How to Make a Bigger Game Alone, with AI Agents as Your Team

Why is a bigger game so hard to make alone?

A bigger game is hard to make alone because a game is many jobs, and a solo developer can only do one at a time. While you write code, nobody writes dialogue. While you balance the shop, nobody tests the last build. The project grows, but your hours do not.

Here is where solo projects usually stall:

• Every character sounds like you. The dialogue reads like programmer dialogue, because a programmer wrote it.
• Balance is guesswork. You tune the shop economy in a spreadsheet late at night and guess every number.
• Help does not stick. You post "looking for a writer, rev share" and the replies fade away.
• The devlog goes quiet. Marketing is one more job, so it is the first one dropped.

AI agents do not remove these jobs. They give you a first draft for each one, so you spend your time judging and fixing rather than starting from a blank page.

How to set up AI agents for solo game development, step by step

You can set up AI helpers for each studio role by hand with Claude Code and a few text files. Each helper gets one job, clear rules and your engine version. Doing this yourself once teaches you what a good agent needs.

1. Write a one-page game design document. List the core loop, which is the action the player repeats most, plus the main characters and the target platform. Every agent will read this page, so keep it short and current.
2. List the roles you are missing. Most solo devs lack a writer, a tester, a balance designer and a sound lead. Pick the two that hurt most. Do not set up ten roles on day one.
3. Create one agent per role. In Claude Code, a subagent is a Markdown file in your project's .claude/agents/ folder. It has a name, a description and its own instructions, and it runs in its own context. The Claude Code subagents docs show the file format.
4. Pin the engine and version. Tell the code agent which engine and version you use, such as Godot 4, and how you name scripts and nodes. Engines change between versions, and a wrong version means code that will not run.
5. Review every answer like a pull request. Run the code in the editor. Read it before you keep it. Commit to git after each working change, so one bad answer is easy to undo.
6. Give the writing agent a character sheet. For each character, write how they talk, what they want and three lines they would never say. Dialogue gets much less flat when the agent has a voice to copy.
7. Model the economy before you ask for numbers. List every source (where coins come from) and every sink (where coins go). Ask the agent to explain its numbers, then test them in a real play session.
8. Write a test checklist. Ask the testing agent to turn your design page into a list of things to try on each build. You still play the build, but now you play it with a list.
9. Watch the cost from day one. Run /usage in Claude Code to see your session usage. The Claude Code cost guide suggests Sonnet for most coding and a smaller model for simple helper tasks.

If you have never built a prototype, start with Godot's free Your first 2D game tutorial before any of this. For more agents and skills worth trying, see the best Claude Code skills for indie developers and where to find Claude Code skills.

AI tools for game art, 3D and animation

Most lists of AI game tools are about assets, not agents. Here are four that come up often, with what each vendor says it does. We have not tested them, so try the free tier on one asset first.

If you ship on Steam, read the AI part of the Steamworks content survey first. It asks you to disclose AI-made content that ships in the game, and content made by AI while the game runs. Tools that only make you faster are not its focus.

The faster way: install 27 game dev agents for Claude Code

The faster way is to install agents that already know their roles and your engine, instead of writing and testing each file yourself. Indie Game Studio Kit is 27 of them: engine specialists for Godot, Unity and Unreal, plus a narrative director, an economy designer, an audio director and more.

The agent files follow one rule: they are pair programmers, not autopilots. The narrative director's file says it plainly: "You are a collaborative consultant, not an autonomous executor." Each agent asks questions, offers options and waits for your call. That keeps you in charge of the part only you can do, which is knowing what makes your game fun.

One gap to know before you buy. Some agents name teammates the kit does not include. The narrative director, for example, hands final dialogue to a writer agent, and there is no writer file in the folder. There is no QA tester either. In practice those jobs come back to you, or to an agent you write yourself with the steps above.

How to use it, step by step

1. Buy and download. Polar or Gumroad sends the files right after payment.
2. Install the agents. Copy the 27 files in the agents folder into ~/.claude/agents, or into your project's .claude/agents folder to keep them to one game. The layout is flat, with no skills folder to set up. If an agent never shows up, check where the file landed first. A skill of mine once sat for weeks in a folder Claude Code does not load, and it only started working after I moved it on September 8, 2026.
3. Add your design page. Put your one-page game design document where every agent can read it.
4. Start with one role. For example, ask the narrative director to plan a scene: who is in it, what each character wants and how it moves the story. Then write the lines yourself, or ask a writing agent of your own.
5. Bring in the engine specialist. Ask the Godot, Unity or Unreal agent for the script that runs the scene. Test it in the editor before you keep it.
6. Check usage after the first week. Use /usage to see what a normal week costs you, then switch simple roles to a cheaper model if you need to.

What the output looks like

The agents work in the same pattern, and it is written into each file. First the agent asks clarifying questions: the player experience you want, your limits, games you love or hate. Then it gives you two to four options, each with pros and cons and a recommendation. Only after you choose does it draft, one section at a time, and it asks before it writes a file:

Explicitly ask: "May I write this section to [filepath]?" Wait for "yes" before using Write/Edit tools.

What you end up with is a document you approved section by section: a story outline, a world rules page, an economy plan with its sources and sinks, or a script for your engine. The narrative director, for example, must give each world element a one-line summary, its rules, its history, its links to other elements and a check that it does not contradict existing lore. Because the agents read the same design page, they talk about the same characters.

How much it costs to run

I have not measured a month of running these agents, so I will not give you a monthly figure. Your cost depends on how often you run them and on how you pay. On a Claude Pro or Max plan, use counts against the plan instead of a per-token bill. My own scheduled Claude Code jobs run under my normal subscription, with no separate API key. On the API, run /usage after your first week and you will see the real number.

For reference, Anthropic's Claude Code cost page puts average enterprise use at about $13 per developer per active day, or $150 to $250 a month. That is a full-time developer at work. A solo project a few evenings a week is a different pattern, so trust your own /usage number over any average.