Turn a single Claude Code session into a full game development studio.
49 agents. 74 skills. One coordinated AI team.
Building a game solo with AI is powerful — but a single chat session has no structure. No one stops you from hardcoding magic numbers, skipping design docs, or writing spaghetti code. There's no QA pass, no design review, no one asking "does this actually fit the game's vision?"
Claude Code Game Studios solves this by giving your AI session the structure of a real studio. Instead of one general-purpose assistant, you get 49 specialized agents organized into a studio hierarchy — directors who guard the vision, department leads who own their domains, and specialists who do the hands-on work. Each agent has defined responsibilities, escalation paths, and quality gates.
The result: you still make every decision, but now you have a team that asks the right questions, catches mistakes early, and keeps your project organized from first brainstorm to launch.
- What's Included
- Studio Hierarchy
- Slash Commands
- Does it produce games, or documents?
- Getting Started
- Configuration
- Upgrading
- Project Structure
- How It Works
- Design Philosophy
- Customization
- Platform Support
- Community
- Supporting This Project
- License
| Category | Count | Description |
|---|---|---|
| Agents | 49 | Specialized subagents across design, programming, art, audio, narrative, QA, and production |
| Skills | 74 | Slash commands for every workflow phase (/start, /design-system, /create-epics, /create-stories, /dev-story, /story-done, etc.) |
| Hooks | 15 | Automated validation on commits, pushes, asset changes, design-doc reads, writes and registries, session lifecycle, agent audit trail, and gap detection |
| Rules | 14 | Path-scoped coding standards enforced when editing gameplay, engine, AI, UI, network code, scene and asset files, and more |
| Templates | 62 | Document templates for GDDs, UX specs, ADRs, sprint plans, HUD design, accessibility, and more |
Agents are organized into three tiers, matching how real studios operate:
Tier 1 — Directors
creative-director technical-director producer
Tier 2 — Department Leads
game-designer lead-programmer art-director
audio-director narrative-director qa-lead
release-manager localization-lead
Tier 3 — Specialists
gameplay-programmer engine-programmer ai-programmer
network-programmer tools-programmer ui-programmer
systems-designer level-designer economy-designer
technical-artist sound-designer writer
world-builder ux-designer prototyper
performance-analyst devops-engineer analytics-engineer
security-engineer qa-tester accessibility-specialist
live-ops-designer community-manager
Each agent declares a model in its frontmatter: the three directors ask for
Opus, lead-programmer and fifteen specialists name Sonnet, and the remaining thirty
inherit your session's model. Treat the tiers as seniority of role, not as
a billing plan — most of the studio runs on whatever model you are already using.
The template includes agent sets for all three major engines. Use the set that matches your project:
| Engine | Lead Agent | Sub-Specialists |
|---|---|---|
| Godot 4 | godot-specialist |
GDScript, Shaders, GDExtension |
| Unity | unity-specialist |
DOTS/ECS, Shaders/VFX, Addressables, UI Toolkit |
| Unreal Engine 5 | unreal-specialist |
GAS, Blueprints, Replication, UMG/CommonUI |
Type / in Claude Code to access all 74 skills:
Onboarding & Navigation
/start /help /project-stage-detect /setup-engine /adopt /settings
Game Design
/brainstorm /map-systems /design-system /quick-design /review-all-gdds /propagate-design-change
Art & Assets
/art-bible /asset-spec /asset-audit
UX & Interface Design
/ux-design /ux-review
Architecture
/create-architecture /architecture-decision /architecture-review /create-control-manifest
Stories & Sprints
/create-epics /create-stories /dev-story /sprint-plan /sprint-status /story-readiness /story-done /estimate
Reviews & Analysis
/design-review /code-review /balance-check /content-audit /scope-check /perf-profile /tech-debt /gate-check /consistency-check /security-audit
QA & Testing
/qa-plan /smoke-check /soak-test /regression-suite /test-setup /test-helpers /test-evidence-review /test-flakiness /skill-test /skill-improve
Production
/milestone-review /retrospective /bug-report /bug-triage /reverse-document /playtest-report
Release
/release-checklist /launch-checklist /changelog /patch-notes /hotfix /day-one-patch
Creative & Content
/prototype /vertical-slice /onboard /localize
Team Orchestration (coordinate multiple agents on a single feature)
/team-combat /team-narrative /team-ui /team-release /team-polish /team-audio /team-level /team-live-ops /team-qa
A fair question, and the answer is a measurement rather than a sample project.
