Prompt Arcade ($ARCADE): type a game, play it, share it, remix it

Live figures
1st	Price	—
2nd	Market cap	—
3rd	Volume 24h	—
4th	Liquidity	—
5th	Change 24h	—

1-2

Make a game

Level select

How your words were read

Click the screen first. Arrows or WASD to move; space or ▲ to jump, fire or flap.

Remix it in words

1-3

Can it be won?

Level select

Before anyone plays a generated game, a bot plays it — twenty times, each with a different level. If the bot never wins, the game is probably too hard as described. The bot is simple; a person may do better or worse.

Bot wins by difficulty, 20 runs each

Press “Every difficulty” to fill the meters.

1-4

Starter cartridges

Level select

One starter for each of the eight genres. Pull one off the shelf to load it into the maker, then remix it.

1-5

Technical appendix

Level select

Instruction booklet · Appendix A

Reading a Game From a Sentence: A Rule-Based Compiler With Its Own Playtester

The promptplay contributors · promptplay 0.1.0, MIT · the figures on these pages are computed from the library on this page

A.1 Summary

promptplay turns a typed description into a small game specification — one of eight genres, the player, the thing to avoid, the thing to collect, the goal, lives, difficulty, speed, gravity and colours — runs it in a canvas engine, packs it into a share code, accepts plain-English edits, and lets a bot play it to see whether it can be won. Every decision is reported with the words that caused it, and every default is listed. On 40 held-out prompts written before the rules, all five parts were right on 34 and the genre on 39.

A.2 Background

On 7 October 2026 Google Labs released Playground, an experimental web platform where people describe a game in a chat box, play it at once, change characters, rules, physics and scenery over several turns, and keep it private, send it as a link or publish it to a public gallery [1, 2]. It runs on Google's models, launched in the United States for adults, and gives game generation a weekly allowance by membership tier; a Unity integration is in testing [1, 2].

Not connected. promptplay, this site and $ARCADE have no connection with Google, Unity or the publications cited, and none of them has seen this appendix. promptplay uses no AI model at all: it is a set of readable rules, which is what makes its decisions explainable — and what limits them.

A.3 Reading the sentence

The prompt is split into words; plurals are made singular (mice → mouse, cherries → cherry). Each genre has words that count for it with a weight, and the highest total wins; a tie is reported as a guess. The roles come from cue words: avoid, dodge, shoot, jump over, stomp, or a thing that chases, roams, patrols, marks the enemy; collect, grab, eat, catch, pick up marks the item; as a, help the, pilot a, with your, is a marks the player, and otherwise the player is the first thing named. Shots after a shooting verb belong to the player; "waves of goblins" are goblins.

A.4 Playing it

Each genre is a small engine on a 480 × 320 screen with a fixed step and a seeded random generator, so a share code always plays the same level. The same code runs without a screen for testing and for the playtest bot, which is a few lines of plain rules per genre (jump at the edge, path to the nearest item while keeping two cells from an enemy, aim where the target will be).

A-1

A.5 Evaluation

Table A1. Parts read correctly.
Set	Prompts	Genre	Player	Enemy	Item	Goal	All five

The development prompts were used to write the rules. The held-out prompts were written first, their SHA-256 recorded, and they were scored once; the rules were not changed afterwards. The misses, as they came out: ….

On the held-out games the bot won at least once on 38 of 40, and … of 400 runs overall. It never won the long shooter survivals (90 seconds, two minutes): the spawn rate climbs and the bot's aim is plain. Those games carry the verdict "probably too hard as described" — the warning working as intended, not proof that no person could win.

Table A2. The held-out games the bot found hardest (wins of 10).
Prompt	Genre	Wins

A.6 Limitations

It knows the words it was written for: open vocabulary works for names, but unfamiliar grammar can put a word in the wrong role, as the misses show. Eight genres, single player, 2D, simple shapes for sprites, English only. The rules were written by the same people who wrote both prompt sets. The bot measures "can be won", not "is fun".

R References

J. Leon, Android Headlines, 7 October 2026: Playground lets you create games using simple text prompts. androidheadlines.com
D. Horetski, Dexerto, 7 October 2026: a new AI platform lets you make games just by describing them. dexerto.com
promptplay 0.1.0, the library on this page (MIT).

A-2

1-6

The ROM

Level select

promptplay 0.1.0: one file, no dependencies, Node 18 and later and every modern browser. MIT licence. Every file of the repository is a bank in the ROM map below, with its size and SHA-256.

The repository
Bank	Address	File	SHA-256

Downloads

Quick start

const P = require("./promptplay.js");
const { spec, why } = P.compile("a snake game, eat 15 apples");
const code = P.encode(spec);              // "pp1.…"
P.remix(spec, "make it faster").changes;
P.playtest(spec, { runs: 20 }).verdict;
P.play(canvas, spec);                     // in a browser

node bin/promptplay.js compile "catch falling stars with a net"
node bin/promptplay.js playtest "a bird flapping through gaps" --runs 20
npm test

The API