> ## Content Index
> Fetch the complete content index at: https://www.siliconwok.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# How I built a hawker policy paper and a game with AI
- URL: https://www.siliconwok.com/how-i-built-a-hawker-policy-paper-and-a-game-with-ai/
- Published: 2026-11-02T13:13:22.000Z
- Updated: 2026-11-02T23:37:28.000Z
- Description: Seven days, thirteen library releases and a playable hawker-centre game, built with a swarm of AI agents and one human. What worked, what broke, and what I would tell you.
- Author: Uncle James
- Tags: Essays, Hawker

## A plate of char kway teow, and a deadline I set myself

Let me start with the key metrics, because they still surprise me. In seven days (27 October to 2 November 2026), I shipped thirteen releases of a 345-page policy paper on saving Singapore's hawker heritage. Seven days ago, I had zero context on the state of our hawker culture. In the same week, beyond a first cut of the plan, I also built the first playable slice of a game about running a hawker centre. It runs at 60 frames a second on my phone.

I did not do this with a team of policy makers, the way it would have been done when I used to work in the Singapore government. I did it with a swarm of AI agents, and one human: me.

I care about our hawkers because the trade is ageing faster than it is renewing. The median hawker was 60 years old in 2021\. I grew up on hawker food and have fond gustatory memories of hawker centres and their uncles and aunties that feed our nation. I wanted a plan that was honest about the money, and a way to make people feel the trade-offs. Writing a paper the old-school way would have taken me months, and a team of like-minded citizens I didn't have on hand. So I tried something else. This is an account of how it went, what broke along the way, and what I would tell anyone thinking of doing the same.

![Rows of hawker stalls with lit signboards and diners at tables under a steel roof at Alexandra Food Centre, Singapore](https://www.siliconwok.com/content/images/2026/11/alexandra-food-centre.webp)

Alexandra Food Centre, Singapore. Photo: [Exec8](https://commons.wikimedia.org/wiki/File:Singapore%5FHawker%5FAlexandra%5FFood%5FCentre.jpg?ref=siliconwok.com), [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/?ref=siliconwok.com), via Wikimedia Commons.

## The setup: one lead, many hands, and Jamie

If you've been giving AI a proper go, as some of us builders have over the past few months, you'll recognise the pattern I'm about to describe. It worked better here because of practice on smaller past projects that went nowhere. It took me several weeks to get right.

1. **I decide.** I set the goal, the rules and every judgement call: what the plan asks of Government, how the game scores a steward, what is fair to a hawker.
2. **Claude leads.** One Claude session acts as the lead. It plans the work, writes the briefs and checks what comes back.
3. **Workers build.** The lead hands each piece to a worker – another, cheaper model, including the DGX Spark pairs in my AI homelab – on its own branch of the code repo. A stronger model takes the work that needs judgement. A cheaper one takes the mechanical work: rebuilds, test runs, version stamps.
4. **Jamie ships.** Jamie is my own OpenClaw agent, on a small server at home. I've been tuning her harness since February. Only Jamie merges and deploys. The AI that writes the code never pushes it live.
5. **I review on my phone.** I read the live site and PDFs and play the live build, then send back feedback and what I notice.

Two rules held throughout. Every commit is signed as me, so I own every line. And nothing goes live without my say.

## The library: from one paper to eleven documents

It began as a single 45-page paper. Version 1.0 went up on 27 October. By 1 November it was version 1.9.3: thirteen releases in six days.

It is now a library of eleven documents and 345 pages, refactored for easier reading:

- **The front:** an Executive Summary and The Plan.
- **Seven volumes, A to G:** who already helps hawkers, the budget, social protection, where a hawker meal's dollar goes, succession, the method, and a technical paper.
- **The back:** Sources, a short Your Table for diners, and a change log that records every edit and why.

The site at [www.siliconwok.com/heritage](https://www.siliconwok.com/heritage/) has 198 pages, an intro film, and two models you can play with. One follows a plate's dollar from the diner to the wok, built from public data where it exists and inferred where there are gaps. The other tests whether the plan's funding holds.

![The Hawker Heritage home page, headed Saving Singapore's Hawker Heritage, with buttons to read the plan and follow the plate](https://www.siliconwok.com/content/images/2026/11/hawker-library-home.webp)

The library's home page at [www.siliconwok.com/heritage](https://www.siliconwok.com/heritage/).

![Follow the Plate bar chart splitting one dollar of sales by where the money goes, for seven kinds of stall from a subsidised NEA stall to a mall food court](https://www.siliconwok.com/content/images/2026/11/follow-the-plate-dollar.webp)

Follow the Plate: one dollar of a hawker's sales, split by where it goes, for each kind of stall.

The AI did the heavy lifting: research, drafting, the models, the typesetting and the checks. Every figure is sourced or marked as an assumption. My job was judgement. I scaled a detailed ballot scheme back to a short note. I made the plan start at one hawker centre, not ten. And on 1 November I asked for a tone pass, because I am talking with NEA's Hawker Centres Group and I would rather change things from within than make enemies by being careless with my tone. The facts stayed; the barbs went.

Opus 5.5 is very good, with occasional use of Fable 5.1\. Astra 6 is alright, but Sol 6.1 is not good enough except as a baseline worker. My locally hosted DeepSeek v4 and Qwen 3.8 Flash Next were relegated to straightforward work. I didn't want to experiment too much on token efficiency. I preferred to sprint hard and spend on frontier cloud models in exchange for progress.

