---
title: "You can build a game from a single prompt. I did, and here's how."
date: 2026-07-19T08:31:50.666+00:00
url: https://hurricane.works/blog/build-a-multiplayer-game-from-a-single-prompt
description: "I pointed an AI coding agent at a plan and said 'build this project'. Out came a complete multiplayer browser game you can play right now. Here is what a single prompt can really do, and the one thing that makes it work."
author: "Hurricane Works"
categories: ["AI", "Engineering"]
---

# You can build a game from a single prompt. I did, and here's how.


I typed "build this project" into an AI coding agent and pointed it at a plan file. A few hours later I had a complete, playable, multiplayer game running in the browser. It is called SALVO, it is live right now, and you can play it here: 

https://salvo-al.netlify.app

![YouTube video](https://www.youtube.com/watch?v=3O-yuzg53ok)

You create a game, send a friend the link, they tap it, and the two of you are playing battleships against each other. No signup, no app install, nothing to download. If there is no willing opponent to hand, there is a solo mode against three difficulties of bot and a daily puzzle you can try on your own in about ten seconds. The whole first version was built in one shot, without me writing a line of code.

That sounds like a party trick, and half of it is. The other half is genuinely useful, and the gap between the two is the entire point of this post.

## How to play

Open the link on your phone and tap PLAY A FRIEND. A game link pops up: send it to someone over WhatsApp or iMessage. They tap it, you both drag your five ships onto the grid (or just hit shuffle), and then you take turns firing at each other's hidden fleet. First to sink all five enemy ships wins. Games are asynchronous, so nobody has to be online at the same time: take your shot, close the tab, and pick it up again whenever your opponent has moved. Want to warm up alone first? PRACTICE VS BOT or the DAILY PUZZLE will do it.

## What "single prompt" actually means

The prompt really was one line. But it was aimed at a plan I had written first: a couple of pages that spelled out the game, the data model, the rules the server would enforce, the visual style, and, crucially, an ordered build sequence with a check after every stage. The setup section ended with a line to the effect of "everything from here is executed without further human input."

So yes, one prompt. The leverage was not the prompt. It was the plan.

This matters, because the popular fantasy of "AI builds your app" is that you type a wish and receive a product. That is not what happened, and if you try it that way you get a pretty toy that falls over the second two real people touch it. What actually works is closer to briefing a very fast, very literal contractor: you decide the hard things yourself, write them down precisely, and then let the machine do the enormous amount of typing.

## What one prompt built

Working straight down the plan, the agent produced, in order:

- The docs first: an architecture note, the game rules, the data model, and a running decision log. Writing the design down before the code is not box-ticking; it is what lets a single long pass stay coherent with itself.
- A database schema with row-level security from the start, so a player can never read their opponent's board.
- A pure game engine (board maths, ship-placement validation, shot resolution, three bot strategies including a probability-density hunter, and a date-seeded daily puzzle) with its own unit tests. Sixty-odd of them, all green, before a single screen existed.
- The server logic, written so the server is the referee: the browser never holds the opponent's ship positions, so there is nothing on the client to cheat with.
- The entire front end from a locked visual design: landing, invite, ship placement, the game board with its flip-to-your-own-fleet animation, results, daily puzzle, leaderboard.
- And then it took screenshots of its own screens at phone size, compared them against the design mockup, and fixed what had drifted before calling itself done.

That last step is the one people do not expect. The agent did not just write the interface; it looked at the interface it had written and corrected it against a reference.

## What playing it added

The one-shot build was complete against the spec. It was not finished as a product, and that difference only showed up when real people started playing. The second phase is where most of the character got added, and none of it was in the original plan:

- **Live video between the two players.** You both tap a camera icon and can see each other's faces right there in the game, your opponent's feed growing larger while they line up their shot, so you get to watch them sweat. It runs as a direct peer-to-peer connection between the two phones.
- **"Your move" nudges.** Because games are asynchronous, the app can notify you when it is your turn, and falls back to firing off a ready-made WhatsApp message if on-device notifications are not available.
- **Never losing a game.** You can wander off mid-battle and a "back to battle" button drops you straight back where you were, or you can resign cleanly if you want out.
- **Readable combat.** Big banners tell you at a glance whether an incoming shot hit or missed, because a subtle splash animation turned out to be far too easy to miss in the heat of it.

Every one of those came from playing, not from planning. The plan got me a demo. Playing it got me a game.

## How you can actually do this

You do not need to be a game developer, and you barely need to be a programmer, but you do need to spend an hour thinking like a designer first. The recipe:

1. **Write the plan, not the prompt.** Describe what you are building, the data it needs, the rules the server must enforce, and how it should look, in enough detail that a literal reader could not go wrong. Anything you leave vague, the machine will decide for you, and not always the way you hoped.
2. **Put the build order in the plan, with a check after each step.** "Write the tests and make them pass." "Screenshot it and compare to the design." The agent follows the sequence, and those checks are what keep a long single pass honest instead of drifting into plausible nonsense.
3. **Lock the look before you build.** I had a visual mockup designed and frozen first. "Make it nice" produces mush. "Reproduce this exact design" produces the design.
4. **Then, and only then, say "build this".**

Do that and the single prompt stops being a trick and becomes a real force multiplier. The thinking stays yours. The typing, all of it, does not.

## The honest headline

It is not "type one sentence, get a finished game." It is "spend an hour on a precise plan, and one prompt will build you a real, working first version in an afternoon, which you then finish by playing it." Which, when you sit with it, is still a faintly absurd thing to be able to do.

Write the plan. Then go and play the thing it built: https://salvo-al.netlify.app