Four games were built from one brief at different modes.rigor levels and the
documents each wrote before its first line of game code were counted:
modes.rigor |
Docs before first line of game code | Time |
|---|---|---|
minimal |
1 (the one-page brief) | 0 min |
standard |
30 | 58 min |
All of them produced a running, tested game. Two blind reviewers and a human
playtest then ranked the standard build last — the extra 29 documents
bought traceability, not a better game.
The measurement above is about cost. The other half is whether the thing works,
and for a game that means someone has to look at it — a passing test says
nothing about a menu drawn off-screen. So a story that changes anything the
player sees is not closed until the game has been launched, observed, and a
screenshot retained in production/qa/evidence/. A parse check is not a run.
This holds at every rigor level, including minimal, where automated tests are
waived but the look is not.
The games themselves are deliberately not shipped here: this is a template you clone to start your own project, and somebody else's game in your repo is clutter, not evidence.
- Git
- Claude Code — the native installer is recommended (
npm install -g @anthropic-ai/claude-codealso works, on Node.js 22+) - Python 3 — reads
project.yamlfor every skill and hook; without it every setting silently falls back to its default - Bash — runs every hook; built into macOS and Linux, and on Windows it comes with Git for Windows
- Recommended: jq (for hook validation)
Hooks fail gracefully if optional tools are missing — you just lose validation — with one deliberate exception: a commit that stages data JSON is blocked when there is no Python to validate it.
-
Clone or use as template:
git clone https://github.com/Donchitos/Claude-Code-Game-Studios.git my-game cd my-game -
Open Claude Code in the project's root folder (the one holding
CLAUDE.mdandproject.yaml) and start a session:claude
Starting it in a subfolder such as
src/skips the project's hooks, deny rules and ask-first permission mode. -
Run
/start— the system asks where you are (no idea, vague concept, clear design, existing work) and guides you to the right workflow. No assumptions.Or jump directly to a specific skill if you already know what you need:
/brainstorm— explore game ideas from scratch/setup-engine godot 4.7— configure your engine if you already know/project-stage-detect— analyze an existing project
project.yaml at the repo root is the single source of truth for how the
studio behaves on your project — engine, naming conventions, and a modes
block that controls how much process every skill enforces. /start writes it
for you; you rarely need to hand-edit it.
The one question that matters most is modes.rigor — minimal, standard,
or full. It's a single front door over six separate knobs (workflow —
which GDD sections are required, docs.density, qa.level,
story_granularity, review_mode — how many director agents review your work,
and team.size), so /start asks it once instead of six times:
minimal(default) — jam-game speed. No GDDs required, terse docs, minimal QA evidence, solo review./start→/setup-engine→/brainstorm(a one-page brief) →/create-stories→/dev-story↔/story-done.standard— 5 required GDD sections, balanced doc depth, standard QA evidence, lean review.full— all 8 GDD sections, thorough docs, full QA evidence on every story type, full director review.
minimal is the default because the heavier tier costs more to reach working
code without producing a better result, and gives nothing back when someone new
picks the project up. Raise it in one /settings call when your project grows
— /help and /gate-check will suggest it when they see the signs.
Two escape hatches keep one setting from being a blunt instrument.
system_overrides holds a single system to a higher standard than the rest
of the project — your combat system gets the full treatment while everything
else stays light. testing.strict sets, per test type (logic, integration,
visual, UI, config), whether missing evidence blocks a story or merely warns.
Any of the six underlying knobs can still be set individually if you want a mix
(e.g. workflow: full with docs.density: terse — comprehensive but compact).
Two of them — review_mode (full / lean / solo) and team.size — are
also personal: rigor sets the team-wide default, but you can override them
per-developer in a gitignored project.local.yaml without touching the shared
config. modes.automation (how often the AI pauses to ask before it acts) sits
outside rigor and is likewise personally overridable.
View or change any setting with the /settings skill:
/settings # view effective config
/settings modes.review_mode # view one setting
/settings modes.rigor=full # write to project.yaml (team-wide)
/settings --local modes.review_mode=solo # write to project.local.yaml (just you)
Full schema and per-setting behavior: .claude/docs/effects-map.md.
