How to make a game with AI, step by step

This is a general process that works with most prompt-driven game tools, not just one. The steps matter more than the tool. The single biggest factor in success is scope: start far smaller than feels ambitious. A tight, finishable game beats a grand idea that never works.

Choose a single core loop you can describe in a sentence. Resist adding menus, levels, and shops on day one. You can layer those in later. A good first idea names one genre and one main action.

• A platformer where you jump between rooftops and collect coins.
• A top-down game where you dodge falling rocks and survive as long as you can.
• A puzzle where you slide blocks to clear a path to the exit.

A strong prompt gives the AI a clear target. Name four things: the genre, what the player does, how they win or score, and the vibe. Keep it short.

For example: a fast neon endless runner where you slide and jump to dodge obstacles, and your score goes up the longer you survive. That names the genre, the action, the goal and the feel, so the AI has a clear target.

Generate the first version and play it. It gives you something concrete to react to. Notice what feels right and what does not, and write down the one change that would help most.

If the build does not run or looks broken, tell the tool exactly what failed and ask for a fix. Strong tools catch and repair many errors on their own.

This is where most of the work happens. Ask for small, specific changes and play after each one. Small requests are far easier for an AI to get right than one giant rewrite.

• Make the jumps a little higher.
• Add a second obstacle type that moves faster.
• Speed up the scrolling as the score climbs.
• Change the palette to something darker.

If a tool offers checkpoints or version history, use it. It lets you try a bold change and roll back if it goes the wrong way.

Play your game as a player would, not as its maker. Better still, have someone else try it. Fresh hands find confusing controls and dead ends fast. Fix what trips them up, then share it. A browser game is easiest to share, since players just open a link.

OddShed runs these steps for you. Describe the game, answer the question cards, approve the plan, and the agent builds a Phaser 2D or Three.js 3D game in the live preview. Change it by chat, roll back to a restore point when a change misses, and share it by link or list it in the public Game Library once it passes review. Each turn shows exactly what it cost.

• Start with one small, finishable idea.
• Write a clear prompt: genre, player action, goal, vibe.
• Play the first build, then refine one specific change at a time.
• Test with fresh hands, then share, ideally as a browser link.