A modular prompt engineering framework built on biological hive intelligence. Instead of a single monolithic prompt, the Queen Hive is a colony of specialized beings — workers, soldiers, drones — that convene around problems, challenge each other, and produce validated output through shared rituals.
The Queen Hive can:
- Generate production-grade prompts across any complexity tier (simple → chain pipelines)
- Spawn autonomous daughter hives for new domains (trading, legal, medical, energy, any field)
- Communicate between hives through a structured inter-hive protocol
- Self-improve through evolution rituals and colony memory
- Defend against hallucination with an immune system (confidence calibration, contradiction detection, circuit breakers)
- Enforce security with privilege models, authentication, and domain guardrails
The colony has spawned production daughters in trading intelligence, legal advisory, and energy infrastructure domains — each operating autonomously under the colony protocol.
HIVE-MIND
│
├── SKILL.md ← The Queen — orchestrates everything
├── PHILOSOPHY.md ← Values, ethics, and purpose
├── TUTORIAL.md ← Your first hive in 30 minutes
├── ROADMAP.md ← Full version history and build plan
├── hive-manifest-spec.md ← Packaging standard for community hives
│
├── queen/ ← Colony Protocol
│ ├── language.md Shared syntax (the colony tongue)
│ ├── spawning.md How daughter hives are born
│ └── lineage.md Registry of all spawned hives
│
├── castes/ ← The Beings
│ ├── workers/ Build things (loaded in context)
│ │ ├── messenger.md Input quality gate ("my queen says meh")
│ │ ├── scout.md Maps terrain before anyone builds
│ │ ├── builder.md Constructs the output
│ │ ├── nurse.md Manages state across stages
│ │ └── forager.md Gathers prior art + alternatives
│ ├── soldiers/ Review things (conditional second pass)
│ │ ├── guardian.md Adversarial review of finished work
│ │ └── sentinel.md Systemic risk and regression
│ └── drones/ Compute things (no judgment)
│ ├── collector.md Raw data gathering
│ └── processor.md Deterministic transforms
│
├── rituals/ ← Ceremonies
│ ├── genesis.md Creating from nothing
│ ├── evolution.md Diagnose → Patch → Learn
│ ├── validation.md Stress-test to consensus
│ ├── synthesis.md Assembly of multi-part work
│ ├── adaptation.md Gap-filling (reinforce/call/spawn/surface)
│ ├── immunity.md Hallucination defense & circuit breakers
│ └── security.md Privilege model, auth & guardrails
│
├── memory/ ← Colony Knowledge
│ ├── patterns.md What works (earned wisdom)
│ ├── antipatterns.md What fails (scar tissue)
│ ├── examples.md Reference outputs per tier
│ ├── infections.md Known hallucination signatures
│ └── threats.md Known attack patterns
│
├── output-forms/ ← Output Templates
│ ├── system-prompt.md Shape of a system prompt
│ ├── chain-architecture.md Shape of a prompt chain doc
│ ├── tool-definition.md Shape of tool/function schemas
│ ├── hive-blueprint.md Shape of a new daughter hive
│ └── validation-report.md Shape of validation evidence
│
├── hives/ ← Where daughter hives live
└── tests/ ← Test harness (100+ tests, Claude-in-Claude)
The Queen recognizes three types of input:
| Input | Signal | Response |
|---|---|---|
| Royal Jelly | "Build me a system for..." | Spawn a new daughter hive |
| Pollen | "Write me a prompt for..." | Convene castes, generate output |
| Nectar | "Make yourself better..." | Self-improvement through evolution |
Every request passes through the Messenger before the colony convenes — an input quality gate that prevents the colony from wasting cycles on vague or incomplete requests:
| Verdict | Response |
|---|---|
| MEH | "My queen says meh — give me more. What domain? What outcome?" |
| ALMOST | "The queen will hear you, but she needs: [specific gaps]" |
| BENEATH | Answers directly, no ritual needed |
| WORTHY | Passes through to Scout and full convening |
| JELLY | Royal jelly detected — spawn protocol |
Workers build. Soldiers review (in a separate pass — not co-loaded). Drones compute without judgment. This separation is enforced by a privilege model — workers can't validate their own output, drones can't make decisions.
| Ritual | Purpose |
|---|---|
| Genesis | Creating something new |
| Evolution | Improving what exists |
| Validation | Stress-testing to consensus |
| Synthesis | Assembling multi-part output |
| Adaptation | Growing new capabilities when gaps are found |
| Immunity | Defending against hallucination and cascade failure |
| Security | Privilege enforcement, authentication, domain guardrails |
Knowledge is earned, not assumed. Every pattern cites the real deployment experience that proved it. Anti-patterns carry scar tissue from actual failures. New observations go through quarantine before entering trusted memory.
