Flows and Prompt Engineering work together seamlessly. Test and refine prompts in the Playground, then deploy them in flows for automated, production-scale execution.

The Workflow

1. Prompt Playground
   ↓ Create and test prompts
   ↓ Run A/B experiments
   ↓ Evaluate quality
   ↓ Iterate and optimize
   
2. Save Prompt
   ↓ Save winning variant
   ↓ Document parameters
   ↓ Set model settings
   
3. Use in Flow
   ↓ Add Prompt node
   ↓ Select saved prompt
   ↓ Map variables
   ↓ Execute at scale

Benefits of This Approach

Quality Assurance

• Test prompts with evaluators before automation
• Validate quality metrics meet thresholds
• Catch issues in controlled environment
• Iterate quickly without flow execution costs

Version Control

• Prompts are centrally managed
• Update once, affects all flows using it
• Roll back to previous versions if needed
• Track changes and improvements over time

Consistent Settings

• Model selection stored with prompt
• Temperature and parameters preserved
• System prompts included
• Response schema enforced

Step-by-Step Integration

1. Create and Test Prompt

In the Prompt Playground:

1. Write your prompt template with variables
2. Configure model settings (model, temperature, etc.)
3. Test with sample inputs
4. Run evaluations to measure quality
5. Iterate until satisfactory

2. Save the Prompt

Prompt Name: "Product Description Generator"
Template: "Write a {{tone}} product description for {{productName}}"
Variables: tone, productName
Model: GPT-4
Temperature: 0.7
Max Tokens: 500

3. Add to Flow

In the Flow Playground:

1. Drag Prompt node to canvas
2. Select "Product Description Generator" from dropdown
3. Node automatically creates inputs for tone and productName
4. Map variables from upstream nodes

4. Map Variables

Dataset Source (products)
  out.name → Prompt.productName
  out.targetAudience → Prompt.tone
  
Prompt Node executes with mapped values
  → out.content contains generated description

Advanced Patterns

Conditional Prompt Selection

Use different prompts based on data characteristics:

Dataset Source
  → Branch by product category
    → Branch A: Prompt "Technical Description"
    → Branch B: Prompt "Marketing Copy"
  → Merge results

Multi-Stage Processing

Chain multiple prompts for complex tasks:

Prompt 1: "Extract Key Facts"
  → output: structured data
Prompt 2: "Generate Summary"
  → uses facts from Prompt 1
Prompt 3: "Create Title"
  → uses summary from Prompt 2

Feedback Loops

Use evaluation results to improve prompts:

Flow Execution
  → Prompt generates outputs
  → Evaluator scores quality
  → Low scores trigger alert
  → Refine prompt in Playground
  → Update saved prompt
  → Re-run flow with improved version

Prompt Variables Best Practices

Clear Variable Names

✅ Good:
{{productName}}, {{targetAudience}}, {{keyFeatures}}

❌ Bad:
{{x}}, {{data}}, {{input1}}

Provide Context

Better: "Write a {{tone}} product description..."
vs
Worse: "Describe {{product}}"

Use Structured Outputs

Configure JSON schemas for reliable parsing:

Schema:
{
  "type": "object",
  "properties": {
    "description": { "type": "string" },
    "keyPoints": { 
      "type": "array",
      "items": { "type": "string" }
    },
    "tone": { 
      "type": "string",
      "enum": ["professional", "casual", "technical"]
    }
  }
}

Monitoring Prompt Performance

In Playground

• Run experiments to compare variants
• Track evaluation scores over time
• Analyze failure patterns
• Measure cost and latency

In Flows

• Monitor execution success rates
• Track API usage and costs
• Review output quality samples
• Identify edge cases for testing

With Tracing

Enable production tracing to monitor deployed flows:

Deployed Flow API
  → Traces every execution
  → Logs inputs and outputs
  → Records latency and costs
  → Flags anomalies
  → Feeds back to Playground for improvement

Common Patterns

Content Generation

Dataset Source (topics)
  → Prompt: "Generate article outline"
  → Array Splitter (split sections)
  → Prompt: "Write section content"
  → Array Flatten (collect sections)
  → Prompt: "Format final article"
  → Dataset Sink

Data Enrichment

Dataset Source (company names)
  → Website Mapper (find pages)
  → Page Scraper (extract info)
  → Prompt: "Summarize company"
  → Dataset Sink (enriched data)

Quality Control

Dataset Source
  → Prompt: "Generate content"
  → Prompt: "Review and score content"
  → Filter: Keep only high scores
  → Dataset Sink (approved content)

Troubleshooting

Prompt Not Found

If a saved prompt is missing from the dropdown:

• Verify prompt is saved in the same workspace
• Check prompt hasn't been deleted
• Refresh the flow page

Variable Mapping Issues

If variables aren't mapping correctly:

• Check variable names match exactly
• Verify data types are compatible
• Test with simple pass-through first
• Inspect intermediate outputs

Unexpected Outputs

If flow outputs differ from Playground results:

• Verify same model and settings
• Check input data format
• Review variable mapping
• Test prompt separately with flow inputs