YetiClaw Game Studio Agents

By Yeti Games, LLC

27 specialized AI agents for game development, AI consulting, avatar creation, and business outreach — running privately on your own hardware via Picoclaw and Qwen3.5 4B.

Docs

README.md — setup and installation
WORKFLOW.md — how to brief agents and run a game project
AGENTS.md — quick slash command reference

What Is This?

A complete Picoclaw skill pack that turns your private hardware into a full AI game studio. Core workflows run entirely locally — no cloud subscription required for day-to-day use. Optional paid APIs (Nano Banana 2, Meshy) are gated behind a human approval step so you stay in control of spend.

Works on:

🍊 Orange Pi 6 Plus (ARM64) — the original YetiClaw setup
🍎 macOS (Apple Silicon or Intel)
🪟 Windows 10/11 (x64) — via Ollama + OpenClaw
🐧 Any Linux x86_64 or ARM64 machine

Creator Workflow — Brief to Build

The producer is the entry point for any new game project. Give it your full creative vision — story, art style, platform, mechanics, everything — and it writes a structured brief directly, then offers to expand any section with a specialist.

Step 1 — Brief creation

Just tell the producer what you want. No need to simplify or break it into steps:

/producer write a brief for a Pac-Man style arcade game for VIVERSE.
Browser-based, mobile and desktop controls. Classic maze gameplay with modern
twists, power-ups, and online leaderboards. Pixel art style, all ages.

The producer reads your full vision and writes the brief immediately — no spawning, no waiting. You get back a structured document with game name, concept, mechanics, visual style, platform, timeline, and budget.

Step 2 — Save the brief

save

Saves to projects/[slug]/brief.md in workspace and syncs to Google Drive.

Step 3 — Expand sections with specialists

Once the brief exists, expand any section one at a time:

expand mechanics
expand narrative
expand three.js
expand levels
expand art

Each expansion spawns ONE specialist agent focused on that section only.

Step 4 — Hand off to specialists

With the brief saved, any agent can read it directly:

/threejs-dev scaffold the project structure for the Pac-Man game
/game-designer design the power-up system
/narrative-director write the game's story and character bios

Agent Roster

🎬 Directors

Command	Role
`/creative-director`	Creative vision, tone, aesthetic, MDA review
`/technical-director`	Unity/Three.js architecture decisions (code-gen)
`/producer`	Game brief creation + sprint planning, milestones, blockers

🎮 Game Design

Command	Role
`/game-designer`	Mechanics, GDDs, MDA framework, Bartle player types
`/level-designer`	Level layouts, flow, pacing, encounters
`/systems-designer`	Economy, progression, data-driven systems

💻 Programming (code-gen — Unity Editor not required on this machine)

Command	Role
`/gameplay-programmer`	Unity C# player mechanics, movement, combat
`/engine-programmer`	Unity C# core: save systems, scene management, Addressables
`/ai-programmer`	Unity C# enemy AI, behavior trees, NavMesh
`/ui-programmer`	Unity UI Toolkit, HUD, menus, accessibility
`/unity-specialist`	Package conflicts, editor scripting, render pipeline

🎨 Art & Audio

Command	Role
`/art-director`	Visual style guide, art briefs, concept generation (Nano Banana 2)
`/sound-designer`	Audio identity, FMOD events, adaptive music
`/technical-artist`	Shaders, VFX Graph, asset import pipeline (code-gen)

📖 Narrative & QA

Command	Role
`/narrative-director`	Story structure, character arcs, narrative design
`/writer`	Dialogue, UI copy, item descriptions, voice guide
`/world-builder`	Lore, factions, geography, world bible
`/qa-tester`	Test plans, bug reports (Critical / High / Medium / Low)

🌐 Web & XR

Command	Role
`/threejs-dev`	TypeScript Three.js games for VIVERSE (WebGL/WebXR)

💼 Business

Command	Role
`/email-writer`	Outreach for game dev · AI integration · AI app dev
`/ai-consultant`	AI strategy and consulting for studios and SMBs

💰 Asset Generation (approval required before any paid API fires)

Command	Role
`/asset-approver`	Budget gate — shows cost, waits for `/approve` or `/deny`
`/meshy`	Meshy.ai text-to-3D and image-to-3D (GLB output)

Asset Generation Workflow

/threejs-dev or /art-director
    → produces Asset Brief
        → /asset-approver [brief]
            → shows cost estimate in Telegram
                → you reply /approve [name] or /deny [name]
                    ✓ approve → Nano Banana 2 or /meshy executes
                    ✗ deny   → placeholder used, no spend

Nano Banana 2 (gemini-3.1-flash-image-preview) handles image generation and understanding.