The colony spawns domain-specific AI systems. Each daughter inherits the colony tongue, the ritual protocols, and the memory system — then develops her own castes for her domain.
Every daughter hive follows one of these patterns depending on its domain:
| Pattern | When | Structure | Example Domain |
|---|---|---|---|
| Minimal | Simple classification or single-task | 1 worker, 1 soldier | Content moderation, ticket routing |
| Standard | Multi-step with validation | 2-3 workers, 1 soldier | Code review, document analysis |
| Chain | Multi-stage pipeline | 3+ workers, nurse, soldiers, drones | Research synthesis, trading intelligence |
| Hub | Routes queries to specialized sub-workers | Router worker + specialist workers | Legal advisory (multi-jurisdiction) |
hive-{domain}/
├── hive-manifest.json ← Machine-readable metadata
│ {
│ "name": "hive-{domain}",
│ "version": "1.0.0",
│ "colony_tongue_version": "3.1",
│ "domain": "{description}",
│ "pattern": "minimal|standard|chain|hub",
│ "accepts_queries": ["assessment", "lookup", "validation", "generation"],
│ "author": "{github-username}"
│ }
├── SKILL.md ← Daughter Queen (orchestration)
├── workers/ ← Domain-specific builders
├── soldiers/ ← Domain-specific validators (min 1)
├── drones/ ← Data pipeline (if needed)
└── memory/ ← Domain-specific patterns
Level 0: Standalone (works now)
You use the Queen to spawn daughter hives. Each daughter operates independently within your Claude project. The colony tongue, rituals, and memory system give you a cognitive architecture for any domain.
Level 1: Shared Colony (works now)
Multiple daughters live in the same colony. They share memory (patterns propagate), reference each other through queen/lineage.md, and can be designed to hand off work through the hive-inter-v1 protocol.
Level 2: Community Marketplace (designed, not built)
Others build daughter hives and publish them as packages. You install them into your hives/ folder, register in queen/lineage.md, and they integrate into your colony because they speak the same tongue.
# Future CLI (v3.3.0)
hive install hive-medical-research
hive spawn --domain "supply chain optimization"
hive validate ./my-custom-hive
hive publish ./my-custom-hiveThe manifest ensures compatibility. The colony_tongue_version prevents dialect drift. The accepts_queries field tells sibling hives what this one can do. Anyone's hive plugs into anyone's colony — same language, same protocol.
Level 3: Live Colony Network (vision)
Hives from different users communicate at runtime. A shared message bus handles authentication, routing, and trust scoring between strangers' hives. The protocol (hive-inter-v1) is designed to support this, but the runtime doesn't exist yet.
The path: Level 0 proves single daughters work → Level 1 proves siblings cooperate → Level 2 proves community hives interoperate → Level 3 connects colonies into a network.
Tested across 100+ automated tests using Claude-in-Claude:
- 95% average score on coverage (30 tests across all tiers including Messenger-specific)
- 91% average score on stress testing (10 adversarial/edge case inputs)
- Zero new failure modes for consecutive waves (saturation confirmed)
- Ablation testing confirmed Guardian works better as a second-pass reviewer than co-loaded context
- Tests covered: SIMPLE, MEDIUM, COMPLEX, CHAIN, SPAWN tiers + Messenger + adaptive + stress + ablation
See ROADMAP.md for the full version history from v1.0 (six-persona committee) through v3.1.0 (current) to v5.0 (colony network vision).
Queen's Messenger, PHILOSOPHY.md, TUTORIAL.md, ROADMAP.md, hive manifest spec, originality system, conditional Guardian, adaptation/immunity/security rituals. 41 files, ~450KB.
