the same PR. A deploy now migrates its own database before the new code goes live.

Decisions

These are the calls I made, in the threads and on the cards and artifacts the threads sent. An override means the pick went against the recommendation. That happened 5 times. Every other card went with the default, and the rest were decided without being asked.

Decisions

29 decisions, 17 of them on a card

5 went against the recommendation

WhenQuestionPick
22 Oct 09:52The nameSunrise Sunset Tap Co.
22 Oct 10:12Buy the domain now?Buy nothing and set up nothing yet overrideRecommended: Buy the domain
22 Oct 10:38Where to put the demoDeploy to a vercel.app address
22 Oct 11:21Golden Hour or Hill RepeatsGolden Hour for the site, Hill Repeats for the admin overrideRecommended: A, Golden Hour
22 Oct 11:41More video now?Put all video on hold
22 Oct 12:55Who can approve a payoutEvery owner is a director overrideRecommended: Three directors
22 Oct 12:55Texts the site promises but never sendsCut the lines overrideRecommended: Simulate them
22 Oct 12:55The other 9 judgement callsDefaults: 7 days' notice, 11 pm finish, one hire a day, money held until the hire, runner fees paid quarterly, retry a failed balance for 3 days
22 Oct 14:46Stills first, or stills and video together?Stills only, then video once they are approved
22 Oct 14:46What the hero showsNo runners in the hero, 8 taps on the rear
22 Oct 15:20Which stillsA2 for the hero, B1 for the truck
22 Oct 17:20Ticket, Hand-cut or Luggage labelHand-cut overrideRecommended: Ticket
22 Oct 21:11The site's chat button"Chat", which opens the SMS demo. No web chat.
22 Oct 21:42Play and pause on the loopsRemove it, and accept the WCAG 2.2.2 miss
22 Oct 21:51Wrap the Marketplace photos for the plan (about $2)Go
22 Oct 22:31Inbox: tidy, split or queueC, a decision queue with drafted replies
22 Oct 23:06How to fix the soft loopsTopaz Starlight Precise 2.6 (about $1.70)
23 Oct 12:55Share imageOption A, the logo on the hero
23 Oct 14:10Font for the white adminGeist
23 Oct 14:27White console or Hill Repeats adminThe white console
23 Oct 14:48The design reviews' findingsAccept every recommendation
23 Oct 15:28Font for the diagramsShantell Sans, for diagrams and margin notes only
23 Oct 16:32How the explainer talks about ownersNever state how many. The byline says "For review".
23 Oct 16:38Search enginesKeep the demo out of search until a real domain
23 Oct 17:09How good the system has to beProduction grade, built so it could serve other businesses later
23 Oct 18:22A demo for other businessesOut of scope, and out of the business plan
23 Oct 18:24Effort per unit of workExtra High for the hard units, Medium for the small ones
23 Oct 23:36Hosted PostgresNeon, not Supabase
24 Oct 09:51Database set-upNeon in Sydney, Neon Auth off

The business rules came first, from three rounds of questions on the first morning. They are still the rules the code runs on:

  • A Pty Ltd, and every owner is a director.
  • Dry hire: the client buys the drinks, we bring the truck.
  • No staff. Owners run deliveries and are paid a runner fee for each.
  • 4 or 8 hours, a deposit now, the balance before, a bond hold.
  • A travel fee more than an hour from the Perth CBD.
  • 7 days' notice, up to 6 months ahead, one hire a day, done by 11 pm.
  • More than 30 days out, a full refund. Inside 30, the deposit is kept.
  • A paid booking beats an owner's own use.
  • Payouts each quarter. Every owner sees everything.
  • The agent may confirm a booking and take payment by text.
  • Texts first, then email, phone and the web form. No web chat.
  • Glassware included, no add-ons.

The artifacts carried decisions too. In the copy deck six edits came back: Perth metro, delivery from the Perth CBD and a dry-hire voice (#4). The business plan was edited in place. Every recommendation in the two design reviews was accepted, and my answer to the red team was "plan the fixes".

What worked

  • One thread per job. Threads ran in parallel without stepping on each other, and each ended in one PR that was easy to read.
  • A default on every question. Work never waited for an answer, and 24 of 29 calls went with the default or needed no card.
  • Reviews by other models. Each review found things the builder missed. The red team found the three that mattered most.
  • Say what is wrong. "Feels AI generated" and "too playful" each led to a better look within hours.
  • Cost first. After the Fal overrun, an estimate before each batch kept spend to a few dollars a round.

The making of the making of

The original write-up was part of the build too. It was made the same way as everything above it: one thread, one pull request (#21) and rounds of review. So it was a dream inside a dream: the truck, then the demo and its explainer, then the page about them, then a section about the page, asked for at 10:18 pm on 24 October while the third round of the page was being built.

The top of the original How this was made page
Level 3: the page, as it stood

The levels leak into each other. The timeline has a bar for the write-up's own thread, and it stops where its data was taken. The token chart counts that thread, and counting it cost tokens. The decisions table doesn't list the decisions about the page, because that would be level 5.

The top is still spinning.