Meshy — text-to-3D and image-to-3D, GLB output saved to Google Drive.

Prerequisites

All platforms

A Telegram bot token
A Gemini API key (for Nano Banana 2)
A Meshy API key (for 3D generation, Pro plan required)
rclone configured with a gdrive remote
Node.js 20+ (for Google Drive and Gmail MCP servers)

Orange Pi 6 Plus only

Ubuntu 24.04, 16GB RAM minimum
Custom Picoclaw binary with 600s timeout fix (see below)

macOS only

Ollama — serves models locally
Apple Silicon recommended

Windows only

Ollama for Windows — OllamaSetup.exe, no admin rights required
NVIDIA GPU recommended (8GB+ VRAM) or modern CPU with AVX2

Installation

🍊 Orange Pi 6 Plus

Note: The base installer (llama.cpp, model download, Picoclaw, systemd services) is not included in this repo. Contact me for installation services.

Step 1 — Deploy the agents

tar -xzf yeticlaw-studio.tar.gz
scp -r yeticlaw-studio/ orangepi@[your_ip]:/tmp/
ssh orangepi@[your_ip]
sudo bash /tmp/yeticlaw-studio/deploy.sh

Step 2 — Configure secrets

nano ~/.picoclaw/.security.yml

channels:
  telegram:
    token: YOUR_TELEGRAM_BOT_TOKEN
model_list:
  nano-banana-2:
    api_key: YOUR_GEMINI_API_KEY
skills: {}
web: {}

chmod 600 ~/.picoclaw/.security.yml

Also set channels.telegram.token in config.json for compatibility.

Step 3 — Add Meshy key

sudo nano /etc/environment
# Add: MESHY_API_KEY=your_meshy_key_here
sudo systemctl restart yeticlaw-gateway

Step 4 — Set up Google Drive

su - orangepi
rclone config  # n → gdrive → drive → follow OAuth
npx @piotr-agier/google-drive-mcp auth
npx @gongrzhe/server-gmail-autoauth-mcp auth

🍎 macOS

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Install other tools
brew install rclone jq node

# Launch OpenClaw with gemma4:e4b
ollama launch openclaw --model gemma4:e4b --yes

# Deploy agents
tar -xzf yeticlaw-studio.tar.gz && cd yeticlaw-studio
bash deploy-mac.sh

🪟 Windows

Step 1 — Install Ollama

Download and run OllamaSetup.exe from ollama.com/download/windows.
No administrator rights required. Or use PowerShell:

irm https://ollama.com/install.ps1 | iex

Ollama runs in the background automatically and serves on http://localhost:11434.

Step 2 — Launch OpenClaw with gemma4:e4b

ollama launch openclaw --model gemma4:e4b --yes

Ollama will download gemma4:e4b (~9.6GB) on first run and launch the OpenClaw setup wizard to configure your Telegram bot token.

Step 3 — Deploy agents

Open PowerShell in the yeticlaw-studio folder:

# Create workspace folders
$workspace = "$env:USERPROFILE\.openclaw\workspace\skills"
New-Item -ItemType Directory -Force -Path $workspace

# Copy SOUL.md + AGENTS.md
Copy-Item SOUL.md "$env:USERPROFILE\.openclaw\workspace\SOUL.md"
Copy-Item AGENTS.md "$env:USERPROFILE\.openclaw\workspace\AGENTS.md"

# Install all skills from skills-mac/ (same skills, works on Windows)
Get-ChildItem skills-mac -Directory | ForEach-Object {
    $dst = "$workspace\$($_.Name)"
    New-Item -ItemType Directory -Force -Path $dst | Out-Null
    Copy-Item "$($_.FullName)\SKILL.md" "$dst\SKILL.md"
    Write-Host "  ✓ $($_.Name)"
}

Write-Host "Done — $((Get-ChildItem $workspace -Directory).Count) skills installed"

Step 4 — Configure secrets

Edit the OpenClaw config (created during ollama launch openclaw setup):

%USERPROFILE%\.openclaw\openclaw.json

Set your Telegram bot token and Gemini API key (for Nano Banana 2 / image generation).

