Blog/
How we built Quizlamp: plan, stack, days, what went wrong
Quizlamp, our pub-quiz web app, went from an approved concept to a working first version in about one day (3 to 4 November 2026), then took six more days to launch on free channels (10 November). It had 325 commits in its first 14 days and has about 33,000 lines of TypeScript. AI agents wrote the code and a separate checker tested it. It runs on one small shared server with no paid ads. This post gives the plan, the stack, the timeline and the mistakes, each checked against our project records and code history.
Quizlamp lets a host build a quiz, show it on a big screen, and have teams answer on their phones. You can try it at zinobits.com/quizlamp/, and the short case is at /work/quizlamp/. We built it as our own product, so we can show real numbers without a client's permission.
The plan: concept first, then one build
Before any code was written, the app went through a pre-production phase that included market research, maps of three competing quiz apps, our own feature list, and a written concept. The owner approved that gate on 3 Nov. We set a simple rule: nothing is built until the gate passes.
After that, a single builder agent (Sonnet 5.5) created version 1 in one pass, using only the feature list. Its log records that the first build passed the project's checks (type checking, linting, layer rules, 13 unit tests, 10 end-to-end tests in a real browser) on the evening of 3 Nov (the build log).
A different agent then evaluated the build with hidden tests that the builder never saw. It failed three items (a silent team in the marking queue, repeated team colours, an expired pass still showing as live). One round of fixes on 4 Nov resolved them, and the owner accepted the first-working-build gate that day.
| Date (2026) | What happened | Source |
|---|---|---|
| 3 Nov | Concept approved; version 1 built; 5 commits | project goal file, commit log |
| 4 Nov | Fix round, rename from Quizhall to Quizlamp, landing page live, shared server set up; 43 commits that day | commit log |
| 6 Nov | Lovable redesign goes live (Vite and React Router) | merge 5395b8e |
| 7 Nov | Front end moves to TanStack Start | commit d550b36 |
| 9 Nov | TanStack Start branch merged into the main line | merge 13836e9 |
| 10 Nov | Launch on free channels only | launch goal file |
| 13 Nov | 10 languages | project notes |
The stack, and why each piece
- Front end: React 19, TanStack Start and Router, Tailwind 4, built with Vite. TanStack Start provides full-document server rendering and server functions in a single project, which fits a quiz page that must load quickly on a phone. The front end changed twice: version 1 used our own small web layer, the Lovable redesign replaced it on 6 Nov, and a TanStack Start version of that design was merged on 9 Nov.
- Server: Hono on Node. It is tiny and based on web standards, so the same handlers run anywhere.
- Data: a single SQLite file accessed through Node's built-in
node:sqlite. We use it as a document store (onedocstable withkind,id,owner, and JSON). The module is marked “release candidate,” so we keep it behind a thin wrapper file. Live quiz sessions stay in memory and are written to SQLite after each change. - Realtime: server-sent events, with the server owning the quiz state. Answers and host notes are omitted from the phone and big-screen views until they are revealed, preventing a team from reading the answer early.
- Payments: Freemius acts as the seller of record, handling tax, fraud, and disputes. Their page states there are no setup or monthly fees; you pay only when you sell. Prices are $6 for a one-night Pass, $15 a month or $150 a year for hosts (set in the app's plan file).
- Hosting: one shared 1 GB Hetzner server running Caddy and a separate service per app, listed on Hetzner's cloud page for a few euros a month excluding VAT. Backups run with restic.
Size today: 440 TypeScript files under src, about 33 000 lines (generated route tree included), and 57 test files covering unit, end-to-end, and mutation tests.
Days and cost: what we can and cannot say
The calendar above is exact. The effort is not: the build was done by agents, not people, and we did not track hours per feature. The first build log notes roughly 23 minutes in its last context window and does not include earlier work. So we do not publish a “build cost.” We publish what can be verified: 325 commits between 3 and 16 Nov, with the busiest day on 13 Nov (52 commits).
For a client, the relevant picture is different. A quiz app that includes accounts, live play, payments, and ten languages is much more than a first version needs. Our own v1 only handled building, running, displaying, answering, scoring, and a few question packs. The additional features were added after launch.
What went wrong
- Bugs the builder missed. The first build review uncovered 8 more numbered bugs beyond the judge’s 3. A quiz save containing a broken shape was still accepted. The test pages
/dev/outboxand/dev/clockcould be accessed on a live server unless a specific flag was set. A quiz that stopped on its auto-timer did not record the pass used. These issues were fixed on 4 November, in small separate commits. - A fix that broke a test. We shortened the sign-in period to 30 days, which caused a 31-day clock-jump test to fail. The sign-in duration was then extended to 90 days.
- Two deployments at the same time. On 13 November, two releases of the app were deployed at once. One deployment removed the other’s release folder, causing the live app to return errors for about 3 minutes. A deploy lock now prevents a second deployment from starting while one is in progress.
- A counter that gave the wrong number. Our internal visit counter showed 52 visits on 15 November, while PostHog recorded only 1. JavaScript-enabled crawlers and link previews inflated our counter. We removed the internal counter on 16 November and now rely solely on PostHog.
- Mail that never sent. The live app was configured with environment variable names that the mailer did not read, so sign-in emails probably never left until 16 November. The mailer now throws a clear error if the required settings are missing (commit 4d21d46).
- Running tests against an old build. An end-to-end test run scored 11 out of 18 because it used an outdated build output. After rebuilding, the same suite reported 26 passed and 3 failed. Rebuild before you judge test results (from our own project notes).
- Naming and payment changes. The product was called Quizhall until the owner chose the name Quizlamp on 4 November. A placeholder checkout in version 1 had to be replaced with a real seller account; the four payment checks performed in the live test were later discarded.
None of these came from the design. The separate judge caught the build bugs (item 1). The live system caught the mail, the deployment clash and the counter (items 3 to 5), which no review before launch would have found. That is why someone other than the builder checks every build, and why you watch the live app after launch.
What we would do for you
Take the plan: a written concept, a gate before any code, and a separate judge with hidden tests. Those three cost us little and caught the build bugs before launch, but not the problems that only showed up live. Skip the 10 languages until someone asks for one. Keep the first version to one loop that works end to end, put the full price in writing, and add after launch.
Built something with an AI builder and want it to hold up? Our App plan is EUR 990 and gives you a written scope and a fixed build quote. Get a fixed quote. See also how it works, pricing and the Quizlamp case.