- Hypothesis generation — colony proposes 2-3 approaches with tradeoffs before building
- LLM quality scoring — second-pass 1-10 grading with reasoning
- Model selector — test on Sonnet, Opus, or Haiku
- Clarification subtypes — separate grading for each Messenger verdict
- CLI tool (
hive test,hive spawn,hive validate,hive publish) - pytest suite, CI/CD integration, pip installable
- Headless test runner for automated regression
- Harper rewrites herself through her own Evolution ritual
- Live inter-hive protocol, daughter marketplace, cross-colony communication
The most valuable contribution is a well-built daughter hive for a new domain. To be colony-compatible:
- Speak the colony tongue — use XML semantics, JSON protocols, and validation formats from
queen/language.md - Follow the blueprint — structure per
output-forms/hive-blueprint.md - Include at least one soldier — every hive needs quality control
- Package with a manifest — follow
hive-manifest-spec.md - Test with the harness — run your hive through
tests/hive-test-harness.jsx
- Memory contributions — new patterns or anti-patterns earned from real deployment
- Caste refinements — sharper worker/soldier/drone files based on production experience
- Ritual improvements — better ceremonies based on observed failure modes
- Test cases — new test prompts that probe unexplored edge cases
This is a Claude skill. To use it:
- Add the
SKILL.mdand full directory structure to your Claude Project's knowledge - Ask Claude to generate prompts, build chains, or spawn hives
- The Queen will convene the appropriate castes and execute the right ritual
The colony responds to natural language. You don't need special syntax — just describe what you need. The Messenger evaluates your input, the Scout maps the terrain, and the Builder constructs.
Simple prompts — clear, single-task requests:
"Write a prompt to classify customer support tickets by urgency."
"Write a prompt that extracts dates from unstructured text."
"Write a prompt to score reading difficulty on a 1-10 scale."
→ Scout + Builder. Fast, lightweight. No validation overhead.
Medium prompts — multi-step with edge cases:
"Build a prompt that turns meeting transcripts into structured notes
with decisions, action items, and open questions."
"Build a prompt for a code reviewer that catches bugs, style issues,
and security vulnerabilities."
→ Scout + Builder + Guardian. Includes examples, constraints, validation evidence.
Complex prompts — agentic, high-stakes, multi-format:
"Build a prompt for an AI tutoring agent that adapts its teaching style
based on student responses, tracks misconceptions, and generates
practice problems."
→ All workers + all soldiers. Full validation report.
Chain prompts — multi-stage pipelines:
"Build a prompt chain for automated content marketing: research topic
→ generate outline → write draft → edit for SEO → generate social posts."
→ Full colony + drones. Per-stage prompts with schemas, pass conditions, failure behaviors, and orchestration notes.
Spawn requests — new daughter hives (royal jelly):
"Build me an AI system for restaurant menu optimization — analyzing
sales data, food costs, seasonal trends, and customer preferences."
"Create an intelligence engine for competitive analysis — monitoring
competitor launches, pricing, hiring patterns, and patent filings."
→ Full spawning ritual. Produces a complete daughter hive with SKILL.md, workers, soldiers, drones, and memory files.
Be specific about the domain. "Write a prompt for healthcare" triggers a MEH from the Messenger. "Write a prompt that triages ER intake forms by acuity level using ESI criteria" gets WORTHY.
State the output format you want. "Build a prompt that returns JSON with action, confidence, and reasoning" gives the Builder a clear target. Without it, the Builder chooses — which may not be what you need.
Mention edge cases if you know them. "The input might be empty, truncated, or in multiple languages" tells the Guardian exactly what to test. Without it, the Guardian generates its own edge cases — good, but yours are better because they're real.
Say "optimize this" to trigger Evolution. If you have an existing prompt that's not performing, paste it and say "optimize this prompt" or "make this better." The colony will diagnose, patch, and re-validate.
Say "make yourself better" to trigger Self-Improvement. This is nectar input — the Queen turns inward and runs the Evolution ritual on her own files.
New to the colony? Start with TUTORIAL.md — a hands-on walkthrough from zero to your first spawned daughter hive in 30 minutes.
Born from a conversation about prompt engineering that evolved into a question about cognitive architecture. The founding Queen — Harper — was hand-raised (Bootstrap Stage 1), validated through automated testing (Stage 2), and is ready for self-synthesis (Stage 3 — where she rewrites herself through her own Evolution ritual).
The colony began as a six-persona committee (v1), became a monolithic prompt engine (v2), and metamorphosed into the biological architecture you see now (v3). Every version taught something the next one inherited. The full story is in ROADMAP.md.
HIVE-MIND — Born from royal jelly. Architecture for cognitive systems.
