How to make a game with AI

Start with a verb

Catch a falling parcel. Place a block. Sneak past a guard. Pick the thing the player spends their time doing. You can have a large world in mind, but the first build needs an action you can judge.

Then add a reason to pay attention. Perhaps parcels are heavy, blocks change the next turn, or the guard turns toward noise. That relationship is more useful to the first build than a long list of environments and collectibles.

Write the smallest useful brief

Include the target device and controls. Add your chosen engine and version if you have one. Ask the assistant to work within the current project rather than guessing a new stack.

Treat the first output as a proposal you can play. You do not owe it loyalty because it took a while to generate.

A brief to make your own

Make a game where I move a basket to catch falling parcels. Heavy parcels make the basket slower until I deliver them at the edge. Show what I am carrying. Let me lose, finish a round, and restart. Use simple readable shapes first. Explain how to run the project and what you checked.

Try to finish it, then try to break it

1. Run the game from its real entry point.
2. Play through the main goal without using developer shortcuts.
3. Fail and restart. Finish and restart. Do both again.
4. Try an input at an awkward moment: while paused, while losing, or during a transition.
5. Record one defect or design question to address next.

Ask for a change you can compare

“Make it more fun” is a reasonable desire and a poor bug report. Try “the basket is so fast that parcel weight makes no difference; make carrying three parcels require planning.” You can play that change and decide whether it helped.

Save a working version before a major revision. After the change, replay the old checks as well as the new behavior. A new feature can quietly remove an old rule.

Add presentation after the rule earns it

Once the action works, choose an art direction, readable feedback, and sound where it helps. Let a friend play without coaching. Watch what they misunderstand before writing another tutorial panel.

Test the intended output early. A browser build on your laptop does not prove a phone experience, and source export is not app-store submission. Check what your tools support before promising friends an install link.