Skip to content

README.md

Latest commit

 

History

History
275 lines (207 loc) · 6.59 KB

README.md

File metadata and controls

275 lines (207 loc) · 6.59 KB

Anthropic Claude + Prompt Fusion Integration

This example shows how to integrate Prompt Fusion with Anthropic's Claude API using the system parameter for dynamic prompt composition.

What This Does

Demonstrates:

  • Using Prompt Fusion with Claude's system parameter
  • Multi-turn conversations with persona switching (via ClaudeFusionChat class)
  • Extended thinking mode integration
  • Streaming responses with fused prompts

Prerequisites

  • Node.js 18+
  • Anthropic API key
  • TypeScript knowledge (or use ts-node)

Installation

npm install @anthropic-ai/sdk

Setup

Create a .env file:

ANTHROPIC_API_KEY=your_anthropic_api_key_here

How to Run

# Using ts-node
npx ts-node claude-with-fusion.ts

# Or using tsx
npx tsx claude-with-fusion.ts

The example includes 4 usage patterns - check the main() function at the bottom of the file.

Code Overview

Basic Usage

import Anthropic from '@anthropic-ai/sdk';
import PromptFusionEngine from '../../core/promptFusionEngine.js';

const client = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });
const fusionEngine = new PromptFusionEngine();

// Fuse your prompts
const systemPrompt = fusionEngine.semanticWeightedFusion(
  baseSystemPrompt,
  brainSystemPrompt,
  personaSystemPrompts['methodologist'],
  { base: 0.2, brain: 0.3, persona: 0.5 }
);

// Use with Claude
const response = await client.messages.create({
  model: "claude-sonnet-4-5-20250929",
  max_tokens: 2048,
  system: systemPrompt,  // Fused prompt goes here
  messages: [
    { role: "user", content: "Your question" }
  ]
});

Pattern 1: Single Message

const response = await claudeWithFusion(
  "What are the key methodological considerations in RCTs?",
  'methodologist'  // Persona type
);

Pattern 2: Extended Thinking Mode

const { thinking, response } = await claudeWithThinkingAndFusion(
  "Analyze validity threats in quasi-experimental designs",
  'methodologist'
);

console.log("Thinking:", thinking);
console.log("Response:", response);

Pattern 3: Streaming

await streamingClaudeWithFusion(
  "Explain mixed-methods research",
  'theorist',
  (chunk) => process.stdout.write(chunk)  // Handle each chunk
);

Pattern 4: Multi-Turn with Persona Switching

const chat = new ClaudeFusionChat();

// Start without persona
let msg = await chat.send("What is meta-analysis?");

// Switch to theorist mid-conversation
chat.switchPersona('theorist');
msg = await chat.send("How does it relate to theory building?");

// Switch back
chat.switchPersona(null);

Key Features

Fusion Timing

With Claude, fusion happens before each API call:

function createFusedSystemPrompt(base, brain, persona, hasPersona) {
  const weights = hasPersona
    ? { base: 0.2, brain: 0.3, persona: 0.5 }
    : { base: 0.4, brain: 0.6, persona: 0.0 };

  return fusionEngine.semanticWeightedFusion(base, brain, persona, weights);
}

Multi-Turn Conversation Management

The ClaudeFusionChat class:

  • Maintains message history
  • Re-fuses prompts when persona switches
  • Passes history to Claude for context continuity
class ClaudeFusionChat {
  private messages = [];
  private currentPersona = null;

  switchPersona(personaType) {
    this.currentPersona = personaType;
    // Next send() will use new persona
  }

  async send(userMessage) {
    this.messages.push({ role: "user", content: userMessage });

    const systemPrompt = createFusedSystemPrompt(/*...*/);

    const response = await client.messages.create({
      system: systemPrompt,
      messages: this.messages  // Full history
    });

    this.messages.push({ role: "assistant", content: response });
    return response;
  }
}

Customization

Add Your Own Personas

const personaSystemPrompts = {
  methodologist: `...`,
  theorist: `...`,
  clinician: `SPECIALIZED ROLE: Clinical Researcher

  EXPERTISE:
  - Clinical trial design
  - Patient-centered outcomes
  - Translational research

  FOCUS AREAS:
  1. Clinical relevance assessment
  2. Safety and efficacy evaluation
  3. Real-world application`
};

Adjust Thinking Budget

const response = await client.messages.create({
  model: "claude-sonnet-4-5-20250929",
  max_tokens: 4096,
  thinking: {
    type: "enabled",
    budget_tokens: 5000  // Increase for complex reasoning
  },
  system: fusedPrompt,
  messages: [/*...*/]
});

Change Weight Patterns

// Research-heavy (brain dominates when no persona)
const weights = hasPersona
  ? { base: 0.2, brain: 0.3, persona: 0.5 }
  : { base: 0.3, brain: 0.7, persona: 0.0 };

// Safety-first (base dominates)
const weights = hasPersona
  ? { base: 0.5, brain: 0.3, persona: 0.2 }
  : { base: 0.6, brain: 0.4, persona: 0.0 };

Extended Thinking Integration

Extended thinking works seamlessly with fused prompts:

const response = await client.messages.create({
  model: "claude-sonnet-4-5-20250929",
  thinking: { type: "enabled", budget_tokens: 2048 },
  system: fusedPrompt,  // Your fused prompt
  messages: [{ role: "user", content: "Complex question" }]
});

// Extract thinking and response
const thinkingBlock = response.content.find(b => b.type === 'thinking');
const textBlock = response.content.find(b => b.type === 'text');

The fused prompt provides context for both thinking and response generation.

Troubleshooting

Persona not applying:

  • Verify hasPersona boolean is true when persona content exists
  • Log the systemPrompt to see the actual fused output
  • Check persona content isn't empty string

Message history growing too large:

  • Implement message trimming:
if (this.messages.length > 20) {
  // Keep system context + recent messages
  this.messages = this.messages.slice(-10);
}

Extended thinking not showing:

  • Make sure you're checking response.content array
  • Thinking blocks have type: 'thinking'
  • Not all queries trigger thinking - complex ones do

Comparison with Other Frameworks

Aspect Claude (system param) LangChain (messageModifier)
Fusion timing Per API call Per message (hook)
Multi-turn Manual history management Framework handled
Streaming Native support Native support
Thinking mode Built-in N/A

Learn More