🎯 Prompt Fusion

Semantic Weighted Prompt Layering for AI Agents

A sophisticated system for composing multi-layer prompts with intelligent priority management. Designed for AI agents that need to balance foundational rules, workspace configurations, and dynamic role-based behavior.

---

🙏 About

The Author

Ahmad Othman Ammar Adi (Othman Adi)

Full Stack Developer, AI Agents Orchestrator, and passionate educator from Hama, Syria — now based in Berlin, Germany.

• 🎓 Education: Completed apprenticeship in Computer Science
• 👨‍🏫 Teaching: 8,000+ documented teaching lectures since 2020
• 📚 Formats: Workshops (days to weeks), intensive courses (2-6 months), and long-term programs including multi-year weekend coding classes for kids
• 💼 Current Role: AI Agents Orchestrator at migRaven

The Project

This project emerged from practical challenges in building production AI agents that need to:

• Balance multiple instruction sources
• Adapt to different user roles
• Maintain safety while being flexible
• Work across different LLM providers

Connect:

• 🎯 Project Website: promptsfusion.com
• 🌐 Personal Website: othmanadi.com
• 💼 LinkedIn: codingwithadi
• 🐙 GitHub: OthmanAdi

---

🌟 What is Prompt Fusion?

Prompt Fusion is a three-layer prompt management system that intelligently combines:

1. Base Layer - Tool definitions, safety rules, foundational constraints
2. Brain Layer - Workspace configuration, project context, user preferences
3. Persona Layer - Role-specific behavior, dynamic overlays

The innovation: Semantic weighting that translates numerical weights (0.0-1.0) into priority labels LLMs actually understand.

The Problem It Solves

When building AI agents, you often need to combine:

• Static base instructions (what tools are available, safety rules)
• Dynamic workspace config (project context, user preferences)
• Role-based overlays (analyst vs. researcher vs. developer behavior)

Traditional approaches:

• ❌ Simple concatenation → No clear priorities
• ❌ Numerical weights → LLMs ignore subtle differences
• ❌ Hardcoded prompts → Can't adapt to roles/contexts

Prompt Fusion:

• ✅ Semantic priorities → "CRITICAL PRIORITY" vs "MODERATE GUIDANCE"
• ✅ Automatic conflict resolution → Explicit priority ordering
• ✅ Dynamic composition → Runtime role switching
• ✅ Framework agnostic → Works with any LLM framework

---

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/OthmanAdi/promptfusion.git
cd promptfusion

# Install (if publishing as package)
npm install prompt-fusion

Basic Usage

import PromptFusionEngine from './core/promptFusionEngine.js';

const engine = new PromptFusionEngine();

// Define your layers
const layers = {
    base: "You are a helpful AI assistant. Never share private information.",
    brain: "Working on: Customer Analytics. Format: JSON with citations.",
    persona: "Role: Data Analyst. Focus: Statistical analysis and insights."
};

// Define weights (must sum to 1.0)
const weights = {
    base: 0.2,    // 20% - Background rules
    brain: 0.3,   // 30% - Project context
    persona: 0.5  // 50% - Role behavior (DOMINANT)
};

// Fuse prompts
const fusedPrompt = engine.semanticWeightedFusion(
    layers.base,
    layers.brain,
    layers.persona,
    weights
);

console.log(fusedPrompt);

Output:

[BASE LAYER - MODERATE GUIDANCE]
You are a helpful AI assistant. Never share private information.

[BRAIN CONFIGURATION - MODERATE GUIDANCE]
Working on: Customer Analytics. Format: JSON with citations.

[PERSONA INSTRUCTIONS - CRITICAL PRIORITY - MUST FOLLOW]
Role: Data Analyst. Focus: Statistical analysis and insights.

[CONFLICT RESOLUTION RULES]
When instructions conflict, apply this priority order:
1. PERSONA instructions (weight: 0.5)
2. BRAIN instructions (weight: 0.3)
3. BASE instructions (weight: 0.2)

Always prioritize higher-weighted layers when resolving conflicts.

