Lunch & Learn with Jobsity: Prompt Engineering Techniques for Achieving Desired Outcomes with LLMs

Welcome to May’s Lunch & Learn! Every month at Jobsity, we have one of our in-house experts cover an industry topic. Last week, we had the pleasure of hosting Node.js developer Miguel Obando. He shared his expertise on prompt engineering techniques for achieving desired outcomes when working with Large Language Models (LLMs). 

Here’s a recap for those of you who missed it:

LLMs are sophisticated artificial intelligence models designed to process complex information, especially human language. By analyzing vast datasets of text and code, LLMs learn to identify patterns and relationships. This allows them to perform tasks like text generation and language translation. 

Miguel broke down the mathematical concepts behind LLMs, explaining that while humans represent words with letters (e.g., C-A-T for cat), AIs use vectors (e.g., [0.0074, 0.0030, -0.0105, …0.443]). This vector representation allows LLMs to understand and generate language in a way that mimics human thought processes.

Transitioning from the basics, the session highlighted the importance of data training. Different approaches to data training can significantly impact the performance of LLMs, so understanding it is a vital part of optimization. Miguel demonstrated how to effectively enter prompts in ChatGPT 3.5. He also discussed the unique capabilities of Meta's LLaMA3, emphasizing how these tools can be harnessed for various applications.

To illustrate the practical uses of LLMs, Miguel took us through several real-world applications. LLMs can track and analyze customer feedback, power customer service chatbots, extract complex data insights, and support internal knowledge bases. These applications show the extraordinary cross-industry versatility and utility of LLMs.

Miguel also discussed the importance of good prompting. He used the Chevrolet of Watsonville case to illustrate how well-crafted prompts can significantly impact outcomes, demonstrating the practical value of this skill. To provide attendees with practical skills, Miguel provided a comprehensive framework for developing effective prompts: 

• Context: Clearly define what the prompt is about and what information is needed.        
• Task: Specify the action you want the model to perform.                                     
• Instruction: Guide the model on how to respond, including the desired format and content.                                                                                     
• Clarify: Offer additional information to enhance the model’s understanding.            
• Refine: Add limitations to ensure accurate responses.

Miguel then shared 3 unconventional, yet effective tips:

1. Address the intended audience in the prompt (e.g., “this response is for VPs of X company”).
2. Use strong directives like “Must” or “Task.”
3. Implement delimiters for clarity.

As a final piece of advice, Miguel emphasized summarizing and expanding content. For summarizing, define what should be considered, request specific details, and limit the response length. For expanding, use factual information, provide context before the user’s question, set the temperature (randomness level) to low, and always have a human backup while logging every prompt sent. 

Finally, he made sure to mention the importance of thorough testing before deploying anything into production. This ensures reliability and accuracy, preventing potential issues down the line. 

The Lunch & Learn is available for on demand viewing.

At Jobsity, we're building more than just a workforce; we're fostering a community of passionate learners and innovators.  Ready to join this dynamic group? Apply on our careers site today!