What should I tell the AI if I do not know what game I want to make?

Start from a game you already like and tell the AI to make something in that direction. A 2D platformer like Celeste or a top down shooter like Vampire Survivors gives the AI a complete mental model to build from, and gives you a running game in minutes that you can react to. Reacting to a real prototype is far easier than designing in your head. Once it is on screen you will immediately know what you want to change, and that is when the real ideas show up.

How detailed does my first message to the AI need to be?

One clear sentence is enough to start. Name the genre or a reference game, the main action the player does, the win or lose condition, and the art style. You do not need to plan levels, story, or menus first. The reliable approach is to get the smallest playable version running, then add one thing at a time. A long, detailed brief up front often makes the AI guess at dozens of decisions at once, and when the result is off you cannot tell which part went wrong.

Why does the AI build a different game than the one I imagined?

Usually because the prompt described a vibe instead of mechanics. The AI cannot turn make it fun or make it epic into code, those are results of rules, not rules themselves. Tell it the concrete behavior instead: the player runs automatically and jumps when I tap, hitting an obstacle ends the run. Naming a reference game also closes the gap fast, because make it like Flappy Bird carries hundreds of unspoken rules the AI already knows.

Do I have to describe the art and sound, or just the gameplay?

Describe the gameplay first and get it playable, then describe the look. When you do describe art, be specific about style and view: pixel art, side view, bright and cartoony beats make it look good. In an AI native engine you can generate the actual sprites, models, music, and sound effects from the same plain language prompts and drop them straight into the running game, so you tell the AI what to make for both code and assets in the same conversation.

Can I just paste my whole game idea as one big message?

You can paste a long idea as background, and it helps the AI understand where you are going. But it should still build one piece at a time, not all of it from a single message. Treat the big description as context, then ask for the core loop first, run it, and add the rest in order. Building the whole thing from one giant prompt means the AI makes many design choices blind, and you lose the ability to catch a wrong turn early.

What is the single most common mistake when telling AI what game to make?

Asking for an outcome instead of an action. Make a challenging RPG, make a relaxing game, make it addictive, none of these tell the AI what to actually build, so it guesses. Replace the outcome with the mechanic that produces it: enemies that take three hits and hit back is what challenging means in code. Name what the player does, the rule that triggers, and what happens as a result, and the AI has something real to build.

Is telling an AI to make a game free to try?

Yes. Summer Engine is free to download and use, including 3D, multiplayer, and a no watermark Steam export, so you can describe a game and watch it get built without paying. The paid plan only raises AI usage limits and adds team features. Many browser based prompt to game tools cap how many prompts you get per day, stamp a watermark on the result, or lock export behind a subscription, so check those limits before you commit to an idea you want to ship.

←Back to Blog

June 6, 2026·Summer Team

How to Tell AI What Game to Make (So It Builds the Right One)

You open the prompt box and freeze. Here is how to tell an AI what game to make, with a fill-in-the-blank brief, three worked examples, and the order to describe things in.

You open the chat box, the cursor blinks, and you have nothing. You know you want to make a game. You might even have a vague picture of it. But the moment you have to type the first sentence to an AI, the idea evaporates and you write something like "make me a cool game" and hope.

That blank-box freeze is the real beginner problem, and it has nothing to do with code. The skill of making a game with AI starts before any mechanic gets built: it is knowing what to tell the AI so it builds the game in your head instead of a generic one. This guide is about exactly that first message and the few that follow it.

Why "make a fun game" gives you a bad game

An AI builds what you describe. If you describe a feeling, it has to invent everything underneath that feeling, and it will invent something that is not yours.

"Make a fun platformer" forces the AI to guess the player's speed, what the enemies do, how jumping feels, what the goal is, and what it looks like. You did not decide any of those, so the AI decides them for you, and you end up reacting to a stranger's game instead of building your own.

The fix is to describe the game as a set of concrete things the AI can actually build, not as the experience you hope it produces. "Challenging" is not a thing you can build. "Enemies take three hits and hit back" is. "Relaxing" is not buildable. "No enemies, no timer, plant seeds and watch them grow" is. The whole skill of telling an AI what game to make is translating the feeling in your head into actions and rules.