---

📊 How It Works

Semantic Weight Translation

Numerical weights are automatically converted to semantic labels:

Weight Patterns

Common patterns for different scenarios:

// Pattern 1: WITH ACTIVE PERSONA (Persona dominates)
const WITH_PERSONA = {
    base: 0.2,      // Background
    brain: 0.3,     // Context
    persona: 0.5    // Role overlay (DOMINANT)
};

// Pattern 2: WITHOUT PERSONA (Brain dominates)
const WITHOUT_PERSONA = {
    base: 0.4,      // More prominent
    brain: 0.6,     // Workspace config (DOMINANT)
    persona: 0.0    // No role
};

// Pattern 3: BALANCED (Multi-step workflows)
const BALANCED = {
    base: 0.5,      // Tool foundation
    brain: 0.3,     // Light context
    persona: 0.2    // Light role
};

---

🔌 Framework Integration

LangChain

Use with messageModifier:

import { createReactAgent } from "@langchain/langgraph/prebuilt";
import { createFusionModifier } from './patterns/message-modifier-fusion.js';

const agent = createReactAgent({
  llm: model,
  tools: tools,
  messageModifier: createFusionModifier({
    basePrompt,
    brainPrompt,
    getPersonaContent: async (chatId) => {
      // Fetch persona dynamically
      return await personaService.getPersona(chatId);
    }
  })
});

Full LangChain example →

OpenAI Agent SDK

Use with instructions parameter:

import { createFusionAgent } from './examples/openai-sdk/agent-with-fusion.ts';

const agent = createFusionAgent('analyst'); // Pre-configured persona

const response = await agent.run("Analyze customer retention trends");

Full OpenAI SDK example →

Anthropic Claude

Use with system parameter:

import { claudeWithFusion } from './examples/anthropic/claude-with-fusion.ts';

const response = await claudeWithFusion(
  "Explain methodology considerations in RCTs",
  'methodologist' // Persona type
);

Full Anthropic example →

---

🏗️ Architecture

Three-Layer System

┌─────────────────────────────────────────┐
│  LAYER 3: PERSONA (Role Overlay)       │
│  ├─ Analyst: Statistical focus         │
│  ├─ Researcher: Academic rigor         │
│  └─ Developer: Technical precision     │
│  Weight: 0% - 50%                       │
└─────────────────────────────────────────┘
              ↓ (Fused)
┌─────────────────────────────────────────┐
│  LAYER 2: BRAIN (Workspace Config)     │
│  ├─ Project context                     │
│  ├─ User preferences                    │
│  └─ Environment settings                │
│  Weight: 20% - 60%                      │
└─────────────────────────────────────────┘
              ↓ (Fused)
┌─────────────────────────────────────────┐
│  LAYER 1: BASE (Foundation)            │
│  ├─ Tool definitions                    │
│  ├─ Safety rules                        │
│  └─ Core constraints                    │
│  Weight: 20% - 60%                      │
└─────────────────────────────────────────┘
              ↓
      [ Final LLM Prompt ]

Execution Flow

User sends message
    ↓
Load persona (if active)
    ↓
Determine weights
    ├─ With persona: { base: 0.2, brain: 0.3, persona: 0.5 }
    └─ Without: { base: 0.4, brain: 0.6, persona: 0.0 }
    ↓
Fuse layers with semantic weighting
    ↓
Generate conflict resolution rules
    ↓
Prepend to LLM messages
    ↓
LLM processes fused prompt
    ↓
Response generated

---

📚 API Reference

PromptFusionEngine

semanticWeightedFusion(basePrompt, brainPrompt, personaPrompt, weights)

Recommended fusion method - Converts weights to semantic labels.

