Skip to content

smashandclash/agent-studio

BranchesTags

Repository files navigation

Agent Studio

Structured multi-agent game development workspace for modern studios.

Agent Studio rebrands and extends the original Claude template into a cross-host workflow system that works across Claude Code, Cursor, and other SKILL.md-compatible hosts.

  • Claude-compatible layer in .claude/
  • Host-neutral compatibility layer in .agents/
  • Cursor-compatible layer in .cursor/
  • Shared collaboration protocol, role hierarchy, and production workflows

Table of Contents

What Agent Studio Is

Agent Studio is an AI-assisted operating system for game development teams. It combines:

  • role-specialized agents,
  • skill-based workflows,
  • collaboration guardrails,
  • and reusable production templates.

The goal is consistency: your ideation, design, implementation, QA, and release process follows one shared structure regardless of host.

Who This Is For

Agent Studio is designed for:

  • solo developers who want studio-grade process,
  • small teams that need clear role boundaries,
  • game studios adopting AI with guardrails,
  • teams shipping on Godot, Unity, Unreal, or web game stacks.

If you want the assistant to behave like a coordinated team (not a single generalist), this repo is for you.

What Is Included

Area What you get
Agent architecture Leadership, department leads, and specialist agents
Workflow skills Structured commands for ideation, design, implementation, QA, and release
Guardrails Hook scripts and path-based rules for safer execution
Templates Reusable docs for GDDs, ADRs, milestones, and reports
Engine tracks Godot, Unity, Unreal, and Web (Three.js + TypeScript)
Compatibility layer Shared standards in .agents/shared/ with host-specific adapters

Engine and Platform Coverage

Stack Lead agent Sub-specialists
Godot 4 godot-specialist GDScript, shaders, GDExtension
Unity unity-specialist DOTS/ECS, shaders/VFX, addressables, UI
Unreal Engine 5 unreal-specialist GAS, Blueprints, replication, UMG
Web game stack web-specialist Three.js, TypeScript web game architecture

Cross-Host Compatibility Model

Claude workflow

  • Start from README.md, CLAUDE.md, and .claude/docs/quick-start.md.
  • Use generated .claude/skills/ adapters and .claude/agents/.

Cursor workflow

  • Start from .cursor/README.md.
  • Use .cursor/rules/ for Cursor-oriented overlays.
  • Keep behavior aligned with canonical standards in .agents/shared/.

Universal layer

  • Canonical commands: .agents/skills/
  • Universal protocol and standards: .agents/shared/
  • Generated Claude adapters: .claude/skills/
  • Generated top-level install surface: skills/

Shared principle

All layers follow the same collaboration model:

ask -> options -> decision -> draft -> approval

Quick Start

Quick Start (Cursor)

  1. Clone this repo and open it in Cursor.
  2. Read these first:
    • README.md
    • .cursor/README.md
    • CLAUDE.md
  3. Review active overlays in .cursor/rules/.
  4. Choose and configure your engine path (use /setup-engine flow in compatible host sessions).
  5. Start from /start if this is your first session.
  6. Move into design with /brainstorm or /map-systems.
  7. Run delivery workflows (/sprint-plan, /team-*, /qa-web, /ship-release) as you progress.

Quick Start (Claude Code)

  1. Clone this repo.
  2. Open in Claude Code.
  3. Read:
    • README.md
    • CLAUDE.md
    • .claude/docs/quick-start.md
  4. Configure engine/version with /setup-engine.
  5. Start with /start.
  6. Use stage-driven skills from docs/WORKFLOW-GUIDE.md.

Cursor Usage Guide

This section is intentionally detailed to make Cursor onboarding straightforward.

1) Understand where behavior comes from

  • .cursor/rules/ = Cursor-specific overlays and routing behavior.
  • .agents/shared/ = host-neutral protocol and values.
  • .agents/skills/ = canonical skill definitions used across hosts.

If something appears inconsistent, treat .agents/shared/ and .agents/skills/ as source-of-truth and regenerate adapters.

2) Use rule overlays as workflow switches

You have dedicated overlays in .cursor/rules/:

  • agent-routing.mdc
  • studio-collaboration.mdc
  • investigation-first.mdc
  • safety-toggles.mdc
  • sprint-pipeline.mdc
  • web-game-track.mdc

These are designed to keep Cursor behavior aligned with the same production model documented in CLAUDE.md and .agents/shared/.

