AI Game Prompts: Turn a Vague Wish Into a Good Request

A good AI game prompt describes one thing the player does, with numbers, a feeling and a clear place to stop. A vague wish gets a dull result, because your AI fills every gap with whatever is most common.

Why does a vague wish go wrong?

Say you are making Rootling Deep, an invented game about a seedling growing through a cave. You type, "Add a grow mechanic." You get a bar that fills up when you press a key, and the plant gets bigger. It works, and it is the most ordinary version anyone could think of.

The AI did not fail. It had one word to go on and many choices to make, so it made the usual ones. The Claude Code docs say a specific prompt gives the AI information it does not already have, and that it often saves you a round of corrections.

What is the four-line card?

Before you ask for anything, fill in these four lines:

• Action: what the player does, in one sentence.
• Numbers: every amount, speed and limit.
• Feeling: how it should feel to play.
• Stop: what you do not want built this time.

For the seedling, that gives:

• Action: Hold the button and the root tip grows towards the pointer. Let go and it stops.
• Numbers: It grows 3 tiles a second. You start with 40 water, and each tile costs 1.
• Feeling: Tense when water runs low, like threading a needle.
• Stop: Only the root and the water meter. No enemies, no menu, no art.

Each line closes a different gap. Action stops the AI from inventing a different game. Numbers give you something to change later. Feeling tells it what to aim for. Stop keeps one request from turning into twelve.

You can guess the numbers. A wrong number is easy to change, and a missing one gets decided for you. Stop is the line people skip, and it is often the one that saves the evening.

What do other wishes look like as cards?

• "Add a shop." Action: spend coins on one of three boots between levels. Numbers: boots cost 20, 50 and 90. Feeling: you can never afford everything. Stop: no selling, no inventory screen.
• "Make a boss." Action: dodge three telegraphed attacks, then hit the glowing eye. Numbers: 5 hits to win, 1.5 seconds of warning. Feeling: hard, but always your own fault. Stop: one room, one boss.

If you are not sure how to start at all, AI prompts for game development goes stage by stage. If you want to know what kind of thing you are describing, read what a core game mechanic is.

Paste this to your AI

I have a vague wish for my game: [your wish].

Turn it into a four-line card before you build anything:
Action: what the player does, in one sentence.
Numbers: every amount, speed and limit, as starting values.
Feeling: how it should feel to play.
Stop: what you will not build this time.

Ask me up to three questions. Where you fill in a value I did not give you, mark it "my suggestion". If it looks like a feature from an existing game, tell me, so I can make it my own.

When I approve the card, draw the feature as a small Mermaid state diagram: every state it can be in, and what moves it to the next one. Walk through each path, including running out of something, and show me any dead end. Build only from that drawing.

Then go through the card line by line and show me where each line is in the game and what you tried. If a line is not done or you could not try it, say so. The Feeling line is mine to judge, not yours.

Or skip this and connect the MCP: your AI then finds all of this by itself.Connect the MCP

How to check it worked

• Read the drawing. In the example, what happens when the water reaches zero while you hold the button? There should be an arrow for that.
• Check the Stop line. Nothing on it should exist in the game.
• Change one number yourself. Ask where the growth speed is set, change it, and see the game change.
• Give the card to a fresh chat. If it would build something quite different, your card still has a gap.