## The game: Full House

A policy paper tells you the trade-offs. A game makes you live them. I grew up on simulation games such as SimCity, Sim Theme Park, Theme Hospital and Cities: Skylines, and wondered what a hawker simulator would be like. So on 30 October I started work on Full House: you run a hawker centre for 20 years, one day at a time.

The first milestone was the simulation itself, with no graphics. It models diners by age and income, menus, queues and seats, rents and ledgers, a business cycle, inflation, government vouchers and Parliament's yearly decisions. Hawkers have households, tempers, faiths, and good and bad weeks. At least one stall in every game becomes famous, and keeping it alive is on you. And yes, of course, I couldn't resist putting myself in the game. Uncle James opens a robot-assisted stall, Silicon Wok, when a stall first falls vacant, some time after the first year. Ignore him and he is friendly. Fight him and he becomes the national debate.

The second milestone put it on screen: an isometric hall built in three.js, with ten stalls, day and night, rain and haze, and crowds driven live by the simulation. On 2 November both milestones were merged and tagged v0.1-m1\. That morning I opened a private preview on my iPhone, and it held 60 frames a second at the lunch rush. I continue to test the graphics at low, medium and high settings.

![Full House seen from the far zoom: the hall's roof opens over the seating, where diners queue, carry trays and eat at the lunch rush](https://www.siliconwok.com/content/images/2026/11/fullhouse-hall-isometric.webp)

The hall at the far zoom. The roof opens over whatever part of the hall you are looking at when you’re zoomed out.

None of the art was drawn by hand. The hall, the figures and the signs are all generated by code. Every name is invented, every hawker is a person with dignity, and no community is ever the joke.

![Full House at dusk: tube lights on over emptying tables, a cleaner standing by the tray return](https://www.siliconwok.com/content/images/2026/11/fullhouse-hall-dusk.webp)

Dusk in Full House. The light follows the simulation's clock, and so do the rain and the haze.

## What worked, what broke, and the rules I set

Granted, it was not always smooth sailing. Three things broke in ways worth telling.

- **The buttons that did nothing.** On my iPhone, tapping a choice card did nothing. The cards were redrawn sixty times a second, so the button changed under my finger. Every automated test had passed, because a test 'clicks' faster than a thumb. It took a real phone to find it.
- **Rough edges on screen.** People walked through each other. Every table was a four-seat square, where real centres mix rectangles, squares and rounds seating four to six or more, packed closer together. The simulation knew where everyone sat; the screen did not stop them passing through one another on the way. After my play-through, routing and personal space are now built in, and a denser hall is next.
- **A score that looked right but wasn't.** The balance bench said doing nothing scored 41 out of 100\. Real 20-year games said about 30\. The bench had been tuned on the first three years only. We are tuning again, on the long games.

![Two choice cards on a phone in Full House, Exam season and Seat turnover falls, each with a let pass button](https://www.siliconwok.com/content/images/2026/11/fullhouse-phone-choice-cards.webp)

The choice cards on a phone, day 1: the buttons that wouldn't take a tap, now fixed.

In all honesty, I'm probably being too demanding. AI has come a long way in the past six to nine months, and I'm probably still prompting it wrong. What worked was the rules. I wrote them down once, and every worker reads them first:

1. **Invent every name.** No real stall, official or brand. The one exception is my own pen name.
2. **Dignity first.** Hardship, crime and gambling are never comic. Faith and community never decide a character's traits.
3. **Even-handed politics.** Quote what was said, attribute it plainly, take no side.
4. **No secrets in chat or code.** Ever.
5. **The AI writes; Jamie ships; I decide.**

## What it cost, and what I would tell you

I spend an average of 3 to 4 hours a day working with my team of AI agents, on this project alone. It has taken an entire week's worth of my Claude Ultra plan (once called Max 20x) and US$600 of extra usage, plus a smaller amount of Astra 6 and Sol 6.1\. I don't have a good tracker for my local models yet, and they still carry other recurring work in my AI homelab.

The honest answer is that most of my time went into reviewing what the agents proposed and deciding, not into keystrokes. I read, I judged, I said yes, no or maybe. The AI is fast. It is also confident when it is wrong. That makes the person in the chair more important, not less.

If you want to try this, here is what I would tell you:

1. **Write your rules before you write anything else.** The agents follow them better than people do, and they never get tired of reading them.
2. **Keep one person who can say no.** For me, that is me. Nothing ships without it.
3. **Spend the clever model on judgement, and the cheap one on chores.** Most of the work is chores.
4. **Test on the real thing.** My phone found what every automated test missed.
5. **Let a separate agent ship.** The one that writes the code should never be the one that pushes it live.

## What comes next

The game's next milestone is a demo you can play from the first morning to the monthly report. After that come sound, polish and a public release at www.siliconwok.com/fullhouse. Down the road, I want to remake the hawker library's intro film with real footage from the game: a hall that fills, a famous stall that sells out, and a robot stall that won't go away.

I am not doing this for profit, and I do not have the final answers. I have a plan, a game and a week's proof that one person with good tools can move quickly. I'm tired of complaining about the dying hawker trade. I want to apply my skills as a forward-deployed engineer – someone who builds with the tools, not just talks about them – to a constructive cause.

If you love your hawker stalls, read the plan and tell me where it is wrong. If you can build, cook or teach, write to me at uncle@siliconwok.com. Join the movement, and let's drive it together.