Step 5 — Set up Google Drive (optional)

winget install Rclone.Rclone
rclone config  # n → gdrive → drive → follow OAuth

Notes

GPU: NVIDIA cards use CUDA automatically. AMD cards use DirectML. CPU-only works but is slower.
Model storage: Ollama stores models in %USERPROFILE%\.ollama\models by default. Set OLLAMA_MODELS env var to change this.
Service install: For always-on use, install as a Windows service with NSSM: nssm install ollama "C:\Users\you\AppData\Local\Programs\Ollama\ollama.exe" serve
No deploy.sh: The bash deploy script won't run on Windows. Use the PowerShell commands above instead.

Important: Custom Binary Required (Orange Pi)

The Orange Pi setup requires a custom-built Picoclaw binary with the HTTP timeout increased from 120s to 600s. The official binary cancels responses before the model finishes generating on slow hardware.

Build it on your Mac:

git clone --depth=1 https://github.com/sipeed/picoclaw.git
cd picoclaw
sed -i '' 's/const DefaultRequestTimeout = 120 \* time.Second/const DefaultRequestTimeout = 600 * time.Second/' pkg/providers/common/common.go
GOOS=linux GOARCH=arm64 go build -tags "goolm,stdjson" -ldflags "-s -w" -o picoclaw-linux-arm64 ./cmd/picoclaw
scp picoclaw-linux-arm64 orangepi@[your_ip]:/tmp/
ssh orangepi@[your_ip]
sudo cp /tmp/picoclaw-linux-arm64 /usr/local/bin/picoclaw && sudo chmod +x /usr/local/bin/picoclaw

On macOS with Ollama, the official binary works fine.

Troubleshooting

Gateway exits silently after loading Check .security.yml exists with correct token. Verify config.json has "version": 1 (integer) and channels.telegram.enabled: true. Clean corrupt cron store:

echo '{"version":1,"jobs":[]}' > /opt/yeticlaw/openclaw/workspace/cron/jobs.json
sudo systemctl restart yeticlaw-gateway

Responses timing out (Orange Pi) Install the custom binary with 600s timeout — see section above.

Skill not loading / description too long Skill descriptions must be under 1024 characters. Check frontmatter:

head -4 /opt/yeticlaw/openclaw/workspace/skills/[skill-name]/SKILL.md

llama-server SEGV on startup Binary is corrupted. Rebuild from source:

sudo rm -rf /opt/llama.cpp
sudo git clone --depth=1 https://github.com/ggml-org/llama.cpp /opt/llama.cpp
sudo cmake -B /opt/llama.cpp/build /opt/llama.cpp -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH="armv9-a+sve+i8mm+dotprod" -DGGML_CPU_KLEIDIAI=ON -DCMAKE_BUILD_TYPE=Release
sudo cmake --build /opt/llama.cpp/build --config Release -j$(nproc)
sudo install -m 755 /opt/llama.cpp/build/bin/llama-server /usr/local/bin/
sudo systemctl start llama-server

rclone not syncing

rclone ls gdrive:YetiClaw
rclone config reconnect gdrive:

Platform Differences

Feature	Orange Pi 6 Plus	macOS	Windows
Model server	llama.cpp (KleidiAI optimized)	Ollama	Ollama
Model	Qwen3.5 4B	gemma4:e4b	gemma4:e4b
Speed	~12 tok/s gen	~57 tok/s (M-series)	varies (GPU recommended)
Picoclaw binary	Custom build (600s timeout)	Official release	Official release
Gateway service	systemd	LaunchAgent or foreground	Foreground or NSSM service
Deploy script	`deploy.sh` (sudo required)	`deploy-mac.sh` (no sudo)	PowerShell (manual)
Workspace path	`/opt/yeticlaw/openclaw/workspace`	`~/.openclaw/workspace`	`%USERPROFILE%\.openclaw\workspace`

