AI Prompts for Game Development, Stage by Stage
The right AI prompt for game development depends on the stage you are in, because an idea needs arguing with while a finished feature needs checking. Keep one prompt per stage in a file, and you will stop asking for the wrong kind of help.
What goes wrong when you use the wrong prompt?
Say you are making Caravan of Salt, an invented trading game about crossing a desert. Here is what the wrong request does at each stage:
| Stage | What you need | The mistake |
|---|---|---|
| Idea | Someone to argue with | "Build it" gives you a pretty version of a weak idea |
| First playable | The roughest thing you can play | "Make it look nice" gives you polished menus and no game |
| New feature | A plan before the work | "Just add camels" gives you camels that break the trading |
| Polish | A list of rough edges | "Make it better" gives you new features instead |
| Release | Promises checked against the game | "Is it ready?" gives you a cheerful yes |
The stages run in order, and the jump between them is a decision you make. Move on from Idea when one sentence describes the fun. Leave First playable when someone asks for a second round. Stop adding features when your list says so, not when you run out of ideas.
The diagram that appears in these prompts carries over from stage to stage. It starts as the main loop of your game, then grows with each feature, so your AI always has one picture of what the game is meant to do. If you do not know what the main loop is, read what a core game mechanic is.
Spielerando can give your AI ready-made blueprints, code and a wiki through an MCP connection, which helps at the feature stage. It takes a free account with Google or e-mail, and it does not look at your game or code.
Paste this to your AI
Idea stage:
Here is my game idea: [one or two sentences]. Do not build anything. Argue with it. Draw the main loop as a Mermaid diagram: what the player does, what they get, why they do it again. Check that the loop closes, and show me where it breaks or gets dull. Mark anything that is only your opinion.
Or skip this and connect the MCP: your AI then finds all of this by itself.Connect the MCP
First playable:
Build the smallest version that can be played: one screen, grey shapes, no menus. First draw every state it can be in as a Mermaid diagram, walk each path and show me any dead end. Build from that drawing. Tell me what you tried and what happened. If you could not run it, say so.
New feature:
I want to add [feature]. Add it to the drawing first: which new states, which new arrows, what it touches. If the game and the drawing already disagree, show me that spot and stop. Build only after I agree, then show me what you tried, or say you could not try it.
Polish:
Add nothing new. Walk the main path in the drawing and list rough edges: missing feedback, confusing moments, places with no way back. Name where each one happens. Whether it is fun is for me to judge, so do not tell me it is.
Release:
List every promise on my game page. Next to each, show the place in the game that keeps it and what you tried. Mark anything you cannot show as "not shown". Do not tell me the game is ready. Tell me what you know and what you do not.
How to check it worked
- Run your idea through the first prompt. You should get at least one objection that stings. If it only praises the idea, ask what would make it fail.
- Keep the drawing up to date. After a new feature, the picture should have changed too.
- Show the first playable to someone. Say nothing. Do they want another round?
- Read the polish list. Nothing on it should be a new feature.
- Pick one promise from the release list and go to the place in the game it names.