Vibecoding: Making Games with AI Tools

This guide walks you through setting up EdenSpark to work with AI coding tools — so you can describe a game in plain words and have an AI agent build it while the editor is running.

We’ll use Claude Code as the primary example, but the setup also works with Cursor, GitHub Copilot, and any MCP-compatible tool.

Prerequisites

Before you start, make sure you have:

  • EdenSpark installed and working (see Start Here if you haven’t set it up yet)

  • An AI coding tool — at least one of the following:

  • VS Code installed (for opening the project code)

Step 1: Create a Project

  1. Open the EdenSpark Launcher.

  2. Click the button under My Games to create a new game.

  3. Pick a template — 2D or 3D. It doesn’t matter which one; the AI will rewrite the code anyway. Clone it to create your project.

"Create new game" page

  1. On the game page, click Edit to open the EdenSpark Editor with your project.

  2. In the editor, go to Edit → Open VS Code Project. This opens VS Code with all the generated configuration files already in place.

Step 2: Set Up AI Tools

When you cloned the project, EdenSpark auto-generated configuration files for several AI tools:

  • Claude Code — CLAUDE.md with MCP tool reference and Daslang language guide

  • Cursor — .cursorrules with Daslang language reference

  • GitHub Copilot — .github/copilot-instructions.md with Daslang language reference

  • Other agents — AGENTS.md with Daslang language guide and Game Designer instructions

  • MCP connection — .mcp.json pointing to the running editor and Context7 documentation server

You don’t need to configure these manually — they’re ready to go.

Setting up Claude Code

If you haven’t installed Claude Code yet, follow the setup guide.

You can run Claude Code in a standalone terminal or inside VS Code’s integrated terminal. Using VS Code is convenient — you can drag the terminal tab up next to your code tabs for a side-by-side view:

  1. Open the terminal with Ctrl + ~

  2. Drag the terminal tab name (e.g. “PowerShell”) up to the code tab bar

  3. Switch to WSL or Git Bash (Claude Code requires a Unix-compatible shell on Windows)

That’s it — Context7 (live EdenSpark and Daslang documentation) and EdenSpark MCP (editor connection) are both configured in the project’s .mcp.json and will be picked up automatically when you launch Claude Code from the project directory.

Setting up Cursor or Copilot

No extra setup needed. Open the project folder in your tool — the configuration files are already generated. For MCP support, check your tool’s documentation on how to connect to an MCP server (the .mcp.json file in your project root has the connection details).

Step 3: Verify the Connection

Before you start generating, make sure everything is connected.

Note

The EdenSpark Editor must be running for MCP to work. The MCP server starts together with the editor. If you have multiple editor instances open, they will compete for the same port — keep only one running.

The launcher does not affect MCP.

In Claude Code, type /mcp to check connected tools. You should see both:

  • EdenSpark — connected to the running editor

  • Context7 — connected to the documentation server

If EdenSpark MCP is not connected, make sure the editor is running and that you opened the project from the launcher.

If Context7 is not connected, make sure you launched Claude Code from the project directory (where .mcp.json is located).

Step 4: Start Building

Navigate to your project directory in the terminal and launch your AI tool:

cd /path/to/your/project
claude

Claude will detect the MCP configuration and offer to use the EdenSpark tools. Confirm to proceed.

Now describe your game. A good starting prompt looks like this:

Create a variation of Flow Free game. Here's a description of the gameplay:

<paste gameplay description here>

Tip

For describing gameplay, copy-pasting a Wikipedia article about the game you want to recreate works surprisingly well. The AI gets a clear picture of the mechanics, rules, and win conditions.

Step 5: Design Before You Code

Don’t jump straight into code. Ask the agent to produce a Game Design Document (GDD) first — a short document describing the core loop, controls, win/lose conditions, game objects, and visual style.

Take a moment to review it:

  • Does the core loop match what you had in mind?

  • Are the controls right?

  • Is the scope realistic for a first version?

Ask the agent to adjust anything that’s off — “change the controls to mouse only”, “remove the timer, make it endless”, “look up more reference games for the scoring system”. You can go back and forth a few times until the GDD feels right.

Once you confirm the design, the agent switches to implementation.

Step 6: Watch It Build

The agent will:

  1. Look up relevant APIs via Context7

  2. Inspect the current scene in the editor

  3. Write Daslang code, save it, and watch the editor hot-reload

  4. Check compilation status — if there are errors, it reads and fixes them

  5. Take a screenshot to verify the result

  6. Simulate input to playtest the game loop

  7. Iterate until the game works

You can keep iterating from there: “make the ball faster”, “add a score counter”, “change the background color” — the agent modifies the code and verifies each change in real time.

Tips

Keep the editor running. The MCP server lives inside the editor process. If you close the editor, the AI loses its connection to the scene.

One editor instance at a time. Multiple editors will fight over the MCP port. Stick to one.

Be specific in your prompts. “Make a platformer” is vague — “Make a single-screen platformer where the player jumps between platforms to collect coins, with a 60-second timer” gives the AI much more to work with.

Iterate in small steps. Rather than describing the entire game upfront, build it feature by feature. Get the core loop working first, then add scoring, then add menus.

Samples as learning material. You can tell the agent to study the bundled sample projects for reference — it will look at working code and apply the same patterns. Be aware that reading sample sources consumes tokens, so point the agent at specific samples rather than asking it to read all of them.