Lesson 2: Custom Prompts

Building on your success from Lesson 1, you'll now learn the art and science of prompt engineering to create more intelligent, context-aware AI workflows. We'll enhance your Personal Goal Tracker with sophisticated prompting techniques.

Learning Objectives

By the end of this lesson, you will:

• ✅ Master prompt engineering fundamentals
• ✅ Use dynamic prompt templates with variables
• ✅ Configure AI behavior through context parameters
• ✅ Implement domain-specific AI assistants
• ✅ Optimize AI responses for different use cases

Overview: What We're Building

We'll enhance the Personal Goal Tracker with:

1. Smart Goal Detection: AI identifies goal categories automatically
2. Specialized Coaches: Different AI personalities for different goal types
3. Dynamic Prompting: Prompts that adapt based on goal content
4. Advanced Configuration: Fine-tuned AI parameters for optimal results

Step 1: Download the Enhanced Workflow

Let's get the upgraded workflow file that demonstrates advanced prompting techniques.

Step 2: Understanding Advanced Prompting

Prompt Engineering Fundamentals

What Makes a Great Prompt?

1. Clear Role Definition: "You are a fitness coach..."
2. Specific Instructions: Exact format and content requirements
3. Context Awareness: Using input variables like {goal}
4. Output Structure: Defined sections and formatting
5. Personality: Tone and style that matches the use case

Dynamic Prompt Templates

Notice how our prompts use variables to create dynamic content:

Your Goal: {goal}
Category: {goal_category}

This allows one prompt template to work with any user input while maintaining personalization.

Temperature Settings Explained

Temperature	Use Case	Example
0.1	Classification, precise answers	Goal categorization
0.3	Professional advice, structured responses	Career coaching
0.4	Creative but focused content	Health coaching
0.5	Balanced creativity and structure	General life coaching
0.7+	Creative writing, brainstorming	(Not used in this lesson)

Step 3: Advanced Agent Configuration

Goal Classification Agent

SmartGoals,ClassifyGoal,AI classifies the goal category,llm,RouteToCoach,ErrorHandler,goal,goal_category,"You are a goal classification expert...","{""provider"": ""anthropic"", ""model"": ""claude-3-5-sonnet-20241022"", ""temperature"": 0.1}"

Why Temperature 0.1?

• Classification needs consistency
• We want the same goal to always get the same category
• Low temperature reduces randomness

Specialized Coach Agents

Each coach has a unique personality and approach:

Health Coach:

• Emoji: 🏃‍♀️ (energetic fitness vibe)
• Temperature: 0.4 (creative but structured)
• Focus: Evidence-based fitness strategies

Career Coach:

• Emoji: 💼 (professional)
• Temperature: 0.3 (structured professional advice)
• Focus: Strategic career advancement

Learning Coach:

• Emoji: 📚 (educational)
• Temperature: 0.3 (methodical learning approaches)
• Focus: Learning optimization techniques

General Coach:

• Emoji: 🌟 (inspirational)
• Temperature: 0.5 (balanced and holistic)
• Focus: Life integration and systems thinking

Step 4: The Orchestrator Agent

Intelligent Routing

SmartGoals,RouteToCoach,Route to specialized coach based on category,orchestrator,HealthCoach,ErrorHandler,"goal,goal_category",coach_selection,,"{""nodes"": ""HealthCoach|CareerCoach|LearningCoach|GeneralCoach""}"

The OrchestratorAgent automatically routes to the appropriate coach based on the goal category:

• HEALTH → HealthCoach
• CAREER → CareerCoach
• LEARNING → LearningCoach
• All others → GeneralCoach

Step 5: Running the Enhanced Workflow

Setup and Execution

# Make sure you have your API key set
export ANTHROPIC_API_KEY="your-api-key-here"

# Save lesson2.csv to your project directory
# Run the enhanced workflow
agentmap run lesson2.csv

Example Interaction

What goal would you like to achieve? Please describe it in detail with any relevant context about your situation:
> I want to lose 20 pounds in 6 months and build a sustainable exercise routine because I've been feeling low energy lately

[AI classifies as HEALTH goal...]
[Routes to HealthCoach...]
[Generates detailed fitness plan...]

🎉 Goal Analysis Complete!

Your HEALTH goal has been analyzed by our specialized coach and saved to your personal database.

✅ What's Been Created:
- Detailed action plan tailored to your goal type
- Specific strategies and techniques  
- Progress tracking framework
- Obstacle anticipation and solutions

📁 Find Your Analysis: data/smart_goals_analysis.csv

Step 6: Examining the Enhanced Output

Check your detailed analysis:

cat data/smart_goals_analysis.csv

You'll see structured output with:

• Goal classification
• Specialized coaching advice
• Action plans with timelines
• Specific strategies and techniques
• Progress tracking methods

Step 7: Prompt Engineering Exercises

Exercise 1: Create a Finance Coach

Add a specialized finance coach by modifying the workflow:

SmartGoals,FinanceCoach,Specialized financial planning coaching,llm,SaveAnalysis,ErrorHandler,"goal,goal_category",detailed_analysis,"💰 **FINANCIAL PLANNING ADVISOR**

I'm your personal finance strategist! Let's turn your financial goal into a concrete action plan.

