Mastering Prompt Engineering in 2024

Image by Editor | Ideogram & Canva

In a previous post, we covered the prompting framework highlighting the role of persona, context, tone, expected output, etc. to design a comprehensive prompt.

However, despite the framework, there are still challenges, such as data privacy, hallucination, and more. This article focuses on various prompting techniques and outlines best practices to nudge the model with the most appropriate response.

Let’s get started.

Types of Prompting Techniques

Image by Author

1. Zero-Shot vs. Few-Shot Prompting

Zero-shot and few-shot prompting are fundamental techniques in the prompt engineering toolkit.

Zero-shot prompting is the easiest way to solicit the model’s response. Given that the model is trained on massive datasets, their response generally works well without any additional examples or specific domain knowledge.

Few-shot prompting involves showing specific nuances or highlighting complexities around the task by showing a few examples. It is particularly useful for tasks that require domain-specific knowledge or the ones that require additional context.

For instance, if I say, ‘cheese’ is ‘fromage’, then ‘apple’ is ‘pomme’ in French, the model learns information about a task from a very limited number of examples.

2. Chain of Thought (CoT) Prompting

In our prompting framework, we prompted the model to show the step-by-step approach to arriving at the answer to ensure it does not hallucinate. Similarly, Chain of Thought encourages the model to break down complex problems into steps, in the same way a human would reason. This approach is particularly effective for tasks requiring multi-step reasoning or problem-solving.

The key highlight of CoT prompting is that the step-by-step thought process ensures the model shows its work, thereby saving it from figuring out the response by itself.

Image by Promptingguide.ai

3. Retrieval-Augmented Generation (RAG)

Retrieval-augmented generation combines the power of large language models with external knowledge retrieval. But, why is external knowledge needed? Aren’t these models trained on large enough data to generate a meaningful response?

Well, despite seeing the massive training data, the model can benefit from additional information that is derived from specialized domains. Hence, RAG helps by providing more accurate and contextually relevant responses, thereby reducing ambiguity and guesswork, mitigating hallucinations.

For example, in legal or medical domains where precise, current information is critical, the domain experts often refer to up-to-date cases or specialized knowledge that helps them make more informed decisions to appropriately handle their tasks. Similarly, RAG becomes the model’s go-to expert that provides specific, authoritative sources.

Watch Out for Data Privacy

Despite the power of these techniques, prompt engineering faces several challenges, data privacy being one of the most prominent.

With growing awareness of how models train and process data, users are increasingly concerned about models even accessing their prompt data to further tune and enhance outcomes. And this fear is legitimate.

The ways of working are fast evolving. Organizations must adopt robust data governance frameworks, thereby ensuring the privacy and security of sensitive enterprise data.

Best Practices for Effective Prompting

Talking about revised ways of working, it is time to follow the best practices to get the maximum out of prompt engineering:

1. Fact-checking

There has been a recent case of a model fabricating a fake legal case, showing the responsible lawyers in a bad light. As reported on Reuters, they confessed to making “a good faith mistake in failing to believe that a piece of technology could be making up cases out of whole cloth.”

This highlights the lack of awareness of the tool at hand. One must not only know what the model is capable of but also its limitations.

Hence, it is recommended to always verify the information generated by AI models, especially for critical or sensitive tasks. Do not just limit your homework to this, but also cross-reference with reliable sources to ensure accuracy.

An example prompt in such a case could be: “Provide three key statistics about AI adoption in the industry of your interest. For each statistic, include a reliable source that I can use to verify the information.”

Image 1 from Guardian | Image 2 from Reuters

Or, you can prompt the model to: “Summarize the latest developments in the AI landscape. For each major development, provide a reference to a relevant research paper or reputable tech news article.”

2. Thorough Thinking

Before generating a response, enforce the model to think through the problem thoroughly by considering various aspects of the task.

For example, you can ask the model: “Consider the ethical, technical, and economic implications before responding. Generate a response only when you’ve thought it through.”

3. User Confirmation

To ensure that the model response aligns with the user’s intent, you can ask it to cross-check and confirm with you before proceeding with the next steps. In case of any ambiguity, you can nudge the model to ask any clarifying questions to better understand the specific task.

For example, you can ask it: “Outline a marketing strategy for an AI-powered healthcare app. After each main point, pause and ask if you need any clarification.”

Or, you can also prompt: “If you need any clarification about specific industries or regions to focus on, please ask before proceeding with the analysis.”

Wrapping Up

I hope these prompting techniques and best practices serve you well in your next best use of AI. All in all, prompting involves creativity and critical thinking, so let’s get your creative hats on and start prompting.

Vidhi Chugh is an AI strategist and a digital transformation leader working at the intersection of product, sciences, and engineering to build scalable machine learning systems. She is an award-winning innovation leader, an author, and an international speaker. She is on a mission to democratize machine learning and break the jargon for everyone to be a part of this transformation.