AI game prompt generator
Describe the game you want. This writes the prompt to paste into Claude Code — one that gives the agent a real quality bar it can't talk its way around, an independent critic that judges the running game against that bar, and a loop that keeps going until yours wins.
- No email
- No account
- No daily limit
- Runs in your browser — your brief never leaves it
Your prompt
Read the four steps below before you paste. The prompt assumes all four — it will underperform badly without them.
How to run it
- Use an agent, not a chat window. It has to open files, run the game, take screenshots, look at them, and spawn other agents. Claude Code with Opus 5 is what we use. A chat window can write code but cannot look at the result, so it cannot run the loop at all.
- Turn the effort up.
/effort→ultracode. On long multi-agent runs this is the difference between "pretty good for AI" and something you'd show someone. - Start in a clean session. Matt Shumer's write-up reports that installed skills and MCP servers hijack the loop and make the prompt work dramatically worse. We haven't measured that ourselves — flagging it as his finding, not ours.
- Paste it and leave it alone. The prompt tells the agent to keep a live progress page. Watch that instead of interrupting — every time you steer mid-run you overwrite the judgment you're paying for. Stop when you love it.
Where this prompt differs from the famous one
The method here is the Gauntlet Loop, published by Matt Shumer after Claude of Duty — a ~55,000-line browser FPS built from a single prompt. We adopted it here on 2026-07-27 and it changed how every game on this site gets built. Credit where it's due: the core idea is his.
We have since run the loop on 19 games here, writing every round's score down. Three things in the prompt above are deliberately not what the famous one says — and the first correction isn't even ours:
| Change | Why, and who measured it |
|---|---|
| One sequential owner per coupled concern — not "fan out sub-agents". | His own repository disproves his most-quoted line. The viral prompt says "Fan out sub-agents and have sub-agents tackle each one individually." The Claude of Duty README then reports that three rounds of six agents each owning a directory moved the score +0.46 and left defects higher than they started (60 → 47 → 66), while one sequential pass with a single owner per coupled concern moved it +1.00 and cut defects 66 → 26. Those are his numbers, not ours — and his method page never mentions them. Split by what interacts, not by what sits in separate folders. The independent critic is what adds the lift; parallel building is what spends the budget. |
| Capture the bar past its gate. | Ours, and it cost us two wasted rounds to learn. Twice we screenshotted a competitor's sign-up modal instead of their board, and the critic correctly judged an empty rectangle. A capture that lands on a gate is evidence about their onboarding — it is not the bar, and a win against it is not a win. |
| Revert on a score drop. | Ours. A round that scores lower than the one before it gets rolled back, not built on. Iterating on top of a regression is how a loop talks itself into believing it is improving. |
| Measure the frame rate every round, not just the look. | Ours, and it is the one the critic cannot catch — a screenshot of a game running at 15fps looks identical to a screenshot of one running at 60. So we loaded 46 of these games in a real browser and measured them: 29 of the 39 we could instrument run slower than 33ms per frame, a median of 66.6ms — roughly 15fps. The worst was 1,433ms per frame, and it is a game that looks completely fine in a still. A critic panel would have passed most of them. |
What we found when we measured 46 of them
These games are all built with this method and hosted by their creators. We ran every live one through a headless browser at 1440×900 and 390×844, nudged past the menus, and recorded what it actually did. The interesting part is which assumption died:
| Measured | Result |
|---|---|
| Frame time (p50, in play) | Median 66.6ms · only 10 of 39 held 50fps or better · worst 1,433ms |
| Console clean on arrival | 14 of 39 ship visible errors |
| First-load weight | Median 475KB, but three exceed 5MB and one is 23MB |
| Renders on a phone | 38 of 39 — this is the assumption that died, see below |
We expected mobile to be the weak spot and it isn't. Only one game failed to render at 390×844 and only two overflowed horizontally. The real, and much less visible, failure is frame rate — which no screenshot, and therefore no visual critic, can see. That is why the generated prompt above asks for a measured number every round instead of only a judged one.
Honesty on the instrument: 7 of the 46 blocked the browser's main thread long enough that our own debugger call timed out, so they are excluded rather than counted as failures — that is our limit, not a verdict on those games. Three games under-report their weight because cross-origin assets do not expose transfer size. Frame time is sampled in a ~10-second window after an automated nudge, so a game whose menu we never got past is measured on its menu.
One more, and it is the reason a weak bar is worse than no bar: most free games are not good enough to be a target. Beating the best free version of a genre is a floor, not a finish line. Where the incumbent is weak, tell the agent to also generate a "northstar" image — what this game would look like if it were the best version of itself in any medium — and judge against that. Never let it trace or ship a pixel of a generated image; it answers how good, never how it's built.
When it's playable, give it somewhere to live
A game in a chat window dies when the tab closes, and a game on a free host dies when the free host does. Upload the HTML here and it gets its own permanent page in the arcade — playable in a sandboxed frame, your name on it, free, and you keep every right to it.
Questions
Does this really not want my email? Really. The prompt is assembled by JavaScript on this page, so there is no cost on our side to cover and nothing to meter. Your brief is never transmitted anywhere — open devtools and check.
Why isn't an AI writing the prompt? Because for games the hard part isn't the prose, it's picking a bar the agent can actually inspect — and this is a games site, so we can just name one. A model guessing at a comparison it has never opened is the weakest link in the whole method.
Will this build a good game on the first try? No, and neither does the original. It builds something mediocre, judges it honestly, and then keeps closing the largest gap until you stop it. The output quality is mostly a function of how long you let it run and how honest the bar is.
I don't have Claude Code. Then start with the plain single-file workflow — copy-paste prompts for Claude, ChatGPT or Gemini, no agent required. It won't reach the same bar, but it finishes a game in an afternoon.