3) Follow this operating loop in Cursor

  1. Frame goal and constraints (what feature, by when, quality bar).
  2. Choose workflow command (/investigate, /design-system, /team-web, etc.).
  3. Request options and tradeoffs before implementation.
  4. Approve a draft direction.
  5. Execute with guardrails (/careful, /freeze, /guard when needed).
  6. Run quality passes (/code-review, /perf-profile, /qa-web, /gate-check).
  7. Prepare release artifacts (/release-checklist, /patch-notes, /team-release).

4) When to use safety toggles

Use these for high-risk periods such as large refactors, release week, or hotfixes:

  • /careful — extra caution before risky operations
  • /freeze — restrict edits to approved boundaries
  • /guard — combined careful + freeze semantics
  • /unfreeze — return to normal behavior

5) Where to find realistic session patterns

See:

  • docs/examples/README.md
  • docs/examples/session-design-crafting-system.md
  • docs/examples/session-implement-combat-damage.md
  • docs/examples/session-scope-crisis-decision.md

These examples show the expected collaborative behavior (questions, options, approval gates).

How Cursor Rules Map to Studio Workflows

  • sprint-pipeline.mdc aligns with: /team-web, /qa-web, /ship-release, /retro-weekly
  • investigation-first.mdc aligns with: /investigate
  • safety-toggles.mdc aligns with: /careful, /freeze, /guard, /unfreeze

This mapping helps you quickly switch process modes without changing core architecture.

Recommended Cursor Session Pattern

Use this script-like pattern in daily work:

  1. "I need to implement X under constraints Y and Z."
  2. Ask for 2–3 options with tradeoffs.
  3. Choose one option explicitly.
  4. Ask for a phased execution plan.
  5. Implement phase-by-phase with review points.
  6. Run verification for that phase.
  7. Continue until done, then run release checks.

This keeps creative control with you while using the system as a disciplined execution layer.

Example Use Cases

1) I’m starting from zero

  • Run /start
  • Run /brainstorm
  • Run /map-systems
  • Run /design-system <system-name> for critical systems
  • Run /prototype <core mechanic>
  • Run /sprint-plan new

2) I have a design doc and need implementation

  • Run /investigate to validate assumptions and constraints
  • Run /team-web (web stack) or domain team skill for implementation
  • Run /code-review
  • Run /qa-web (if browser-facing) or QA workflow skills
  • Run /ship-release

3) My sprint is slipping

  • Run /scope-check production/sprints/<sprint-file>.md
  • Run /sprint-plan status
  • Run /milestone-review "<milestone>"
  • Run /retrospective or /retro-weekly

4) I need release readiness confidence

  • Run /release-checklist <version-tag>
  • Run /launch-checklist
  • Run /patch-notes <version-tag>
  • Run /changelog <version-tag>
  • Run /team-release

5) Production incident / emergency fix

  • Run /hotfix "<incident summary>"
  • Run /investigate
  • Apply fix with safety toggles enabled (/guard)
  • Re-run release validation (/ship-release, /launch-checklist)

Workflow by Project Stage

Use docs/WORKFLOW-GUIDE.md as the authoritative phase guide:

  • Phase 0: Setup & Configuration
  • Phase 1: Ideation & Concept
  • Phase 2: Pre-Production & Design
  • Phase 3: Prototyping & Validation
  • Phase 4: Production Sprint Workflow
  • Phase 5: Implementation Deep-Dive
  • Phase 6: Testing & QA
  • Phase 7: Polish & Optimization
  • Phase 8: Localization & Accessibility
  • Phase 9: Release & Launch
  • Phase 10: Post-Launch & Live Ops

If uncertain where to start, run /start or /project-stage-detect.

Key Commands You Will Use Most

Goal Command
Onboard and route workflow /start
Configure engine/version references /setup-engine
Create game concepts /brainstorm
Decompose systems /map-systems
Author a system GDD /design-system
Run focused implementation teams /team-combat, /team-ui, /team-web, etc.
Investigate before fixing /investigate
Plan and track sprint /sprint-plan, /scope-check
Review technical direction /architecture-decision, /code-review
Verify gameplay quality /balance-check, /playtest-report, /perf-profile
Prepare release artifacts /release-checklist, /launch-checklist, /patch-notes, /changelog
Stabilize risk /careful, /freeze, /guard, /unfreeze

For a full command walkthrough, use docs/WORKFLOW-GUIDE.md and .claude/docs/quick-start.md.

