How to Make a Game With AI Prompts (The Prompts That Work, 2026)

You have probably seen the demo: someone types "make a platformer" into an AI, and a few seconds later a little game appears. It looks like magic, and then you try it on your own idea and get something that is almost right but not what you pictured, and you have no idea which words to change.

The problem is rarely the AI. It is the prompt. Making a game with AI is a real skill, and most of that skill is knowing what to type, in what order, and how to react when the result misses. This guide is about the prompts specifically: the anatomy of one that works, the patterns you reuse all day, the copy-paste examples, and the prompts that quietly waste your turns.

If you want the full idea-to-build walkthrough instead, the step by step guide covers the whole loop. This post zooms in on the part everyone gets wrong first: the prompting itself.

{/* IMAGE: Hero. A single chat prompt on the left ("make the player jump when I press space") and the running game on the right showing the jump. Clean editor screenshot style, 1200x630. */}

The one rule that fixes most bad prompts

There is a single mistake behind nearly every disappointing AI game result: asking for the whole game in one prompt.

"Make a fun 2D RPG with combat, a shop, and quests" feels efficient. It is the worst thing you can type. That prompt asks the AI to make a hundred design decisions you never specified: how combat works, what the player controls, what a quest is, how the shop opens. The AI guesses all of them, and when the result is wrong you cannot tell which guess caused it, so you cannot fix it.

The rule that fixes this: one prompt, one mechanic. You build a game out of many small, specific prompts, running the game after each one. This is slower to describe and far faster in practice, because every prompt produces something you can see, test, and correct before moving on. Everything else in this guide is built on that rule.

The anatomy of a prompt that works

A good game prompt has three parts, even when it is one sentence long:

1. The target. What does this prompt affect? The player, an enemy, the score label, the camera. Name it.
2. The behavior. What does it do? Move, jump, spawn, take damage, change a value.
3. The trigger and result. When does it happen, and what happens? On space, jump. On contact, lose a life. At 100 points, win.

Put together: "Make the player jump when I press space." Target (the player), behavior (jump), trigger and result (on space, it jumps). The AI has nothing to guess, so it builds exactly that.

Compare it to "make the controls feel good." There is no target, the behavior is a feeling, and there is no trigger. The AI will invent something, and it almost never matches the feeling in your head. Whenever a prompt is not working, check it against these three parts. Usually one is missing.

Behavior converts. Mood does not.

This is the distinction that separates people who get what they pictured from people who keep fighting the AI.

Behavior is anything with a clear trigger and result. "Enemies move left until they hit a wall, then reverse." "The door opens when the player has the key." "The timer counts down from 60 and ends the game at zero." The AI converts these cleanly because there is nothing ambiguous.

Mood is "make it tense," "make it cozy," "make it challenging," "make it juicy." These are real and they matter, but they are outcomes of mechanics, art, and pacing, not instructions the AI can build directly. Type them as prompts and you get a literal, usually wrong, interpretation.

The fix is to translate mood into behavior yourself before you prompt. "Make it tense" becomes "the enemy speeds up when it is within three tiles of the player." "Make it juicy" becomes "shake the camera for 0.1 seconds and flash the enemy white when it takes a hit." That translation is the actual game design work, and it is the part the AI cannot do for you. Numbers help enormously here: "slowly" is a guess, "three times slower than the player" is an instruction.

The five prompt patterns you will reuse all day

Most of building a game is four or five kinds of prompt, used over and over. Learn these and you have the whole vocabulary.

The build prompt adds one new mechanic. Target, behavior, trigger, result.

"Add coins that the player collects by walking into them, and increase the score by one each time."

The tune prompt changes something that already works. It references the current behavior and gives a direction.

"The jump feels floaty. Make the player fall faster after the top of the jump."

"The enemies are too easy. Make them move 50 percent faster and spawn one every two seconds instead of every four."

Direction matters. "Make the jump better" is a guess. "Make the player fall faster after the peak" tells the AI what to change and which way.

The fix prompt reports a bug like you would to a developer: what you did, what you expected, what happened instead.

"I pressed space while in the air and the player jumped again. It should only jump when standing on the ground."

The asset prompt generates art or sound. Be specific about style, view, and use.

"Generate a 16 by 16 pixel art coin, gold, front facing, for a 2D platformer."

