The Secret Weapon: Prompt Engineering: 🪄
Table of Contents
Introduction
Crafting the perfect prompt is like whispering secret instructions to your AI muse. But there’s no one-size-fits-all approach! Exploring different types of prompts unlocks a spectrum of possibilities in your Generative AI creations. Here’s a sneak peek:
Different Types of Prompt Engineering
1. The Classic Prompt: ✍️
- Straightforward and clear, laying out the desired task or output.
- Think: “Write a poem about a robot falling in love with the moon.”
- Ideal for beginners and specific, focused tasks.
2. The N-Shot Prompt:
- Provide a few examples (N) of the desired output to guide the AI.
- Imagine: “Show me 3 examples of poems about nature written in free verse, then write your own.”
- Useful for refining style and direction, fostering consistency.
3. The Chain-of-Thought Prompt:
- Break down complex tasks into smaller, sequential steps.
- Picture: “1. Identify emotions in this image. 2. Generate captions expressing those emotions in different tones. 3. Merge captions into a cohesive paragraph.”
- Perfect for tackling intricate tasks, and promoting logical flow.
4. The Conditioned Prompt:
- Sets specific constraints or parameters to shape the output.
- Think: “Write a children’s story filled with puns, but avoid using clichés.”
- Adds a layer of challenge and control, encouraging creative exploration within boundaries.
5. The Open-Ended Prompt:
- Offers minimal guidance, leaving the AI to fill in the blanks.
- Imagine: “Create a world where animals speak and build cities. Tell me a story about its most daring explorer.”
- Sparks innovation and unexpected results, ideal for brainstorming and discovery.
6. Interactive Prompts: ️
- Leverage user input or external data to dynamically adjust the narrative.
- Picture: “You are a detective in a cyberpunk city. Tell me what you see as I describe the crime scene.”
- Creates an immersive, personalized experience, blurring the lines between creator and creation.
Points to remember while writing prompt
- Use clear and concise language in prompts.
- Provide relevant context and data.
- Adjust temperature for creativity or accuracy.
- Experiment with different prompt formats and lengths.
- Integrate model responses into Salesforce workflows using provided APIs.
Examples of Different Prompts
Classic Insight Prompt
Prompt: Analyze historical sales data of tech clients and identify key trends in purchasing behaviour.
// Query tech client sales data
List<Opportunity> techSales = [
SELECT Id, CloseDate, Amount, Product.Name
FROM Opportunity
WHERE AccountId IN (SELECT Id FROM Account WHERE Industry = 'Technology')
AND IsClosed = true
ORDER BY CloseDate DESC
];
// Calculate quarterly sales trends
Map<Date, Decimal> quarterlySales = new Map<Date, Decimal>();
for (Opportunity opp : techSales) {
Date quarterStart = Date.newInstance(opp.CloseDate.year(), opp.CloseDate.month() - opp.CloseDate.month() % 3, 1);
quarterlySales.put(quarterStart, quarterlySales.get(quarterStart) + opp.Amount);
}
// Identify top products by revenue
Map<String, Decimal> productRevenue = new Map<String, Decimal>();
for (Opportunity opp : techSales) {
productRevenue.put(opp.Product.Name, productRevenue.get(opp.Product.Name) + opp.Amount);
}
// Display insights for developers to interpret
System.debug('Quarterly Sales Trends: ' + quarterlySales);
System.debug('Top Products by Revenue: ' + productRevenue);
Personalized Customer Prompt
Prompt: Recommend three relevant products for Customer X based on recent website visits and past purchases.
// Gather customer data
Contact customerX = [SELECT Id, Name, RecentVisitedPages, PastPurchases FROM Contact WHERE Name = 'Customer X'];
// Collect product recommendations from Einstein (replace with your actual Einstein model)
List<ProductRecommendation> recommendations = EinsteinRecommendations.getRecommendationsForContact(customerX);
// Display top 3 recommendations
System.debug('Recommended Products for Customer X:');
for (Integer i = 0; i < 3; i++) {
System.debug('- ' + recommendations[i].Product.Name);
}
Data-Driven Content Prompt
Prompt: Create a blog post addressing top customer concerns from surveys and support tickets.
// Gather customer feedback data
List<Case> supportTickets = [SELECT Id, Subject, Description FROM Case WHERE Status = 'Closed' AND Type = 'Problem'];
List<SurveyResponse> surveyResponses = [SELECT Id, Question, Answer FROM SurveyResponse];
// Extract common topics and concerns
Set<String> topConcerns = new Set<String>();
for (Case ticket : supportTickets) {
topConcerns.addAll(extractKeywords(ticket.Subject, ticket.Description));
}
for (SurveyResponse response : surveyResponses) {
topConcerns.addAll(extractKeywords(response.Question, response.Answer));
}
// Generate blog content
String blogPost = '**Top Customer Concerns and How We're Addressing Them**nn';
for (String concern : topConcerns) {
blogPost += '* **' + concern + ':**n';
blogPost += ' - Explanation of the concernn';
blogPost += ' - Steps we're taking to address itn';
}
// Display blog post for developers to review and publish
System.debug(blogPost);
Predictive Scenario Prompt
Prompt 1: Predict the likelihood of a lead converting to an opportunity based on historical data and current lead attributes.
