Building games with AI

Prompt patterns that get good results, and how the AI verifies its own work.

Godot MCP Studio gives your AI 265 tools, but the results depend on how you

direct it. These patterns come from building real games with it.

Start from what exists

Ask the AI to look before it builds: "check the scene tree and project settings

first" makes it call get_scene_tree and get_project_info instead of

guessing. The generated AGENTS.md teaches this, but saying it in your first

prompt helps every client.

Good: "Make a third-person character that walks around this terrain with a

camera that doesn't clip through walls." The AI picks add_body_3d,

add_spring_arm, create_terrain, and wires input.

Less good: dictating each node by hand. The one-call builders exist so the AI

can compose whole features; let it.

The strongest pattern is the verify loop: "run the game and show me a

screenshot", or "prove the jump works by monitoring the player's position".

The AI can run the project, watch properties over time, simulate input, read

the game's errors and output, and screenshot what the player sees. Ask for

proof and it stops hallucinating success.

Ask for placeholder textures early. A playable gray-box game you can iterate

on beats a beautiful scene that doesn't run. create_placeholder_texture,

create_material_preset, and add_shader_preset keep things visible before

real assets exist.

Every scene edit goes through the editor's undo system: Ctrl+Z reverses the

AI's last change like your own. Keep the project in git and let the AI commit

working checkpoints; a clean history is the best undo of all.

Visual authoring belongs in the editor. Materials, textures, lights, collision,

and layout should be saved into scenes or resources so you can inspect them

before pressing Play. Runtime tools are for testing the saved game, not for

creating permanent content that disappears when the game stops.

If a live tool reports "editor not connected", open the project in Godot with

the addon enabled. If the AI can't find a class or property, ask it to use

describe_class instead of guessing. And for anything the fixed tools don't

cover, Full Control mode lets it write a one-off editor script.