"Generate a short looping background track, calm and a little mysterious, for a forest level."

The structure prompt sets up a scene or system before you fill it in.

"Create a main menu scene with a Play button that loads the level, and a Quit button."

Almost everything you build is a sequence of these. You are not searching for one magic prompt. You are stringing together small, specific ones in an order that lets you check each result.

{/* IMAGE: A labeled grid of the five prompt types, each with a one-line example and a tiny screenshot of its result. 1000x1000, illustration. */}

A real game, prompt by prompt

Here is a full tiny game built the way it actually works. The plan, in plain words, is a Flappy Bird clone: a square that jumps, pipes that move, death on contact, a score. Five prompts.

Start from a platformer template so a moving player already exists, then:

"Make the player jump upward a fixed amount each time I press space, and fall back down with gravity."

Run it. The square flaps. Behavior matches the prompt.

"Add pipes that spawn at the right edge and move left at a steady speed, with a gap the player can fit through."

Run it. Pipes scroll. Matches.

"End the game and show a Game Over label when the player touches a pipe or the ground."

Run it. The death rule fires. Matches.

"Add a score in the top corner that goes up by one each time the player passes a pipe."

Run it. The score ticks. Matches.

"When the game is over, show a Restart button that resets the player, pipes, and score."

Run it. You now have the loop end to end.

Notice what each prompt has: a target, a behavior, a trigger and result. Notice what you did between each one: you ran the game and confirmed the behavior before adding the next. When the third prompt produced death on touching a pipe but not the ground, you knew instantly it was the third prompt to fix, because nothing else changed. That traceability is the entire payoff of one prompt per mechanic.

Prompts that waste your turns

These are the prompt shapes that consistently disappoint, and what to write instead.

The whole game prompt. "Make a roguelike with procedural levels, items, and permadeath." Too many decisions, no traceability. Write the first mechanic only: "Generate a room with four walls and a door, and place the player inside."

The mood prompt. "Make it feel epic." Not convertible. Translate it: "Add a slow zoom on the boss when it appears, and swap to a louder music track."

The vague tune. "Make the combat better." The AI does not know what better means to you. Name the change and the direction: "Make the attack hit a wider area and add a short pause when it lands."

The kitchen sink fix. "Nothing works, fix everything." The AI cannot see what you see. Report one bug at a time, with what you expected and what happened.

The re-paste-louder. Retyping the same prompt with more emphasis changes nothing. The words were the problem, so rewrite the behavior more specifically, usually with numbers and a clear trigger.

How to react when a prompt misses

It will happen often, and the move is always the same. Do not argue with the AI, and do not re-paste the same prompt. Ask one question: which of the three parts was unclear?

If the wrong thing got affected, the target was vague. If the behavior is roughly right but feels off, you described a mood, not a mechanic, so add numbers. If it does the right thing at the wrong time, the trigger was loose. Rewrite that one part and run again. Because you only built one mechanic in that prompt, the fix is one prompt too. This tight loop, prompt to run to correct, is the real rhythm of making a game with AI, and it is much faster than it sounds.

What better prompts will not do

Being honest about the ceiling saves a lot of frustration.

Prompts do not decide if your game is fun. You can write perfectly clear prompts that build a technically working game nobody enjoys. Only playtesting tells you that, and only you can act on it. No prompt produces taste.

Prompts do not manage scope. The AI will happily start building an open world MMO one clear prompt at a time and never warn you it is a multi year project. Scope lives in which mechanics you choose to prompt, not in the prompts themselves.

Prompts do not read intent. The gap between what you typed and what you meant is invisible to the AI. The clearer the prompt, the smaller that gap, but it never fully closes, which is why playtesting between prompts is not optional.

The people who build real games with AI did not find a secret prompt. They write small, specific, testable prompts, run the game after each, and rewrite the ones that miss. The AI makes every step fast. It does not make any of the decisions.

Write your first five prompts

The fastest way to internalize all of this is to do it once. Pick a tiny idea, write its core loop as one sentence, then break that sentence into five build prompts, each with a target, a behavior, and a trigger and result. Run the game after every one.

Try the AI game maker and start from a template close to your idea so your first prompt has something to reshape instead of a blank page. Summer Engine is free to download, the export has no watermark or revenue share, and the game you prompt into existence is genuinely yours. Write the smallest prompt first. Then write the next one.