Repository Layout

CLAUDE.md
.claude/
  agents/
  skills/
  hooks/
  rules/
  docs/
.agents/
  shared/
  skills/
  setup/
.cursor/
  README.md
  rules/
skills/
docs/
  WORKFLOW-GUIDE.md
  examples/
  engine-reference/

Canonical vs Generated Skills

Skill sources and adapters follow this model:

  • Canonical skill definitions: .agents/skills/
  • Generated Claude adapters: .claude/skills/
  • Generated top-level install surface: skills/

Generation scripts:

  • .agents/setup/sync-skills.sh
  • .agents/setup/sync-skills.ps1
  • .agents/setup/generate-all-adapters.sh
  • .agents/setup/generate-all-adapters.ps1

Common commands:

# Linux/macOS
bash .agents/setup/sync-skills.sh generate-claude
bash .agents/setup/sync-skills.sh generate-toplevel
bash .agents/setup/generate-all-adapters.sh
# Windows PowerShell
pwsh .agents/setup/sync-skills.ps1 generate-claude
pwsh .agents/setup/sync-skills.ps1 generate-toplevel
pwsh .agents/setup/generate-all-adapters.ps1

Optional Skills.sh Extensions

This project can consume public SKILL.md extensions from https://skills.sh/.

  • Curated list: .agents/skills/skills.sh-curated.md
  • Installer pattern: npx skills add <owner/repo>

Use these as optional add-ons while keeping game-studio workflow canonical in .agents/skills/.

Customization Guidance

You can tailor Agent Studio safely by:

  • keeping only the agent domains you use,
  • keeping canonical skill definitions in .agents/skills/,
  • regenerating adapters after skill edits,
  • preserving shared protocol files in .agents/shared/.

Recommended customization sequence:

  1. Pick your engine track.
  2. Prune unused domains conservatively.
  3. Update technical preferences (.claude/docs/technical-preferences.md).
  4. Validate command/rule consistency.
  5. Regenerate adapters.

Documentation References

  • Main workflow phases: docs/WORKFLOW-GUIDE.md
  • Collaboration principle: docs/COLLABORATIVE-DESIGN-PRINCIPLE.md
  • Cursor compatibility notes: .cursor/README.md
  • Host-neutral compatibility layer: .agents/README.md
  • Claude quick start: .claude/docs/quick-start.md
  • Engine references: docs/engine-reference/README.md
  • Example sessions: docs/examples/README.md

Project Notes

  • Organization: Smash&Clash
  • Author: HKTITAN
  • Product/site: smashandclash.in
  • This repo is the AI studio operating system maintained by Smash&Clash.

Design Value Sources

We incorporate practical insights from:

These are codified in .agents/shared/experience-values.md and reflected in UI rules.

Upgrading

See UPGRADING.md for template upgrade strategy and merge guidance.

FAQ / Troubleshooting

Which docs should I read first?

For most users:

  1. README.md
  2. .cursor/README.md (if using Cursor) or .claude/docs/quick-start.md (if using Claude)
  3. CLAUDE.md
  4. docs/WORKFLOW-GUIDE.md

I changed a canonical skill but behavior looks stale.

Regenerate adapters:

  • .agents/setup/sync-skills.sh full (or PowerShell equivalent)
  • .agents/setup/generate-all-adapters.sh

Where should I add new skills?

Add/modify skills in .agents/skills/, then regenerate adapters. Do not treat generated adapter folders as canonical sources.

How do I keep sessions disciplined in Cursor?

Use the collaborative loop:

  • define constraints,
  • request options,
  • choose explicitly,
  • approve drafts,
  • execute in phases,
  • verify before release.

Use safety toggles (/careful, /freeze, /guard) during risky operations.

What if I already have a project in progress?

Run /project-stage-detect or /start, then follow phase-appropriate workflows from docs/WORKFLOW-GUIDE.md.

Original Project and Attribution

This project is based on the original upstream template: Donchitos/Claude-Code-Game-Studios.

The actively maintained Agent Studio repository is: smashandclash/agent-studio.

Inherited foundation includes:

  • agent hierarchy and domain-specialist model,
  • skill-driven game-dev workflows,
  • hook/rule-centered quality gates.

Agent Studio adaptations include:

  • rebrand to Agent Studio,
  • Cursor-first compatibility layer while preserving .claude support,
  • expanded engine coverage with web game specialists (Three.js + TypeScript).

License

MIT License. See LICENSE.

About

multi-agent game studio workflow

smashandclash.in

Topics

gamedev ai studio workflows agents

Resources

Readme

License

MIT license
Activity
Custom properties

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages