Quick Start Guide

Deploy Guardian agent prompts in under 5 minutes with any LLM.

Option 1: Single Agent (Any LLM)

Works with Claude, ChatGPT, Gemini, Llama, Mistral, or any LLM that supports system instructions.

Step 1: Download a Prompt

Free sample:

Full bundle (49 prompts):

Step 2: Configure for Your System

Each prompt uses consistent placeholder variables. Replace these with your values:

[PROJECT_NAME]     -> your system name (e.g., "Acme AI")
[DOMAIN]           -> your domain or workspace (e.g., "acme.com")
[AGENT_INBOX]      -> path to message queue (e.g., "/data/messages/")
[TASK_REGISTRY]    -> path to coordination DB (e.g., "/data/tasks.db")
[SLACK_CHANNEL]    -> team channel (e.g., "#ai-agents")

A single find-and-replace across all files configures the entire system for your environment.

Step 3: Load as System Instruction

Claude (claude.ai):

  1. Start a new conversation
  2. Click the system prompt area or use Projects
  3. Paste the prompt content
  4. Start chatting

ChatGPT:

  1. Go to “Create a GPT” or use the API
  2. Paste the prompt as the system instruction
  3. Save and test

Gemini:

  1. Use Google AI Studio
  2. Paste as system instruction
  3. Test with sample tasks

API (any provider):

response = client.messages.create(
    model="claude-sonnet-4-20250514",
    system=open("prompts/orchestrator.md").read(),
    messages=[{"role": "user", "content": "Route this: we have a security vulnerability in our API"}]
)

Step 4: Test

Send this test message to verify the agent is working:

Route this request to the right specialist:
we have a security vulnerability in our API endpoint.

The Orchestrator should:

Option 2: Multi-Agent System (Claude Code)

For deploying the full 49-agent network with Claude Code (2.1.77+).

Prerequisites

Step 1: Clone and Install

# Clone this repository
git clone https://github.com/milkomida77/guardian-agent-prompts
cd guardian-agent-prompts

# Copy prompts to Claude agents directory
mkdir -p ~/.claude/agents
cp prompts/*.md ~/.claude/agents/

Step 2: Launch an Agent

# Launch the Orchestrator with a named terminal
claude -n "Orchestrator" --agent guardian-orchestrateur

# Launch a specialist agent
claude -n "Codex" --agent guardian-codex

# List all available agents
claude agents

Step 3: Configure Communication

For agents to communicate, set up at least one of these:

Option A: File-based (simplest)

# Create message directories
mkdir -p /path/to/your/project/messages/

# Each agent reads/writes to its own file
# Built into the prompts -- just set the path

Option B: Slack (recommended for teams)

# Set up Slack MCP or webhook
# Configure channel ID in each prompt
# Agents post status updates and delegation requests

Option C: N8N (recommended for automation)

# Deploy N8N (self-hosted or cloud)
# Create webhook workflows for each agent
# Agents call webhooks for cross-agent communication

Step 4: Set Up Task Registry

The task registry prevents duplicate work:

# Create SQLite database
python3 -c "
import sqlite3
conn = sqlite3.connect('task_registry.db')
conn.execute('''CREATE TABLE IF NOT EXISTS tasks (
    id TEXT PRIMARY KEY,
    description TEXT,
    agent TEXT,
    status TEXT DEFAULT 'claimed',
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    completed_at TIMESTAMP
)''')
conn.commit()
print('Task registry created')
"

Step 5: Verify

# Check agent loaded correctly
claude -n "Test" --agent guardian-orchestrateur

# Send a test task
# "What agents are available and what are their specialties?"

# The Orchestrator should list its delegation taxonomy

Option 3: Start Small, Scale Up

You do not need all 49 agents. Here are recommended starting configurations:

Solo Developer (2-3 agents)

Orchestrator + Codex (code) + one domain specialist

Good for: personal projects, learning the multi-agent pattern

Small Team (5-7 agents)

Orchestrator + Codex + Security + Business + Infrastructure + Memory

Good for: startups, side projects with revenue goals

Full Stack (all 49)

All agents, organized by the 8 categories in the architecture doc

Good for: production systems, complex automation, 24/7 operations

Supported Platforms

Platform Deployment Method Notes
Claude Code --agent flag Native, full support
Claude API System instruction All features work
ChatGPT (GPT Builder) System instruction All features work
ChatGPT API System message All features work
Gemini (AI Studio) System instruction Good, delegation precision varies
Gemini API System instruction Good
Llama (local) System prompt 70B+ recommended
Mistral System instruction Strong instruction-following
N8N AI Agent node System message Workflow integration

Common Issues

Agent does not delegate

Cause: The agent does not know about other agents in the system. Fix: Ensure the agent prompt includes the delegation taxonomy (which agents exist and what they handle).

Agents duplicate work

Cause: No task registry configured. Fix: Set up the SQLite task registry (Step 4 above). Each agent checks before starting work.

Agent loses context between sessions

Cause: No memory persistence configured. Fix: Set up a memory system (knowledge graph, file-based memory, or database). The prompts include memory protocol instructions.

Agent ignores instructions

Cause: System prompt too long for the model’s effective instruction-following window. Fix: For smaller models (7B-13B), use the condensed versions. For larger models, the full prompts work as-is.

Next Steps

  1. Read the Architecture doc: architecture.md for the full system design
  2. Read the Protocol: agent-protocol.md for the shared rules all agents follow
  3. Join the community: Star the GitHub repo for updates
  4. Get the full bundle: 49 prompts on Gumroad (code LAUNCH49 = $10 off)

Built by the Guardian AI team. Running in production since January 2025.