Making a game with vibe coding can be a lot of fun — but if you don’t know how to start and you lack technical background, it becomes frustrating fast. This is the workflow that keeps development smooth.

Define the game before you prompt

Before writing a single prompt, know exactly what you want: the theme, the type of view (2D, 3D, isometric?) and the design style. Defining this upfront saves you a lot of trouble later.

If you are not sure what you are looking for, let the AI define it. Give it names of games you like or want to be inspired by, or just hand it a theme and ask which format would work best.

Write the requirements as documentation

Once the game is clear in your head, formalize it: what levels it will have, what items, what characters. This is the moment to make everything explicit.

Ask the AI to generate markdown documentation detailing all the mechanics, characters and technical requirements, and iterate with it until it matches exactly what you have in mind. Split the documentation into several files — one for characters, another for items, and so on.

Plan tasks before generating code

This step makes the biggest difference. Instead of jumping straight to code, plan the tasks step by step. You gain precision and save time, because you can parallelize a lot of the work.

A task-board tool helps here — I use Kanvas (by XMihura), which is as easy to set up as copying a Python file into your directory plus an AGENTS.md definition. With it configured, ask the agent to generate the tasks for you (the prompt above). Two details matter:

• Mention the AskUserTool. Instead of deciding everything by itself, the agent asks your opinion on key decisions.
• Demand a testable version from the very first moment. You play the game as you go and verify everything follows the plan.

Review the plan, then start building

With the tasks defined, review that everything is in order. If something doesn’t fit, fix the task now — or ask the agent to do it for you.

Then start, and parallelize. Modern environments like Claude Code or Codex often do this automatically, but an explicit prompt still helps: “Do tasks X, Y, Z in parallel, use sub-agents to parallelize the work.”

The key is iterating: review the result in the game after finishing each task, and don’t mark it as approved until it is truly perfect.

Graphics make or break the game

If a game isn’t visually appealing, nobody wants to play it. Fortunately, getting assets is easier than ever. Three options:

• Use existing assets. Repositories like Itch.io or Kenney.nl have all kinds of assets, free and paid, made by artists — great visual coherence guaranteed.
• Generate assets with AI. Dozens of models produce high-quality images or 3D models at very low prices. The hard part: finding a prompt (and maybe a reference image) that keeps every asset visually harmonious.
• Let your agent generate them. Whether the game is 2D or 3D, agents can build assets in code — SVG sprites for 2D, Three.js objects for 3D. Decent results are possible, but it takes a lot of iteration and today it’s still hard to outperform an artist.

That is the whole workflow: define, document, plan, parallelize, and never ship ugly graphics.

Analyze all the game documentation thoroughly, think of the best way to
execute the project from scratch step by step, and generate the necessary
tasks in a way that I have a testable version from the very first moment.
If you have any doubts or need to clarify anything, use the AskUserTool
to ask me.