From aca4fce362e7b58f09b1158a40d08d07937569f3 Mon Sep 17 00:00:00 2001
From: milkomida77 <145509032+milkomida77@users.noreply.github.com>
Date: Mon, 6 Apr 2026 14:12:29 +0200
Subject: [PATCH] docs: add README for Guardian AI directory

---
 Open Source prompts/Guardian AI/README.md | 41 +++++++++++++++++++++++
 1 file changed, 41 insertions(+)
 create mode 100644 Open Source prompts/Guardian AI/README.md

diff --git a/Open Source prompts/Guardian AI/README.md b/Open Source prompts/Guardian AI/README.md
new file mode 100644
index 00000000..18d39ded
--- /dev/null
+++ b/Open Source prompts/Guardian AI/README.md	
@@ -0,0 +1,41 @@
+# Guardian AI
+
+**Type**: Multi-agent orchestration system (open source)
+**Model**: Claude (Opus/Sonnet), also works with GPT, Gemini, Llama, Mistral
+**Agents**: 57 specialized agents coordinated by a single orchestrator
+**Production**: 10,000+ tasks over 6+ months
+**Source**: https://github.com/milkomida77/guardian-agent-prompts
+
+## Prompts in this directory
+
+| File | Description |
+|------|-------------|
+| orchestrator-system-prompt.txt | The main orchestrator that routes all 57 agents. Handles task decomposition, anti-duplication, quality gates, and parallel agent execution. |
+
+## Architecture
+
+Guardian uses a hub-and-spoke model:
+- **1 Orchestrator** (this prompt) routes ALL incoming tasks
+- **57 Specialized Agents** handle specific domains (code, security, trading, OSINT, business, etc.)
+- **Task Registry** prevents duplicate work across agents
+- **Quality Gates** require verification evidence before marking tasks done
+- **Knowledge Graph** provides persistent memory across sessions
+
+## Key Patterns
+
+1. **Identity + NOT-block**: Each agent defines what it IS and what it IS NOT (reduces task drift ~35%)
+2. **Task Registry**: SQLite-based anti-duplication with similarity matching
+3. **Setup Master**: Every task gets a blueprint before delegation (agents, tools, risks, order)
+4. **Quality Gate**: Agent output is a CLAIM; test output is EVIDENCE
+5. **30-minute Heartbeat**: Orchestrator checks progress and reassigns stale tasks
+
+## Differences from other agent systems
+
+| Feature | Guardian | Typical Agent Frameworks |
+|---------|----------|------------------------|
+| Prompt length | 200-800 lines | 20-50 lines |
+| Constraint ratio | 20-30% | <5% |
+| NOT-blocks | Every agent | Rare |
+| Task registry | Built-in | Not included |
+| Quality gates | Mandatory | Optional |
+| Error handling | Explicit per failure mode | Generic retry |