**Your Goal**: {goal}

## 💡 Financial Strategy Analysis
**Goal Type**: [Investment, saving, debt reduction, etc.]
**Risk Assessment**: [Conservative, moderate, aggressive approach]

## 📊 Financial Action Plan
**Month 1**: [Immediate financial steps]
**Months 2-3**: [Building momentum]
**Months 4-6**: [Milestone achievements]
**Long-term**: [Sustained financial health]

## 🎯 Specific Recommendations
- [Concrete financial actions]
- [Tools and apps to use]
- [Tracking and budgeting methods]

## ⚠️ Risk Management
**Potential Pitfalls**: [Common financial mistakes]
**Protection Strategies**: [How to avoid setbacks]

## 📈 Success Metrics
[Specific financial KPIs to track]

Remember: Building wealth is about consistent, smart decisions over time! 💪","{""provider"": ""anthropic"", ""model"": ""claude-3-5-sonnet-20241022"", ""temperature"": 0.3}"

Exercise 2: Modify Classification Logic

Update the ClassifyGoal prompt to include FINANCIAL as a category:

You are a goal classification expert. Analyze this goal and classify it into ONE category: HEALTH (fitness nutrition mental-health), CAREER (professional job business), LEARNING (education skills knowledge), FINANCIAL (money investment savings budgeting), PERSONAL (relationships hobbies lifestyle), or OTHER.

Exercise 3: Experiment with Temperature

Try different temperature settings for the coaches:

More Creative Health Coach (temperature 0.6):

{"provider": "anthropic", "model": "claude-3-5-sonnet-20241022", "temperature": 0.6}

More Precise Career Coach (temperature 0.2):

{"provider": "anthropic", "model": "claude-3-5-sonnet-20241022", "temperature": 0.2}

Step 8: Advanced Prompting Techniques

1. Structured Output Formatting

Use markdown and emojis to create visually appealing responses:

## 🎯 Goal Analysis
**Why this matters**: [explanation]

## 📋 Action Plan
**Week 1**: [specific tasks]

2. Role-Based Prompting

Define clear expertise areas:

You are a certified fitness trainer with 10 years of experience specializing in sustainable weight loss...

3. Context-Aware Prompting

Use multiple input variables:

Your Goal: {goal}
Category: {goal_category}
Current Situation: {context}

4. Output Constraints

Specify format requirements:

Respond with exactly 3 bullet points, each starting with an action verb...

Key Concepts Learned

1. Prompt Engineering Principles

• Clear role definition creates better AI responses
• Structured output formatting improves readability
• Variable injection enables dynamic personalization
• Temperature controls creativity vs. consistency

2. Specialized AI Assistants

• Different domains need different approaches
• Personality and tone matter for user experience
• Context-specific knowledge improves relevance
• Template reusability across similar use cases

3. Intelligent Workflow Routing

• OrchestratorAgent enables smart routing decisions
• Classification enables specialized processing
• Multi-path workflows handle diverse inputs
• Error handling maintains workflow robustness

4. Configuration Optimization

• Temperature settings affect response style
• Model selection impacts capability and cost
• Provider choice depends on specific needs
• Context parameters fine-tune behavior

Troubleshooting

Issue: Classification is inconsistent

Solution: Lower the temperature for the classification agent:

{"temperature": 0.1}

Issue: Coach responses are too generic

Solution: Add more specific role definition and examples in prompts.

Issue: Routing doesn't work correctly

Solution: Check that node names in orchestrator context match actual node names.

Best Practices Learned

1. Prompt Design

• Start with role definition
• Provide specific instructions
• Use structured output formatting
• Include examples when helpful

2. Temperature Selection

• Use low temperature (0.1-0.2) for classification
• Use medium temperature (0.3-0.4) for structured advice
• Use higher temperature (0.5+) for creative content

3. Workflow Design

• Separate classification from processing
• Use specialized agents for different domains
• Implement error handling at each step
• Provide clear user feedback

Next Steps

Outstanding work! You've mastered:

• ✅ Advanced prompt engineering techniques
• ✅ Dynamic template creation with variables
• ✅ AI behavior configuration through context
• ✅ Intelligent workflow routing with orchestration
• ✅ Domain-specific AI assistant creation

In Lesson 3, we'll take the next major step: creating your own custom agent by extending the BaseAgent class and implementing custom business logic.

Continue to Lesson 3: Custom Agent Development →

---

Additional Resources

• Prompt Engineering Guide
• LLMAgent Configuration
• OrchestratorAgent Reference
• AI Provider Setup

---

💡 Pro Tip: Save examples of effective prompts you create - they become valuable templates for future workflows! Also, experiment with different personalities and tones to see how they affect user engagement.