A lightweight solo AI-coding studio template for game developers.
Build your game — not your framework.
v1.2 — Session-stop hook now warns about uncommitted changes. GDD gains a monetization section for mobile devs.
game-designergets mobile interruption design guidance.new-featureskill scans Known Issues before starting work.qa-checkcross-references Known Issues to avoid duplicate bug reports.
Most AI coding templates burn your token budget on coordination before a single line of code gets written. 49 agents debating with directors and leads just to decide what to do next works great for a studio with unlimited API credits.
For a solo dev on a weekly limit, it burns your cap with nothing to show.
This template strips everything down to what actually matters:
| Typical multi-agent | Solo Studio | |
|---|---|---|
| Agents | 20–49 | 6 |
| Hierarchy | Director → Lead → Worker | None — direct specialists |
| Inter-agent calls | Constant | Zero |
| Slash commands | 20+ | 10 practical ones |
| Token overhead | High | Near zero |
| Token budget going to code | ~30–50% | ~95% |
6 specialists. No hierarchy. No coordination overhead. Nearly 100% of your tokens go to actual game code.
- Claude Code —
npm install -g @anthropic-ai/claude-code - Or another coding agent that reads project instructions, such as Codex/GPT-style agents using
AGENTS.md - Godot 4 or Unity installed
- Git
git clone https://github.com/Ghraven/Claude-Code-Solo-Studio my-game
cd my-game
claudeThen in Claude Code:
/setup
Answer 4 questions. Your project is ready in seconds.
This repository is Claude Code-first because the slash commands, hooks, and specialist prompts use Claude's .claude/ convention. The underlying workflow is not Claude-only:
CLAUDE.mdis the Claude project brain.AGENTS.mdis the portable/Codex-friendly project brain.- The six specialists still map cleanly to Codex/GPT work: game design, programming, level design, UI, QA, and release.
- Slash commands can be used as plain prompts if your coding agent does not support Claude-style commands directly.
For Codex-style work, start by copying the relevant project facts from CLAUDE.md into AGENTS.md, then ask your coding agent to follow the same sprint, known-issues, and routing sections.
The setup command asks:
- Platform — Android · PC · Web
- Dimension — 2D · 3D
- Genre — Platformer · RPG · Puzzle · Shooter · Other
- Engine — Godot 4 · Unity (with a tip card to help you choose)
Then generates in one shot:
CLAUDE.mdfilled with your project config- Full folder structure (
src/·assets/·design/·docs/·tests/) - Starter scene and script files for your engine
design/gdd.md— Game Design Document stub ready to fill in
| Command | What it does |
|---|---|
/setup |
Initialize project — platform · genre · engine · starter files |
/new-feature |
Start a feature (breaks into max 5 focused tasks) |
/level-design |
Design a level with scene structure |
/ui-design |
Design a menu · HUD · screen |
/code-review |
Review code for bugs, engine-specific patterns, quality |
/qa-check |
Full quality check before moving on |
/bug-report |
Log a bug with severity, diagnosis, and regression test |
/performance |
Profiling check, bottleneck diagnosis, and optimization strategy |
/build |
Export/build checklist for your target platform |
/status |
Sprint snapshot — tasks, open bugs, next focus |
/help |
Show all commands with descriptions |
Use agents when you want deep focused work on a single domain. Slash commands handle most everyday tasks — agents are for when you need a specialist to go all-in on one area.
| Agent | Use for |
|---|---|
game-designer |
Mechanics · balancing · GDD updates |
programmer |
All code — GDScript or C# |
level-designer |
Level layouts · scene structure |
ui-designer |
Menus · HUD · screen design |
qa |
Code review · bugs · test cases |
release |
Builds · exports · versioning |
Every agent reads the project brain for context — CLAUDE.md for Claude Code, or AGENTS.md for Codex/GPT-style agents — so you do not need to re-explain your engine, platform, or genre every session.
my-game/
├── CLAUDE.md Your project brain — update this as you build
├── AGENTS.md Portable project brain for Codex/GPT-style agents
├── .claude/
│ ├── agents/ 6 specialist agent prompts
│ ├── skills/ 10 slash commands
│ ├── hooks/ 3 hooks (session start/stop, commit check)
│ └── settings.json Hook wiring and permissions
├── design/
│ └── gdd.md Game Design Document
├── src/ Game source code
├── assets/ Art · audio · fonts
├── docs/ Technical notes
└── tests/ Test cases
Agents don't talk to each other. Each specialist has full context baked into its prompt and reads the relevant project brain for project state. There's no coordination layer, no approval chains, no "consulting" step.
You stay in control. The coding agent starts with the specialist context it needs and gets to implementation quickly.
- Hooks only fire when relevant (commit hook exits instantly if not a commit)
CLAUDE.md/AGENTS.mdcarries context across sessions — no context re-setup- Specialists go straight to work — no meta-discussion about what to build
Everything is plain Markdown — edit freely:
# Add a new agent
touch .claude/agents/my-agent.md
# Edit a slash command
nano .claude/skills/new-feature/prompt.md
# Add a hook
touch .claude/hooks/my-hook.sh
# Wire it in .claude/settings.jsonThe default Claude Code permissions allow common read-only navigation tools, including rg, so agents can search the project quickly without broad shell access.
If you adapt the template for Codex or another GPT coding agent, keep the same safety idea: grant enough filesystem/tool access for fast search and local verification, but keep destructive actions and public pushes explicit.
Inspired by Claude-Code-Game-Studios by Donchitos.
Rebuilt from scratch for solo devs on limited API budgets.
See CONTRIBUTING.md — adding a slash command or agent is straightforward.
See CHANGELOG.md for release history.
I use AI-assisted development tools, including Codex and Claude, while building and maintaining this project. All code, design decisions, testing, commits, and releases are reviewed and shipped by me as the repository owner; AI tools are not listed as project contributors.
MIT — see LICENSE