What is prompt engineering?

Very few companies have the budget to train a new large language model (LLM) from scratch. Most data science teams work with commercial and open-source LLMs like GPT, LLaMA, Mistral, etc., as they build AI applications for their enterprise.

Customizing these LLMs for your use case requires significant prompt engineering. It is the process of “communicating” with the LLM so your users get the output they expect. Prompt engineering is also one of the ways to prevent unauthorized, false, or inappropriate LLM output from reaching the end user.

This article explores prompt engineering in detail, including beginner and advanced prompting strategies for ML/AI engineers.

Prompt engineering explainedPrompt engineering explained

Prompt engineering is the process of creating and optimizing “prompts”, the natural language inputs needed to use the LLM across a range of tasks. To understand prompt engineering, you first need to know how LLMs work.

Think of the autocomplete feature on your phone. As you type, it suggests the next word, generally based on the latest/frequent word combinations. For example, if you type “Hello sir” frequently, it suggests sir every time you type “Hello.” LLMs work on the same principle, just on a much (much!) larger scale. They make next-word predictions based on training data from millions of existing documents.

To make better outputs, they need detailed input. Prompt engineering gives the context the LLM needs to respond to your user and effectively complete tasks. It maps user input from your AI app (user domain) to the LLM’s document domain so the LLM makes better predictions. AI app developers often have to develop innovative prompts so their AI apps respond correctly to end users.

Prompt engineering examplePrompt engineering example

Consider you are building an AI chatbot that troubleshoots IT problems in an organization called XYZ Inc. Let’s consider the following situation:

Bot: — Hello! How may I help you today?

User: — The printer stopped working during the meeting. I couldn’t print out store analysis reports for my manager.

Feeding this user prompt directly to the LLM will not be helpful. The LLM has no context about what the user expects from your app. It may get side-tracked or start discussing the reports instead of the printer.

That’s where prompt engineering comes in. The app code wraps the user’s prompt in a “script” before feeding it to the LLM. The script may include additional documents to better guide the LLM.

Example prompt engineering script:

You are an IT support specialist for XYZ Inc. The organization has four printers: A, B, C, and D. You are on a call with an XYZ employee. The employee says: “The printer stopped working during the meeting. I couldn’t print out store analysis reports for my manager.” How will you respond? Your first goal is to identify the printer that needs troubleshooting. Then, refer to the relevant printer manual and suggest a solution.

For this example, the two different prompts were given to GPT 4. As you can see, the response with prompt engineering is more focused and results in better outcomes.

Prompt engineering basicsPrompt engineering basics

Prompt engineering includes many principles and practices that guide prompt development. The term “engineering” indicates a structured approach to the generally creative field of writing a prompt. The structure is necessary to ensure consistency in quality and define measurable success parameters. Before diving into different strategies, let’s overview the basic prompt structure. Think of prompts as blueprints — clear, well-organized ones lead to better results, while vague or poorly structured ones might yield irrelevant or inaccurate outcomes.

Prompt structurePrompt structure

Prompt engineering looks at the same prompt from different abstraction levels. At its most abstract, prompt components present the basic building blocks of any AI prompt. They give a broad framework you can apply to any AI system, regardless of the specific platform or model you’re using.

You can also break down a prompt into elements that define the specific instructions that guide the AI. Elements are more customized to the AI system you’re interacting with.

You can look at a prompt as a recipe at a more detailed level. Recipes combine elements in specific ways to achieve the desired outcome. This is where the structure of the prompt becomes more refined and intentional.

The prompt instance is the actual prompt given to the LLM. It contains real data, objectives, and context for the LLM.

Prompt componentsPrompt components

The essential components of a prompt are summarized in the table below. Whenever you write any prompt, try defining it using these five components.

The final prompt:

Summarize the following text in three sentences. The text is a report on customer feedback for a retail store in 2023. Keep your response under 50 characters and in a professional tone. Customer feedback: ‘The staff was helpful, but I had to wait too long for assistance.’

Prompt engineering techniques for beginnersPrompt engineering techniques for beginners

Prompt engineering techniques present specific approaches to prompting for different tasks. Depending on the context, you can keep your prompt short and sweet or write a lengthy paragraph with examples and details. Promptingguide.ai gives comprehensive explanations for each technique. We present a “prompt-engineering cheat sheet” below so you can quickly find the prompt strategy you need for your use case. These prompts are typically used with a single LLM.

Advanced prompt engineering techniquesAdvanced prompt engineering techniques

Some prompt engineering techniques integrate LLMs with other technologies, additional LLMs, or intricate workflows to address more complex problems. This section highlights a few advanced methods, but it is not an exhaustive list; it is an introduction to the many possibilities in the field.

RAG prompt engineeringRAG prompt engineering

Retrieval Augmented Generation (RAG) closes gaps in the LLM’s training data without re-training the entire model. For example, let’s say you want to develop a chatbot that answers questions about your organization’s HR policies. The LLM is generic and has obviously not been trained on your organization’s HR policy documents.

In this case, you can:

1. Take the user’s prompt.

2. Identify keywords and do a smart search on your knowledge source.

3. Pull out relevant documents.

4. Engineer a new prompt enhanced with relevant retrieved information for additional context. Pass the new prompt to the LLM.

You typically need smart search technologies like Amazon Kendra to set up RAG workflows in your organization.

Automatic prompt engineeringAutomatic prompt engineering

In automatic prompt engineering, one LLM generates the prompts, and another LLM (or human) evaluates the prompts before passing them on to a third LLM as input. The prompt-generating LLM is given an example user input/output data set to guide the prompt-generation process.
The workflow typically consists of 2 steps:

1. The prompt-generating LLM generates a set of candidate prompts.

2. A prompt evaluation function (another LLM or human in the loop) evaluates each candidate prompt and returns the prompt with the highest score.

The final prompt is used downstream with a third LLM or the existing ones.

Program-aided language modelsProgram-aided language models

Program-Aided Language Models (PALM) combine LLMs with external programming logic or tools to enhance their performance on more complex, reasoning-based tasks. It integrates language models with structured code execution so the model can handle tasks that require precise calculations, algorithmic thinking, or detailed data manipulation.

Initially, the LLM is given input explaining how to convert the input to code. Then, the user is asked to do the same for a new problem and run the code. In the below figure, the LLM interacts with the Python interpreter to solve mathematical problems.