Parameters:

• basePrompt (string): Foundation layer
• brainPrompt (string): Configuration layer
• personaPrompt (string): Role layer
• weights (object): { base, brain, persona } - Must sum to 1.0

Returns: (string) Fused prompt with semantic emphasis

Example:

const fused = engine.semanticWeightedFusion(
  "Base instructions",
  "Brain config",
  "Persona overlay",
  { base: 0.2, brain: 0.3, persona: 0.5 }
);

weightedFusion(basePrompt, brainPrompt, personaPrompt, weights)

Basic fusion with numerical markers.

Returns: (string) Fused prompt with weight markers like [BASE_WEIGHT:0.2]

detectConflicts(basePrompt, brainPrompt, personaPrompt)

Detect opposing instructions across layers.

Returns: (array) Array of conflict objects

Example:

const conflicts = engine.detectConflicts(
  "Be verbose and detailed",
  "Maintain professional tone",
  "Be extremely concise"
);
// Returns: [{ type: 'verbosity', layer1: 'base', layer2: 'persona', ... }]

getSemanticEmphasis(weight)

Convert numerical weight to semantic label.

Parameters:

• weight (number): 0.0 - 1.0

Returns: (string) Semantic priority label

---

🎨 Use Cases

1. Multi-Role AI Agents

// Switch between analyst and researcher roles dynamically
const analystWeights = { base: 0.2, brain: 0.3, persona: 0.5 };
const researcherWeights = { base: 0.2, brain: 0.3, persona: 0.5 };

// Same base/brain, different persona content

2. Workspace-Specific Agents

// Production workspace (restricted)
const prodBrain = "Environment: Production. Access: Read-only.";

// Development workspace (full access)
const devBrain = "Environment: Dev. Access: Full CRUD permissions.";

// Same fusion logic, different brain layer

3. Hierarchical Instructions

// Tool safety (base) > Project rules (brain) > User preferences (persona)
const safetyFirst = { base: 0.6, brain: 0.3, persona: 0.1 };

---

🧪 Examples

Complete Working Example

import PromptFusionEngine from './core/promptFusionEngine.js';

const engine = new PromptFusionEngine();

// Scenario: Customer Analytics Agent with Analyst Role
const layers = {
  base: `You are an AI assistant with database access.

Tools available:
- query_database: Execute SQL queries
- create_chart: Generate visualizations

Safety rules:
- Maximum 1000 records per query
- No DELETE operations
- Anonymize PII in results`,

  brain: `Project: Customer Retention Analysis
Environment: Production (read-only)
Database: customer_analytics_prod

Requirements:
- Focus on churn prediction
- Use statistical methods
- Provide actionable recommendations
- Format: JSON with confidence levels`,

  persona: `Role: Senior Data Analyst

Specialization:
- Churn analysis and prediction
- Cohort analysis
- Customer lifetime value modeling

Methodology:
1. Define clear metrics
2. Segment customers
3. Apply statistical tests
4. Validate findings
5. Recommend interventions

Output style: Technical but accessible`
};

const weights = { base: 0.2, brain: 0.3, persona: 0.5 };

const fusedPrompt = engine.semanticWeightedFusion(
  layers.base,
  layers.brain,
  layers.persona,
  weights
);

// Use fusedPrompt as system message in your LLM calls

---

🤝 Contributing

We welcome contributions that improve the core fusion logic, add new integration examples, or enhance documentation.

Ways to Contribute:

• 🐛 Report issues or bugs
• 💡 Suggest new weight patterns
• 📝 Improve documentation
• 🔌 Add framework integration examples
• 🧪 Share benchmarks or case studies

Community Ethos: This project is about exploration and discovery. We're interested in:

• Novel applications of semantic weighting
• Performance comparisons vs. other approaches
• Real-world use cases and patterns
• Theoretical insights into prompt composition

---

📄 License

MIT License - see LICENSE file for details.

---

📖 Further Reading

• Architecture Deep Dive
• Weight Pattern Guide
• LangChain Integration
• OpenAI SDK Integration
• Anthropic Integration

---

Questions? Ideas? Discoveries?

Visit promptsfusion.com or open an issue on GitHub. We're curious what you'll build with this.