You don't need to code. You need a small idea, one magic phrase, and the patience to playtest. This is the workflow behind every original game on this site — the same one that built MOGUL and LEXEME — scaled down to one sitting.

Step 1 — pick an idea smaller than you want to

The graveyard of AI games is full of open-world MMOs that died in message four. Pick one mechanic you can describe in one sentence: "dodge falling rocks that speed up", "merge numbers to reach 2048 but hexagonal", "type the word before it hits the ground." You can always add to a working game; you can't play a pile of ambition.

Step 2 — the magic phrase: one self-contained file

Whatever AI you use, the request that makes your game portable is:

"Build this as a single self-contained HTML file — all CSS and JavaScript inline, no external libraries, no CDN links, no network requests. It should work opened from a file with no internet."

Without it you get code split across files or importing libraries from the internet — which breaks the moment the game leaves the chat. With it, your entire game is one file you can save, send and host anywhere.

A fuller starter prompt to copy:

"Make a browser game: [your one-sentence idea]. Single self-contained HTML file, all CSS/JS inline, no external requests. Canvas rendering, keyboard AND touch controls, a start screen with the title and how to play, a score, and a game-over screen with a restart button. Make it feel juicy — screen shake, particles, sound with the WebAudio API."

Step 3 — playtest like a designer, not a coder

The first version will run and feel wrong. That's normal — the difference between a demo and a game is the next 3–5 messages. Describe what's wrong in feel terms, not code terms:

"The jump floats too long — make it snappier."
"I died and didn't understand why. Make hits more obvious."
"It's too easy until it's suddenly impossible — smooth the difficulty."
"I can't tell whose turn it is."

The AI can fix anything you can name. It just can't feel it for you.

Step 4 — get the file out

Claude: the game runs in an artifact panel — use the download button to save the .html file directly. (More on Claude games →)
ChatGPT: ask for "the complete final file", copy it into a text editor, save as mygame.html. Watch for truncation on long files. (More →)
Gemini: same as ChatGPT — always ask for the complete updated file after edits, never assemble snippets by hand. (More →)

Sanity check before you call it done: the file opens from your desktop with no internet and ends with </html>.

Step 5 — publish it (the part everyone skips)

A game only exists if other people can play it. Chat share links expire, require logins, or aren't playable. Your options:

Upload it to the EGM AI arcade — free, no account. A human reviews it, then it goes live on a permanent page with your name and link, in a directory search engines index.
Host it yourself — GitHub Pages, itch.io and Neocities all take a single HTML file. More setup, more control.

Do both if you like — you keep all rights either way.

Common failure modes (learned the hard way)

The 90%-done trap: "add multiplayer" is where solo projects go to die. Ship the single-player version first.
Base64 bloat: AI-generated images embedded in the file balloon it past hosting limits. Ask the AI to draw with canvas and CSS instead — it's better at that anyway.
Desktop-only controls: half your players are on phones. Ask for touch controls from the first prompt, not as a retrofit.
Rewriting instead of fixing: if the AI keeps breaking things it already fixed, say "change ONLY the jump physics, keep everything else identical."

Made one? Upload it — free hosting →