Prompts old

Prompt Studio

Design, test, and optimize prompts across multiple models.

---

Overview

Prompt Studio transforms prompt engineering from guesswork into a systematic process. Test prompts across different models, compare results side-by-side, and identify the optimal configuration for your use case.

Design Prompt → Add Variables → Select Models → Run Tests → Compare Results → Deploy

---

Key Features

• Multi-model comparison: Test up to 5 models simultaneously
• Variable support: Batch process prompts with different inputs
• 65+ templates: Pre-built prompts for common use cases
• Version control: Track prompt iterations
• JSON schemas: Structure model outputs
• Dataset import: Load test data from CSV files

---

Workflow

1. Create a Prompt

Start with one of three approaches:

Generate: Describe what you want and let AI expand it into a detailed prompt.

From scratch: Write your prompt manually with full control.

From library: Use a pre-built template as a starting point.

Prompt Structure

# System Prompt (optional)
You are a helpful customer service agent for an e-commerce company.
Be concise, professional, and friendly.

# Human Prompt
Customer inquiry: {{inquiry}}
Order history: {{order_history}}

Please provide a helpful response addressing their concern.

---

2. Define Variables

Variables enable testing with different inputs.

Syntax: {{variable_name}}

Example:

Summarize this product review:

Product: {{product_name}}
Review: {{review_text}}
Rating: {{rating}}/5

Variable sources:

• Manual entry
• CSV import
• AI-generated test data

---

3. Select Models

Compare performance across models:

Provider	Available Models
OpenAI	GPT-4, GPT-4o, GPT-3.5-turbo
Anthropic	Claude 3 Opus, Sonnet, Haiku
Google	Gemini 1.5 Pro, Flash
Azure	GPT-4, GPT-3.5-turbo
Custom	Fine-tuned models

Model Parameters

Adjust settings per model:

temperature: 0.7      # Creativity (0-1)
top_p: 0.9           # Nucleus sampling
top_k: 50            # Token selection
max_tokens: 1000     # Output length

---

4. Run Tests

Click Run to execute prompts across selected models.

Batch processing: Run up to 10 data rows simultaneously.

Results include:

• Generated response
• Tokens sent/received
• Response time
• Cost estimate

---

5. Compare Results

Evaluate models across dimensions:

Criteria	Description
Accuracy	Factual correctness
Relevance	Addresses the prompt
Tone	Appropriate style
Completeness	Thorough coverage
Latency	Response speed
Cost	Token economics

---

JSON Schema Output

Structure model responses with schemas:

Definition

{
  "type": "object",
  "properties": {
    "sentiment": {
      "type": "string",
      "enum": ["positive", "negative", "neutral"]
    },
    "confidence": {
      "type": "number",
      "minimum": 0,
      "maximum": 1
    },
    "key_topics": {
      "type": "array",
      "items": { "type": "string" }
    }
  },
  "required": ["sentiment", "confidence"]
}

Result

{
  "sentiment": "positive",
  "confidence": 0.92,
  "key_topics": ["product quality", "fast shipping", "customer service"]
}

---

Test Data

Import from CSV

Upload a CSV with columns matching your variables:

product_name,review_text,rating
"Wireless Headphones","Great sound quality!",5
"Phone Case","Broke after a week",2
"Laptop Stand","Sturdy and adjustable",4

Generate Synthetic Data

Let AI create test data:

1. Define your prompt with variables
2. Click Generate test data
3. AI analyzes context and generates appropriate values
4. Review and edit as needed

---

Prompt Library

Access 65+ pre-built templates:

Categories

• Content generation: Blog posts, emails, marketing copy
• Summarization: Documents, meetings, articles
• Analysis: Sentiment, classification, extraction
• Code: Generation, explanation, debugging
• Customer service: Responses, FAQs, escalation
• Data: Formatting, transformation, validation

Using Templates

1. Browse the library by category
2. Preview template content
3. Click Use template
4. Customize for your use case

---

Version Control

Track prompt evolution:

v1.0 - Initial prompt
v1.1 - Added tone guidelines
v1.2 - Improved error handling
v2.0 - Restructured for better accuracy

Best Practices

• Document changes in version notes
• Test before promoting to production
• Keep previous versions for rollback

---

Integration

Export to Tools

Deploy optimized prompts to:

• Agent instructions
• AI nodes in workflow tools
• Code tool prompts

API Access

Use prompts programmatically:

const response = await client.prompts.execute({
  promptId: "prompt_123",
  variables: {
    inquiry: "Where is my order?",
    order_history: orderData
  },
  model: "gpt-4o"
});

---

Best Practices

Be Specific

# Vague
Summarize this text.

# Specific
Summarize this customer support ticket in 2-3 sentences.
Focus on: the issue, attempted solutions, and current status.
Format as bullet points.

Provide Examples

Classify the sentiment of this review.

Examples:
- "Love this product!" → positive
- "Worst purchase ever" → negative
- "It's okay, nothing special" → neutral

Review: {{review_text}}
Sentiment:

Use System Prompts

Set consistent behavior:

# System
You are a technical documentation writer.
- Use clear, concise language
- Include code examples where helpful
- Avoid jargon unless necessary

Test Edge Cases

Include challenging inputs:

• Empty values
• Very long text
• Special characters
• Ambiguous requests

---

Related

• Evaluation Studio
• Agents
• AI Nodes
• Models