File Structure

yeticlaw-studio/
├── README.md
├── AGENTS.md               ← Quick reference for all slash commands
├── SOUL.md                 ← Workspace routing rules (loaded every message)
├── config.json             ← Picoclaw config
├── deploy.sh               ← Orange Pi deploy script
├── deploy-mac.sh           ← macOS deploy script
└── skills/
    ├── producer/           ← Brief creation + sprint planning
    ├── creative-director/
    ├── technical-director/
    ├── game-designer/
    ├── level-designer/
    ├── systems-designer/
    ├── gameplay-programmer/
    ├── engine-programmer/
    ├── ai-programmer/
    ├── ui-programmer/
    ├── unity-specialist/
    ├── art-director/
    ├── sound-designer/
    ├── technical-artist/
    ├── narrative-director/
    ├── writer/
    ├── world-builder/
    ├── qa-tester/
    ├── threejs-dev/
    ├── ai-consultant/
    ├── email-writer/
    ├── asset-approver/
    ├── meshy/
    ├── game-namer/         ← Atomic: generates game name options
    ├── concept-writer/     ← Atomic: writes 2-sentence concept
    ├── mechanics-designer/ ← Atomic: lists 5 core mechanics
    └── style-writer/       ← Atomic: describes visual style

License

MIT — see LICENSE

Skills are compatible with both Picoclaw and OpenClaw SKILL.md format.

Unity MCP (Mac Mini only)

On a Mac Mini with Unity Editor installed, you can connect Picoclaw directly to Unity via MCP. This enables agents to write C# scripts directly into your Unity project, trigger compilation, read console errors, and fix them automatically.

Setup

Install Unity 6+ with the MCP package from the Package Manager
Enable the Unity MCP server in ~/.picoclaw/config.json:

"unity": {
    "enabled": true,
    "command": "npx",
    "args": ["-y", "@unity/mcp-server"]
}

Open your Unity project before starting the gateway

What this enables

/gameplay-programmer implement the ghost AI behaviour
→ writes PlayerBrewing.cs directly into Unity via MCP
→ triggers compile
→ reads console errors
→ fixes errors automatically
→ "✅ PlayerBrewing.cs compiled successfully, no errors"

No copy-paste required. The agent writes, compiles, and debugs in a live Unity session.

Skill Tiers

Directory	Target	Model	Writing style
`skills/`	Orange Pi	Qwen3.5 4B	Chunked — one file at a time
`skills-mac/`	Mac / Windows	gemma4:e4b	Full context — complete files in one shot

The Pi deploy (deploy.sh) uses skills/. The Mac deploy (deploy-mac.sh) uses skills-mac/.

Name		Name	Last commit message	Last commit date
Latest commit History 30 Commits
docs		docs
skills-mac		skills-mac
skills		skills
.gitignore		.gitignore
AGENTS.md		AGENTS.md
LICENSE		LICENSE
PROJECTS.md		PROJECTS.md
README.md		README.md
SOUL.md		SOUL.md
WORKFLOW.md		WORKFLOW.md
config.json		config.json
deploy-mac.sh		deploy-mac.sh
deploy.sh		deploy.sh

Folders and files

Latest commit

History

Repository files navigation

YetiClaw Game Studio Agents

By Yeti Games, LLC

Docs

What Is This?

Creator Workflow — Brief to Build

Step 1 — Brief creation

Step 2 — Save the brief

Step 3 — Expand sections with specialists

Step 4 — Hand off to specialists

Agent Roster

🎬 Directors

🎮 Game Design

💻 Programming (code-gen — Unity Editor not required on this machine)

🎨 Art & Audio

📖 Narrative & QA

🌐 Web & XR

💼 Business

💰 Asset Generation (approval required before any paid API fires)

Asset Generation Workflow

Prerequisites

All platforms

Orange Pi 6 Plus only

macOS only

Windows only

Installation

🍊 Orange Pi 6 Plus

Step 1 — Deploy the agents

Step 2 — Configure secrets

Step 3 — Add Meshy key

Step 4 — Set up Google Drive

🍎 macOS

🪟 Windows

Step 1 — Install Ollama

Step 2 — Launch OpenClaw with gemma4:e4b

Step 3 — Deploy agents

Step 4 — Configure secrets

Step 5 — Set up Google Drive (optional)

Notes

Important: Custom Binary Required (Orange Pi)

Troubleshooting

Platform Differences

File Structure

License

Unity MCP (Mac Mini only)

Setup

What this enables

Skill Tiers

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages