DEV Community

muhamadrizki025
muhamadrizki025

Posted on

Buku Bon: building a voice-first ledger skeleton for Indonesia with Kiro (Kiroween 2026)

This post is written as part of the **Kiroween 2026 hackathon* (Skeleton Crew category).*

In Indonesia, a lot of money is tracked on paper or not at all. The warung (corner shop) owner remembers who bought rice on credit this week. The ojol (ride-hailing) driver guesses at night whether today's orders covered fuel. Nobody wants to type numbers into an app with one hand while serving a customer or waiting at a red light.

But everyone can say one sentence:

  • "Bu Sri ngutang lima belas ribu" — Bu Sri owes fifteen thousand
  • "isi bensin goceng" — fuel, five thousand (in street slang)

That is Buku Bon: a voice-first ledger skeleton, and two very different apps built on it.

One skeleton, two apps

The Skeleton Crew challenge asks for a template that is lean but flexible enough for two distinct apps. My rule was strict: an app is only a config file.

skeleton/      parser.js · store.js · voice.js · app.js · style.css
app-warung/    index.html + config.js   (Kasbon Warung)
app-ojol/      index.html + config.js   (Dompet Ojol)
Enter fullscreen mode Exit fullscreen mode
  • Kasbon Warung — a shop's credit book. Say who owes what; balances per customer update instantly. "Pak Budi lunas" settles his whole balance.
  • Dompet Ojol — a driver's daily net. "dapat orderan dua puluh tiga ribu", "isi bensin dua puluh ribu", "target hari ini 200 ribu", with progress and a 7-day chart.

A config declares the intents, their keywords and sign, and a summary view. This is the whole "debt" model of the warung app:

intents: [
  { id: 'utang', label: 'Utang', sign: 1,
    keywords: ['ngutang','utang','hutang','kasbon','ngebon','bon','pinjam'] },
  { id: 'bayar', label: 'Bayar', sign: -1, settle: true,
    keywords: ['bayar','nyicil','cicil','lunas','lunasin','setor'] },
],
Enter fullscreen mode Exit fullscreen mode

No framework, no build step, no runtime dependencies. Data stays in the browser, the app works offline through a service worker, installs as a PWA, and exports CSV. If the browser has no speech recognition, typing the same sentence works.

The hard part: Indonesian numbers, spoken

Speech recognizers return a mess of words and digits. The parser has to understand all of these as the same kind of thing:

  • words: dua ratus lima puluh ribu (250,000), satu juta dua ratus ribu
  • contractions: sejuta, seratus, setengah
  • digits and units: 15rb, Rp 15.000, 1,5 juta
  • slang: cepek (100), gopek (500), goceng (5,000), ceban (10,000), gocap (50,000)

It uses three accumulators (millions, thousands, the current group) and is checked by a property-based test: for random n, parseAmount(say(n)) === n, where say() speaks the number in Indonesian words. That one property found more edge cases than a page of unit tests would have.

How I used Kiro

Everything Kiro reads lives in /.kiro at the repo root.

Steering: the rule that shaped every change

Three steering files are always included: product.md (who the users are, speak first / type second, confirm before saving, Indonesian copy and rupiah), tech.md (plain ES modules, a pure DOM-free parser, Web Speech id-ID with typed fallback) and structure.md (the app-is-a-config contract).

The single most useful line was: "extend the skeleton generically instead of special-casing an app."

When I asked Kiro to add "Pak Budi lunas" (settle a whole debt) and a driver's daily target, it did not write if (app === 'warung'). It added two generic intent flags to the skeleton — settle and meta — and switched them on only in each app's config.js. That is exactly the design I wanted, and I didn't have to ask for it in the prompt; the steering did that.

Spec-driven: requirements, design, tasks

The spec in .kiro/specs/voice-ledger-skeleton/ has EARS-style requirements, a design with the parser algorithm and correctness properties, and tasks.md. Working from tasks instead of vibe prompts made each Kiro turn small and checkable. Kiro also kept the spec honest: while implementing it added R7 (offline), R8 (settle balance) and R9 (daily target metadata) and ticked the tasks it finished.

In Kiro CLI sessions, Kiro implemented tasks 10–14:

  1. Fast speech like "satu juta dua ratus ribu", with new test cases and the property test widened to 9,999,999
  2. An offline service worker
  3. The lunas settle intent
  4. The ojol daily target and 7-day chart
  5. UI polish: an SVG mic, a landing page and a PWA manifest

The condensed session logs with my exact prompts are in docs/kiro-sessions/ in the repo.

Debugging with Kiro: the collapsing bar

After the first turn, a headless smoke test showed the weekly chart drew no bar for today and the mic emoji rendered as an empty box. Kiro fixed both, but the re-test still showed every bar at the same tiny height.

I described the symptom and asked for the root cause in one line. Kiro's answer:

Percentage height on .bar collapsed to min-height because .day-bar flex column had no explicit height, making the percentage relative to nothing.

It fixed the CSS in the skeleton (keeping it generic) and re-ran the tests. Asking for the root cause, not just a fix, is a habit I'm keeping.

Hooks

The repo has two agent hooks in .kiro/hooks/: run parser tests on save and keep the skeleton generic (flag words like warung, ojol or utang inside skeleton/ and move them into config hooks). Hooks fire in the Kiro IDE; I worked in the CLI, so I ran the same loop by hand — asking Kiro to run npm test after every task and to do the generic-skeleton check. That check caught a leftover "tidak punya utang" ("has no debt") string in skeleton/app.js, and Kiro moved it into a settleEmpty hook in the warung config.

What changed in how I build

  • Write the rule once, not in every prompt. Steering turned "please keep it generic" from something I repeat into something the agent just does.
  • Small tasks beat big prompts. A task list gave every turn a clear finish line and a test to pass.
  • Properties over examples. For a parser, one round-trip property is worth dozens of hand-picked cases.

Honest notes

The first scaffold (parser, store, voice wrapper, both configs and tests) and the first drafts of the .kiro files were prepared earlier in the submission period with help from an AI assistant; Kiro CLI sessions then built tasks 10–14 on top. No third-party frameworks or starter kits are used.

What's next

A third app (a market stall or an arisan savings group) — which should be one new config.js — plus voice read-back of balances and WhatsApp-friendly reminders for customers who owe.

If you build something for small businesses in your own language, I'd love to hear how you handle spoken numbers.

#KiroweenHackathon26 — written for the Kiroween 2026 hackathon.

Top comments (0)