Prompt Engineering: Your key to unlocking AI’s potential

Let’s begin with “What is Prompt engineering?” In simple words, it is designing the best prompt you provide to AI to get the desired outcome. Prompt engineering is more than just asking AI questions; it’s about unlocking the full potential of AI models by providing them with the right instructions and context. Consider how you use language to ask questions and get answers from people around you. Sometimes you may want to use certain words, particular tone, specific directions in your question to get an answer.

For example:

Generic question:

“Hi Kelly, can you get me something to eat?” More Clear instructions:

“Hi Kelly, can you go to the kitchen? See if there are any fruits – apples or bananas. Get one of each.” In this example Kelly is more likely to get exactly what you want. While today’s LLMs offer remarkable capabilities, it’s important to remember that their effectiveness is directly influenced by the quality of both the training data and the user’s prompts. Thoughtful prompt construction is essential for harnessing the full potential of these models. Being a Sr. Customer Engineer at Google Cloud, I’m every single day talking to companies who are incredibly eager to explore the ability of these new AI models. They often come in with very specific goals and sophisticated evaluation metrics, which is fantastic. But I alwaysy find myself emphasizing one crucial point: the importance of a strong foundation which is 1. Refined knowledge base and 2. Prompt Engineering. Think of it this way – imagine you are a master chef with the most advanced kitchen imaginable. But you don’t have the ingredients you need, would you be able to make the best recipe? Most likely not. Similarly, even the most powerful AI needs high-quality data and carefully crafted prompts to deliver meaningful results. So, before diving into complex testing and analysis, we encourage our clients to focus on building a comprehensive knowledge base and mastering the art of prompt engineering. It’s about understanding how to communicate effectively with the AI, guiding it towards the insights they seek

Before we get into techniques of Prompt Engineering it is important to understand how Large Language Models work and be aware of their limitations as well. Think of Large Language Models (LLMs) as incredibly gifted individuals who have absorbed a vast library of knowledge. These AI systems use a powerful technique called “deep learning” to understand and process information from countless books, articles, websites, and even code. Just like we learn a language by observing patterns and structures, LLMs analyze this massive amount of text to grasp the nuances of human communication. This allows them to engage in conversations, answer questions, and generate text that is remarkably human-like in its fluency and coherence. Essentially, LLMs are sophisticated communicators, capable of understanding complex ideas and expressing them in a clear and engaging manner. This makes them invaluable tools for a wide range of applications, from enhancing customer service interactions to assisting with creative writing and translation. But the results you get are dependent on the data you feed and prompt you provide. So coming back to prompt engineering and its techniques. Below are some widely accepted tech

• Be as specific as possible: When working with AI, clear communication is key. Think of it like asking a friend for help – the more specific you are with your request, the better they can understand your needs and provide help/ useful responses. Just like with a friend, providing detailed instructions to the AI ensures it has all the necessary information to deliver what you’re looking for. Avoid ambiguity and be explicit about the desired format, length, and content, just as you would in a professional setting. This precision streamlines the process and helps the AI generate accurate and relevant output efficiently. Just as I explained in the example at the beginning, Kelley is more likely to provie me with the right fruit only with clear instructions

• Utilize contextual prompts: The second important technique is to use contextual prompts. AI models thrive on context. It means we need to provide background or a framework to our prompts. Think about watching a movie, when we start we generally see where the movie is set, location, time period it gives us an idea what we can expect next. Another example would be working on a marketing campaign and you are looking for slogan for shoes. If you simply ask “write a slogan” it might give out generic sentences however if you provide more context which age group is this targeted at, what materials are used which makes it special and durable or if it is weather resistant, this contextual information will act as a creative brief and will provide more targeted and effective slogans. During a recent pilot program at Google, we collaborated with a manufacturing industry client to develop a generative AI-powered chat assistant. This tool aimed to guide technicians through troubleshooting and resolving automobile issues. Throughout the testing phase, we focused on refining the prompt engineering process, ensuring that the assistant asked precise questions and provided relevant contextual information, similar to how one would guide someone in fixing a hammer drill. This meticulous approach enabled us to optimize the assistant’s effectiveness in diagnosing and resolving problems efficiently.

• Provide Examples: Providing examples in your prompts is incredibly important especially in complex tasks. It is similar to providing a preview or crash course to the AI to spit out accurate results. This is also called “Shot Prompting”. A zero shot prompting is giving AI tasks without any examples, this means the AI model will answer your question based on its general knowledge. It is possible AI will not be able to provide specific answers with zero shot prompting..Google Cloud also provides an option of “grounding with google search” which means it can provide results grounded on google search only if required. One shot prompting is when you provide one example to tell the model what are you expecting, how close the model’s response should be. Few shot prompting is providing more than two examples of the desired output before giving it the actual prompt. In human world the simplest example would be showing a toddler a few images of an umbrella and then asking them to draw it. Similarly, providing more examples to the model will provide efficient, more accurate responses. Be aware that providing too many examples might make it inflexible which might be a deviation from actual desired output.

• Experiment with prompts Tuning: Prompt tuning is a newer approach in natural language processing that focuses on optimizing the prompt itself rather than fine-tuning the entire language model. Instead of adjusting the model’s weights, prompt tuning introduces a small set of learnable parameters within the prompt. These parameters are then adjusted during training to elicit the desired output from the language model. This technique offers several advantages, including reduced computational costs, improved efficiency, and the ability to adapt large language models to specific tasks without modifying the core model. Prompt tuning is showing promising results in various applications, from text classification and question answering to code generation and machine translation. This can be used in several applications.

• Iterate: Lastly it is important to repeat all these techniques and let the AI model learn. As rich the data is and strong the prompts are the AI model will be close to accurate always. Prompt Engineering technique overall isn’t a one-and-done process, it needs to be refined continuously. Each iteration brings you closer to the desired outcome. Perhaps you need to add more context, adjust the tone, or provide specific examples. By iteratively tweaking your prompts, you guide the AI towards a deeper understanding of your intentions and unlock its full potential.

In the grand symphony of AI, where data is the melody and LLMs the orchestra, prompt engineering emerges as the conductor’s baton, orchestrating a harmonious outcome. It’s the key to unlocking the true potential of these powerful language models, transforming them from mere tools into collaborative partners. As we venture deeper into the age of AI, remember that effective communication is paramount. By mastering the art of prompt engineering, we not only enhance the quality of AI’s output but also elevate our own understanding of these complex systems. It’s a journey of continuous exploration and refinement, where each interaction brings us closer to a future where humans and AI work together seamlessly to achieve extraordinary things.