public class LeadConversionPredictor {
public PredictionResult predictLeadConversion(Lead lead) {
// 1. Gather historical data for lead conversion analysis
List<Lead> historicalLeads = [SELECT Id, ConvertedDate, ... FROM Lead WHERE ...];
// 2. Train a prediction model (e.g., logistic regression, decision tree)
Model model = trainPredictionModel(historicalLeads);
// 3. Extract relevant lead attributes
Map<String, Object> leadFeatures = extractLeadFeatures(lead);
// 4. Generate prediction using the model
Probability prediction = model.predict(leadFeatures);
// 5. Return prediction result
return new PredictionResult(prediction);
}
// Helper methods for model training and feature extraction
private Model trainPredictionModel(List<Lead> leads) {
// Implement model training logic
}
private Map<String, Object> extractLeadFeatures(Lead lead) {
// Map lead attributes to features for the model
}
}
Prompt 2: Identify customers most likely to churn within the next 6 months, considering past purchase patterns, engagement metrics, and support interactions.
public class ChurnPredictor {
public List<Customer> predictHighChurnRiskCustomers() {
// 1. Gather relevant customer data from multiple objects
List<Customer> customers = [SELECT Id, ... FROM Customer];
List<Order> orders = [SELECT Id, CustomerId, ... FROM Order];
List<Case> cases = [SELECT Id, CustomerId, ... FROM Case];
// 2. Calculate churn risk scores based on defined criteria
Map<Id, Decimal> churnScores = calculateChurnScores(customers, orders, cases);
// 3. Identify customers with scores above a threshold
List<Customer> highChurnRiskCustomers = new List<Customer>();
for (Customer customer : customers) {
if (churnScores.get(customer.Id) > churnThreshold) {
highChurnRiskCustomers.add(customer);
}
}
return highChurnRiskCustomers;
}
// Helper method for churn score calculation
private Map<Id, Decimal> calculateChurnScores(List<Customer> customers, List<Order> orders, List<Case> cases) {
// Implement churn score logic
}
}
N-Shot Prompt
Prompt:
Generate a list of 3 email subject lines that are likely to get high open rates, based on the following examples:
- Example 1: Your free trial is about to expire!
- Example 2: Don’t miss out on this exclusive offer!
- Example 3: 5 secrets to boosting your sales productivity
Response
- “Unlock VIP Access: Special Benefits Await You!”
- “Flash Sale Alert: Limited-time Savings Inside!”
- “Double Your Productivity with These Proven Strategies!”
Prompt: Write an Apex trigger that updates the ‘Last_Reviewed_Date’ field on the Account object to today’s date whenever a new Opportunity is created for that account.
trigger UpdateContactLastContactedDate on Contact (after insert) {
for (Contact c : Trigger.new) {
c.Last_Contacted_Date__c = Date.today();
}
}
Few Short Example
Prompt: Write an Apex method that calculates the total amount of all open opportunities for a given account, using only a few keywords as guidance: account ID, opportunity amount, and open opportunities.
Keywords:
- account ID
- opportunity amount
- open opportunities
public static Decimal calculateTotalOpenOpportunityAmount(Id accountId) {
List<Opportunity> openOpps = [
SELECT Amount
FROM Opportunity
WHERE AccountId = :accountId AND IsClosed = false
];
Decimal totalAmount = 0;
for (Opportunity opp : openOpps) {
totalAmount += opp.Amount;
}
return totalAmount;
}
Resources
Exploring the world of prompt engineering can be exhilarating and empowering. To ignite your journey, here’s a curated list of resources and links to help you learn and master this fascinating skill:
Interactive Platforms:
- Hugging Face’s Prompt Engineering Tutorial: Hands-on exploration with interactive challenges and code examples. https://huggingface.co/learn/nlp-course
- TensorFlow Playground: Experiment with prompts and different AI models in a user-friendly environment. https://playground.tensorflow.org/
- Colab Notebooks: Get creative with pre-built notebooks on various prompt engineering techniques. https://research.google.com/colaboratory/
Comprehensive Guides:
- The Prompt Manual: Deep dive into the theory and practice of prompt engineering, with detailed examples. https://huggingface.co/docs/peft/conceptual_guides/prompting
- AI Dungeon’s Prompt Engineering Guide: Practical tips and strategies for crafting effective prompts. https://help.aidungeon.com/
- PromptCraft: A community-driven resource with curated prompts, tutorials, and challenges. https://medium.com/@mcraddock/the-art-of-prompt-crafting-3fe057b8b566
Blogs and Articles:
- Two Minute Papers: Bite-sized explainer videos on exciting advancements in AI, including prompt engineering. https://www.youtube.com/channel/UCbfYPyITQ-7l4upoX8nvctg
- Hugging Face Blog: Stay updated on the latest developments and research in the field. https://huggingface.co/blog
- Lex Fridman Podcast: Profound conversations with leading AI researchers, often delving into prompt engineering. https://lexfridman.com/podcast/
Online Courses:
- Udacity’s Intro to Generative AI: Structured learning path covering concepts like prompts and GANs. https://www.udacity.com/collection/generative-ai-courses/all/any-price/any-school/any-skill/any-difficulty/any-duration/any-type/most-popular/page-1
- DeepMind’s Deep Learning Specialization: Dive into various AI techniques, including prompt engineering for text generation. https://www.coursera.org/specializations/deep-learning
- Kaggle Learn: Free tutorials and courses on AI and machine learning, often touching upon prompt engineering. https://www.kaggle.com/learn
Community and Discussion:
- Hugging Face Forum: Connect with other prompt engineering enthusiasts, share ideas, and learn from each other. https://discuss.huggingface.co/
- AI Dungeon Discord: Join a vibrant community discussing AI creativity and experimenting with prompts. https://www.reddit.com/r/AIDungeon/comments/qepgu4/does_ai_dungeon_discord_server_still_exist/
- Reddit’s r/PromptEngineering: Stay abreast of the latest trends and discussions on prompt engineering. https://www.reddit.com/r/PromptEngineering/
- Prompt Engineering – https://www.pantherschools.com/how-to-start-your-ai-journey/
