🐝 Hive-Synthesis-Engine

A multi-agent prompt engineering toolkit for Claude.

Six specialized AI researcher personas collaborate through a structured pipeline to generate production-grade prompts, strategies, skill files, and prompt chains — optimized specifically for Anthropic's Claude.

---

What This Is

Most prompt engineering is ad-hoc. You write a prompt, test it, tweak it, repeat. Hive-Mind systematizes that process through:

1. Structured decomposition — Every request is analyzed across 6 specialist dimensions before a single word of prompt is written
2. Complexity calibration — Prompts are sized to the task (50 tokens for simple, 1500 for complex). No bloat.
3. Adversarial validation — Every output is stress-tested for failure modes, edge cases, and ambiguity
4. Iterative refinement — Built-in critique loops that progressively harden outputs

This isn't a wrapper around "write me a prompt." It's an engineering methodology.

---

Repository Structure

Hive-Mind/
│
├── README.md                          # You are here
├── LICENSE                            # Apache 2.0 License
├── .gitignore
│
├── HIVE_SYNTHESIS_ENGINE.md           # ⭐ Core engine — the master prompt
│
├── prompts/
│   ├── README.md                      # How to use prompt templates
│   ├── templates/
│   │   ├── simple.md                  # Simple task template (50-150 tokens)
│   │   ├── medium.md                  # Medium task template (150-500 tokens)
│   │   └── complex.md                 # Complex task template (500-1500 tokens)
│   └── chains/
│       ├── research-and-write.md      # Multi-step research → synthesis chain
│       └── code-review-pipeline.md    # Code analysis → fix → verify chain
│
├── strategies/
│   ├── README.md                      # Strategy index
│   ├── chain-of-thought.md            # CoT patterns for Claude
│   ├── few-shot-design.md             # Example selection & formatting
│   ├── xml-tag-architecture.md        # Structural optimization for Claude
│   └── adversarial-hardening.md       # Making prompts robust
│
├── skills/
│   ├── README.md                      # How to create & use skill files
│   ├── SKILL_TEMPLATE.md              # Blank skill file template
│   └── examples/
│       ├── code-reviewer.md           # Example: code review skill
│       └── technical-writer.md        # Example: documentation skill
│
├── docs/
│   ├── QUICKSTART.md                  # Get started in 2 minutes
│   ├── CHEATSHEET.md                  # One-page reference card
│   └── ARCHITECTURE.md                # How the multi-agent pipeline works
│
├── examples/
│   ├── README.md                      # Example index
│   ├── debugging-prompt.md            # Worked example: Python debugger
│   ├── content-generation-prompt.md   # Worked example: LinkedIn posts
│   └── strategic-analysis-prompt.md   # Worked example: business risk analysis
│
└── app/
    ├── README.md                      # How to run the interactive app
    └── hive-mind-v2.jsx                 # React app — multi-agent pipeline UI

---

Quick Start

Option 1: Use the Core Engine Directly

Copy the contents of HIVE_SYNTHESIS_ENGINE.md and paste it as a system prompt (API) or into a Claude Project's knowledge base. Then:

Acting as the Hive Synthesis Engine, generate a [system prompt] for:
[describe your goal]. Domain: [domain]. Calibrate complexity appropriately.

Option 2: Use Individual Templates

Grab any template from prompts/templates/ and fill in the bracketed fields for your use case.

Option 3: Run the Interactive App

The app/hive-mind-v2.jsx is a React component that runs a real multi-agent pipeline — 6 independent API calls → synthesis → adversarial critique → refinement. Drop it into any React project or use it as a Claude Artifact.

Option 4: Claude Project Setup (Recommended)

1. Create a new Claude Project
2. Add HIVE_SYNTHESIS_ENGINE.md to the project knowledge
3. Optionally add any strategy files or skill templates you use frequently
4. Every conversation in that project now has the full engine loaded

---

The Six Researchers

Researcher	Specialty	What They Contribute
⛓ Dr. Wei	Chain-of-Thought	Reasoning scaffolding, step-by-step decomposition, self-verification
🎯 Dr. Brown	Few-Shot Design	Example selection, demonstration formatting, pattern induction
⚙️ Dr. Askell	System Prompts	Persona architecture, behavioral constraints, role definition
🏗 Dr. Reynolds	Structure	XML/markdown optimization, tag architecture, output formatting
🛡 Dr. Perez	Adversarial Testing	Failure modes, guardrails, edge cases, robustness
🧬 Dr. Zhou	Meta-Prompting	Self-improvement loops, prompt chaining, iterative refinement

---

Pipeline Architecture

┌─────────────────────────────────────────────────────┐
│                   USER REQUEST                       │
└──────────────────────┬──────────────────────────────┘
                       │
         ┌─────────────┼─────────────┐
         ▼             ▼             ▼
    ┌─────────┐  ┌─────────┐  ┌─────────┐
    │ Dr. Wei │  │Dr. Brown│  │Dr.Askell│   ← Batch 1 (parallel)
    └────┬────┘  └────┬────┘  └────┬────┘
         │            │            │
         ▼            ▼            ▼
    ┌──────────┐ ┌──────────┐ ┌──────────┐
    │Dr.Reynold│ │Dr. Perez │ │ Dr. Zhou │  ← Batch 2 (parallel)
    └────┬─────┘ └────┬─────┘ └────┬─────┘
         │            │            │
         └─────────┐  │  ┌────────┘
                   ▼  ▼  ▼
            ┌──────────────┐
            │  SYNTHESIS   │  ← Merge 6 outputs
            └──────┬───────┘
                   ▼
            ┌──────────────┐
            │   CRITIQUE   │  ← Adversarial stress-test
            └──────┬───────┘
                   ▼
            ┌──────────────┐
            │  REFINEMENT  │  ← Incorporate fixes
            └──────┬───────┘
                   ▼
            ┌──────────────┐
            │ FINAL OUTPUT │
            └──────────────┘
                   │
                   ↻ (optional: iterate phases 3-4)

---

Contributing

This is a living toolkit. If you develop new strategies, templates, or skill files that work well with Claude, open a PR.

---

License

APACHE 2.0 — see LICENSE for details.