The four things to tell the AI, in order

Every useful first message answers four questions. You can fit all four in one sentence.

What kind of game is it? Name the genre, or better, name a real game it can copy. "A 2D platformer like Celeste" or "a survivors-style game like Vampire Survivors" gives the AI an enormous amount of information for free, because it already knows how those games play.
What does the player do over and over? This is the core action, the thing you repeat for the whole game. Run and jump. Shoot. Place a tile. Talk to an NPC. Pick the one verb that matters most.
How do you win or lose? The goal or the fail condition. Reach the flag. Survive ten minutes. Drop to zero health. This is what turns a toy into a game.
What does it look like? One or two words. Pixel art. Top-down. Hand-drawn. Dark and moody. You can skip this on the first message and add it once the game runs.

Put together, a complete first message is short: "A 2D platformer like Celeste where I collect coins and dodge spikes, reach the flag to win the level, pixel art." That single sentence is enough for an AI native engine to build something you can play in minutes.

Start from a game you already like

If you are stuck on what to even ask for, the single most useful move is to name an existing game. It does three things at once: it tells the AI the genre, the core action, and roughly how it should feel, all in two or three words.

"Make it like Flappy Bird" carries a hundred unspoken rules the AI already understands: tap to flap, gravity pulls you down, pipes scroll toward you, touch anything and you die, score one per pipe. You did not have to write any of that. The reference did it.

This is also the cure for the blank box. You do not need an original idea to begin. Start from a clone of something you love, get it running, and your own ideas will show up the moment you are reacting to a real game on screen instead of an empty imagination. Reacting is easy. Inventing from nothing is hard. Get to the reacting stage as fast as possible.

If you genuinely have no idea where to point, our I have a game idea walkthrough helps turn a half-formed thought into a buildable one, and the game templates each come with a working version you can describe changes to instead of starting from a blank project.

Three worked examples

Here is the difference between a message that makes the AI guess and one that makes it build the right thing.

Endless runner. Weak: "make a phone game that is addictive." The AI has no idea what you touch, what kills you, or what the screen shows. Strong: "An endless runner like Subway Surfers, side view. The character runs automatically and jumps when I tap. Hitting an obstacle ends the run. Score goes up the longer I survive. Bright and cartoony." Every sentence is something the AI can turn into a rule.

Top-down shooter. Weak: "make an epic action game." Strong: "A top-down shooter like Vampire Survivors. I move with WASD and the gun fires automatically at the nearest enemy. Enemies spawn in waves and chase me. I have a health bar, and the run ends at zero. Pixel art, dark dungeon." This is a real game scoped down to its loop, which is exactly what you want first. The survivors-like template starts you here so you can describe tweaks instead of building the base.

Cozy sim. Weak: "make a relaxing game with no stress." "Relaxing" is a feeling, not a build instruction. Strong: "A top-down farming sim like Stardew Valley. I walk around a small farm, till soil, plant seeds, water them, and they grow over a few days into crops I can sell for coins. No combat, no timer. Soft pixel art." The absence of pressure is now described as concrete rules (no combat, no timer) instead of a mood.

Notice the pattern. In every strong version you can point at each sentence and say what the AI builds from it. That is the test. If a sentence in your message does not map to something on screen, the AI is going to guess at it.

After the first message: tell it what is wrong, not what you wish

The first build will be close but not right, and that is normal. This is where most people get stuck again, because they go back to vague: "make it better." The AI cannot act on "better."

Tell it what is wrong and which direction to move. "The jump feels floaty, make the player fall faster after the peak." "Enemies are too fast, slow them to half speed." "The level is too short, add two more platforms before the flag." Each of these names the current behavior and the change, so the AI knows exactly what to touch.

Change one thing per message, then run the game and look. If you ask for five changes at once and the result is wrong, you cannot tell which change broke it. One change, run, check, next change. This loop is slower to describe than it feels in practice, and it is the entire job of telling an AI what game to make: many small, specific, testable instructions, with a playtest between each. If you want the prompt-craft side in depth, the prompts that actually work guide is the companion to this one.

Telling it what art and sound to make

Once the game plays right, you tell the AI what it should look and sound like, and the same rules apply: be specific. "A 16 by 16 pixel art coin, gold, front-facing, for a 2D platformer" gets you a usable coin. "Make a coin" gets you a guess.

