Claude Code for Game Makers: A Simple Setup

Claude Code works best on a game when it has a short CLAUDE.md file to read and has to show what it tried before it says something is done. It reads your project, changes files and runs commands, so your setup is mostly about telling it what it cannot see and asking it to back up what it says.

The file names and commands below come from the Claude Code documentation as of 2 November 2026. The tool changes quickly, so compare them with the current docs.

What does Claude Code do in a game project?

Claude Code is an AI coding tool. You can use it in a terminal, in a code editor, in a desktop app or on the web. It can change many files at once, run commands, read the results and fix what failed.

It does not see your game run unless you give it a way to. For a game that matters a lot, because much of what counts happens on screen.

How should I set up my project first?

Ask Claude Code to keep the game's rules, such as movement, damage and inventory, in their own files, apart from graphics and sound. Then it can try the rules by itself and show you what happened. When things break anyway, read My game keeps breaking.

What belongs in CLAUDE.md?

CLAUDE.md is a file in your project folder. Claude Code reads it at the start of every session. Put in what it cannot find out from your files. The docs recommend under 200 lines, because longer files cost more and get followed less.

# My game

- Start the game: `pnpm dev`.
- Game rules live in `src/rules/`. Keep graphics out of that folder.
- The story lives in `story-bible.md`. Read it before you write any story.
- Before you say something is done, show me what you tried and what happened.
- Never just say something fits. Show me where that is written. If you cannot, say "I do not know".
- Do not use names, art, text or numbers from other games.

Your commands may differ, so ask Claude Code which ones your project uses. If your project has a command that runs automatic tests, add it too. The docs say that line is often the most useful one.

They also say Claude treats this file as background and not as a hard rule. That leads to the next step.

How do I make it show its work?

The docs explain why this matters. Claude stops when the work looks done, and if nothing tells it whether the work really is done, "looks done" is all it has to go on. So ask it what it tried and what happened. You can do that in three ways, each stricter than the last:

• In your message. Ask before you accept "done".
• In CLAUDE.md. Add the lines above, so it does this without being asked.
• As a hook. A hook is a command that Claude Code itself runs at a fixed moment, so it does not depend on Claude remembering. A Stop hook runs when Claude tries to finish, and it can send Claude back to work.

Whether your project needs a hook or automatic tests is up to Claude Code. Paste this, and it will tell you what it thinks:

Look at this project. Is there something you could run by yourself each time you try to finish, to find out if the game still works? If so, tell me what it is and whether a Stop hook in .claude/settings.json makes sense here. Show me the files you looked at. If you do not know, say "I do not know". Do not set anything up until I say OK.

There is a limit on how many times in a row a hook can send it back, so this is a strong guard and not an endless loop.

The docs' rule of thumb: if something has to happen every time, make it a hook or a similar setting, because Claude may decide that a line in CLAUDE.md does not apply.

How do I start a bigger feature?

The docs suggest letting Claude interview you until it can write a plan to a file, and then starting a fresh session to build from that plan. The build then begins with a clean head and a written description of what done means.

I want to add a [your feature] to my game, and I have not thought it through. Interview me until you can write a plan, then write it to PLAN.md. Include what it should do, which numbers it uses, what must always be true, and what happens in odd cases. For anything that depends on how the game works today, name the file you read. If you do not know something, write "I do not know" and list it as an open question. Do not guess.

Read the plan, then let it build. When it says it is done, ask what it tried and what happened.

Do subagents help?

Yes, for a second opinion. A subagent is a helper that works in its own separate chat with its own tools. A helper that did not write the code looks at it with fresh eyes. Project helpers live in the .claude/agents/ folder. Paste this and Claude Code can set one up:

Create a read-only project subagent called game-reviewer. After changes to the game rules, it compares them with PLAN.md. For every problem it finds, it quotes the line in PLAN.md and the line in the change. It also reports anything the plan does not cover, and any name or number taken from another game. If it cannot tell, it says "I do not know".

How do I connect Spielerando?

MCP is a way to plug extra knowledge into Claude Code. The MCP guide explains it, and what to look at before you connect any server.

Log in on our site by e-mail or Google, then create a key on the keys page. Add it like this:

claude mcp add --transport http spielerando https://spielerando.com/api/mcp \
  --header "Authorization: Bearer spl_your_key"

Our server gives your AI our wiki, examples and tested values. Usage is counted in tokens, though no AI runs on our side. It never takes or runs your game or your code.

What will Claude Code not do?

In October 2026 Anthropic confirmed that it added two sentences to Claude Code's prompt on Opus 5.5 to steer it away from recreating someone else's specific game. It said that using Claude to infringe IP is against its usage policy. According to Game Developer, original games, genre games and games built from your own designs and assets are meant to keep working. That matches what you should want anyway. See Can you copy game mechanics?

What should stay with me?

• Feel and fun. A game can work and still play badly, so play it.
• Design decisions. Where your plan is silent, the AI picks something. Read what it picked.
• Licences and origins of everything that goes into the game.
• Secret keys. Keep them out of files that get shared, and ask Claude Code how.
• The final say. An AI saying it is done tells you nothing. What it tried and what happened tells you something, but only about what it tried.