Already using an older version of this template? See UPGRADING.md for step-by-step migration instructions, a breakdown of what changed between versions, and which files are safe to overwrite vs. which need a manual merge.
CLAUDE.md # Master configuration
.claude/
settings.json # Hooks, permissions, safety rules
agents/ # 49 agent definitions (markdown + YAML frontmatter)
skills/ # 74 slash commands (subdirectory per skill)
hooks/ # 18 scripts (bash) — 15 event hooks, two helpers they
# source (yaml-helper.sh, update-check.sh), and one
# opt-in diagnostic you wire yourself
rules/ # 14 path-scoped coding standards
statusline.sh # Status line script (context%, model, stage, epic breadcrumb)
docs/
workflow-catalog.yaml # 7-phase pipeline definition (read by /help)
templates/ # 62 document templates (+ per-section guidance)
src/ # Game source code (Godot). The code root is
# ENGINE-SPECIFIC: Unity compiles only
# Assets/, Unreal builds from Source/<Module>/
assets/ # Art, audio, VFX, shaders, data files
design/ # GDDs, narrative docs, level designs
docs/ # Technical documentation and ADRs
tests/ # Test suites (unit, integration, performance, playtest)
tools/ # Build and pipeline tools
prototypes/ # Throwaway prototypes (isolated from src/)
production/ # Sprint plans, milestones, release tracking
studio-tests/ # QA for the skills and agents themselves — see below
CCGS is a template you are meant to edit. studio-tests/ is how
you check that an edited or newly written skill still holds up — a catalog of all
74 skills and 49 agents, per-category quality rubrics, behavioral specs, and
templates for writing specs of your own.
/skill-test static [name] # structural linter — 7 compliance checks
/skill-test category [name] # category-specific quality rubric
/skill-test spec [name] # behavioral spec assertions
/skill-test audit # coverage report across skills and agents
/skill-improve [name] # test-fix-retest loop; keeps or reverts on score change
It tests the framework, never your game — game tests live in tests/.
Agents follow a structured delegation model:
- Vertical delegation — directors delegate to leads, leads delegate to specialists
- Horizontal consultation — same-tier agents can consult each other but can't make binding cross-domain decisions
- Conflict resolution — disagreements escalate up to the shared parent (
creative-directorfor design,technical-directorfor technical) - Change propagation — cross-department changes are coordinated by
producer - Domain boundaries — agents don't modify files outside their domain without explicit delegation
This is not an auto-pilot system. Every agent follows a strict collaboration protocol:
- Ask — agents ask questions before proposing solutions
- Present options — agents show 2-4 options with pros/cons
- You decide — the user always makes the call
- Draft — agents show work before finalizing
- Approve — nothing gets written without your sign-off
You stay in control. The agents provide structure and expertise, not autonomy.
That is how a terminal session started at the project root behaves out of the
box. The guided and autonomous automation settings, and Claude Code's own
auto mode, trade some of these prompts for speed; see
setup-requirements.md for the routes into
auto mode.
Hooks run automatically on every session:
| Hook | Trigger | What It Does |
|---|---|---|
validate-commit.sh |
PreToolUse (Bash, PowerShell) | Checks for hardcoded values, TODO format, JSON validity, design doc sections — exits early if the command is not git commit |
validate-push.sh |
PreToolUse (Bash, PowerShell) | Warns on pushes to protected branches, and on git tag, git merge, gh release, butler push and steamcmd +run_app_build — exits early for any other command |
validate-read.sh |
PreToolUse (Read) | Refuses a whole-file read of a design doc over 40 KB and answers with its headings, so the retry reads one section (features.read_guard) |
validate-write.sh |
PreToolUse (Write/Edit) | Warns, never blocks, when a write would break a rule the skills state: a file they never overwrite, the config guard, the prototype boundary, a localization freeze |
validate-registry.sh |
PreToolUse (Write/Edit) | Warns when a write to the entity or TR-ID registry would delete or renumber an entry, or set a status its header does not list |
validate-assets.sh |
PostToolUse (Write/Edit) | Checks data-file JSON under assets/, Assets/ or Content/, and asset naming in Godot's assets/ only — exits early if the file is under none of them |
session-start.sh |
Session open | Shows current branch and recent commits for orientation |
detect-gaps.sh |
Session open | Detects fresh projects (suggests /start) and missing design docs when code or prototypes exist |
pre-compact.sh |
Before compaction | Logs the compaction (its output does not reach Claude — session-start.sh restores context afterwards) |
notify.sh |
Permission prompt, or idle about a minute | Shows Windows toast notification via PowerShell |
session-stop.sh |
Every response end, in the background | Archives active.md to session log and records git activity |
log-agent.sh |
Agent spawned | Audit trail start — logs subagent invocation |
log-agent-stop.sh |
Agent stops | Audit trail stop — completes subagent record |
validate-agent-artifact.sh |
Agent stops | Warns when a subagent stops without writing the file its brief named |
validate-skill-change.sh |
PostToolUse (Write/Edit) | Advises running /skill-test after any .claude/skills/ change |
Note:
validate-commit.sh,validate-push.sh,validate-read.sh,validate-write.sh,validate-registry.sh,validate-assets.shandvalidate-skill-change.shfire on every Bash, PowerShell, Read, Write or Edit tool call and exit immediately (exit 0) when the command or file path is not relevant. This is normal hook behavior — not a performance concern.