In an AI native engine, the assets come from the same conversation as the code. You describe a sprite, a 3D model, a sound effect, or a music track in plain language, the engine generates it, and it drops into the running game so you see it in context. You are not switching tools or hunting asset stores. You tell the AI what to make, for both the gameplay and the things you see and hear, in one place. The 2D asset generator and 3D asset generator posts go deeper on describing art well.

A fill-in-the-blank first message

Copy this, fill the blanks, and you will never face the blank box again:

A [genre or reference game] where the player [main action]. You win by [goal] and lose if [fail condition]. Make it [art style, optional for now].

For example: "A puzzle game like Tetris where the player rotates and drops falling blocks to clear full rows. You win by surviving as long as possible and lose if the blocks reach the top. Make it clean and minimal." That is a complete instruction, and an AI native engine can start building from it immediately.

You do not need to know your whole game before you start. You need to know enough for the first playable version, because once it is running, the game tells you what it wants to be next.

Try it

Summer Engine is an AI native game engine, compatible with Godot 4, where you describe a game in plain language and watch it get built, mechanics and art together, in a real editor you can ship from. It is free to download and use, including 3D, multiplayer, and a no-watermark Steam export, so you can take a single sentence and see how far it goes without paying anything.

Pick a game you like, fill in the blank-message template above, and send it. The hardest part of making a game with AI is the first sentence, and now you have it.

Frequently asked questions

What should I tell the AI if I do not know what game I want to make?: Start from a game you already like and tell the AI to make something in that direction. A 2D platformer like Celeste or a top down shooter like Vampire Survivors gives the AI a complete mental model to build from, and gives you a running game in minutes that you can react to. Reacting to a real prototype is far easier than designing in your head. Once it is on screen you will immediately know what you want to change, and that is when the real ideas show up.
How detailed does my first message to the AI need to be?: One clear sentence is enough to start. Name the genre or a reference game, the main action the player does, the win or lose condition, and the art style. You do not need to plan levels, story, or menus first. The reliable approach is to get the smallest playable version running, then add one thing at a time. A long, detailed brief up front often makes the AI guess at dozens of decisions at once, and when the result is off you cannot tell which part went wrong.
Why does the AI build a different game than the one I imagined?: Usually because the prompt described a vibe instead of mechanics. The AI cannot turn make it fun or make it epic into code, those are results of rules, not rules themselves. Tell it the concrete behavior instead: the player runs automatically and jumps when I tap, hitting an obstacle ends the run. Naming a reference game also closes the gap fast, because make it like Flappy Bird carries hundreds of unspoken rules the AI already knows.
Do I have to describe the art and sound, or just the gameplay?: Describe the gameplay first and get it playable, then describe the look. When you do describe art, be specific about style and view: pixel art, side view, bright and cartoony beats make it look good. In an AI native engine you can generate the actual sprites, models, music, and sound effects from the same plain language prompts and drop them straight into the running game, so you tell the AI what to make for both code and assets in the same conversation.
Can I just paste my whole game idea as one big message?: You can paste a long idea as background, and it helps the AI understand where you are going. But it should still build one piece at a time, not all of it from a single message. Treat the big description as context, then ask for the core loop first, run it, and add the rest in order. Building the whole thing from one giant prompt means the AI makes many design choices blind, and you lose the ability to catch a wrong turn early.
What is the single most common mistake when telling AI what game to make?: Asking for an outcome instead of an action. Make a challenging RPG, make a relaxing game, make it addictive, none of these tell the AI what to actually build, so it guesses. Replace the outcome with the mechanic that produces it: enemies that take three hits and hit back is what challenging means in code. Name what the player does, the rule that triggers, and what happens as a result, and the AI has something real to build.
Is telling an AI to make a game free to try?: Yes. Summer Engine is free to download and use, including 3D, multiplayer, and a no watermark Steam export, so you can describe a game and watch it get built without paying. The paid plan only raises AI usage limits and adds team features. Many browser based prompt to game tools cap how many prompts you get per day, stamp a watermark on the result, or lock export behind a subscription, so check those limits before you commit to an idea you want to ship.