Permission rules in settings.json auto-allow test runs and block dangerous operations (force push, rm -rf, reading .env files) in both the Bash and PowerShell tools. Read-only commands such as git status, git log and ls need no rule: Claude Code runs them without asking. The git rules match anywhere after git, so a git command whose text only mentions push followed later by + or -f, clean followed by -f, or reset followed by --hard — a commit message such as git commit -m "push + pull" included — is denied too: reword the message.
Coding standards are automatically enforced based on file location. The paths
below are Godot's. Each rule also matches the Unity and Unreal layouts: the code
rules match C# under Assets/**/Gameplay/ and C++ under Source/**/Gameplay/ (and Core,
AI, Networking, UI), data matches JSON under Assets/**/Data/ and
Content/**/Data/, shaders match by file type, and tests match
C# under Assets/Tests/ and C++ under Source/**/Tests/.
| Path | Enforces |
|---|---|
src/gameplay/** |
Data-driven values, delta time usage, no UI references |
src/core/** |
Zero allocations in hot paths, thread safety, API stability |
src/ai/** |
Performance budgets, debuggability, data-driven parameters |
src/networking/** |
Server-authoritative, versioned messages, security |
src/ui/** |
No game state ownership, localization-ready, accessibility |
design/gdd/** |
Required sections per modes.workflow, formula format, edge cases |
design/narrative/** |
Lore consistency, character voice, canon levels |
assets/data/**/*.json |
JSON validity, naming conventions, schema rules |
assets/shaders/** |
Naming conventions, performance targets, cross-platform rules |
**/*.tscn, **/*.tres |
Scene and resource files changed in the editor where possible; safe hand-edit rules |
tests/** |
Test naming, coverage requirements, fixture patterns |
prototypes/** |
Relaxed standards, README required, hypothesis documented |
This template is grounded in professional game development practices:
- MDA Framework — Mechanics, Dynamics, Aesthetics analysis for game design
- Self-Determination Theory — Autonomy, Competence, Relatedness for player motivation
- Flow State Design — Challenge-skill balance for player engagement
- Bartle Player Types — Audience targeting and validation
- Verification-Driven Development — Tests first, then implementation
This is a template, not a locked framework. Everything is meant to be customized:
- Add/remove agents — delete agent files you don't need, add new ones for your domains
- Edit agent prompts — tune agent behavior, add project-specific knowledge
- Modify skills — adjust workflows to match your team's process
- Add rules — create new path-scoped rules for your project's directory structure
- Tune hooks — adjust validation strictness, add new checks
- Pick your engine — use the Godot, Unity, or Unreal agent set (or none)
- Tune process weight —
modes.rigorand its underlying knobs inproject.yaml(see Configuration). Override per-run with--review soloon any skill.
Primary development on Windows 10 with Git Bash. The hooks and scripts are also tested on macOS — its stock bash 3.2 and BSD tools, with and without jq — and on Linux. All hooks use POSIX-compatible patterns (grep -E, not grep -P) and include fallbacks for missing tools. The notify.sh hook uses PowerShell for Windows toast notifications and is a no-op elsewhere — desktop notifications on macOS/Linux are not yet wired. Please file issues for any platform-specific breakage.
- Discussions — GitHub Discussions for questions, ideas, and showcasing what you've built
- Issues — Bug reports and feature requests