What is prompt engineering?

Previously, in our introduction of domain-driven AI programming, we talked about prompts and how AI-generated code is only as good as the input—or prompt—it receives. In this article, we’ll expand on this with the concept of “prompt engineering”.

Recap: What is a prompt, and why are prompts so important?

Prompts are the set of instructions you provide AI tools as input in order to produce the desired output. 

Think of a prompt as a cooking recipe. Just as a recipe guides you with specific instructions and requires certain ingredients to create the desired dish, a prompt provides the necessary information and structure to generate the desired output from an AI model. This output can vary widely in quality and relevance based on the recipe (prompt) you provide.

What is prompt engineering?

If prompts are like recipes, prompt engineering is the skill of crafting these recipes. 

The key to prompt engineering is to write the prompts in such a way that the AI not only understands what is being asked, but also produces the most accurate and useful responses.  It involves understanding how AI models interpret prompts and using that knowledge to format and structure prompts to achieve the desired results.

This practice is crucial as AI technologies, particularly those based on large language models like GPTs, become more sophisticated and integral to a wider range of applications, from creative writing to complex decision-making to software development.

As AI models, especially those based on large language models like GPT (Generative Pre-trained Transformer), become more advanced, the ability to effectively communicate with these models becomes critical. A well-engineered prompt can mean the difference between receiving a generic, irrelevant response and obtaining a nuanced, highly relevant answer.

Just as a well-written recipe can help a novice cook create a delightful meal, a well-engineered prompt ensures that AI produces outputs that are precise, relevant, and directly applicable to the user's needs. The better the prompt, the better the AI's response, enhancing the effectiveness of AI applications in business, creative industries, and beyond.

Key tips for effective prompt engineering

Be specific

Remember that AIs can’t read your mind, and that they are generally tuned to service anyone. The more specific your instructions, the better the AI can understand and execute the task. Specify the context, the problem to be solved, any constraints, and the expected outcome.

Use clear and concise language

Avoid ambiguities and complexities in language that might confuse the AI. Straightforward and clear prompts help the AI process your request more effectively and reduce the risk of misinterpretation. Think of it as giving directions; the clearer and more direct they are, the easier it is to reach the destination without unnecessary detours.

Contextualize your prompts

Context is key in prompt engineering. Provide background information that helps the AI understand not just the ‘what’ but the ‘why’ behind a prompt. This could involve outlining the industry context, user needs, or specific challenges that the application needs to address. This information can significantly influence the relevance and applicability of the AI-generated outputs.

Iterate and refine

Prompt engineering is rarely a one-shot process. It often requires iterative refinement. Start with a prompt, analyze the output, and refine your prompt based on what was or was not achieved. This iterative cycle helps hone in on the most effective way to communicate with the AI, much like tuning an instrument to get the right pitch.

Leverage examples

When possible, provide examples of what you consider a successful output. Examples can serve as concrete guides for AI systems, illustrating exactly what you expect. This technique can be particularly useful when dealing with more creative or subjective outputs where "correctness" might be open to interpretation.

Understand AI capabilities and limitations

A thorough understanding of what the AI can and cannot do is essential. This knowledge can help in crafting prompts that play to the AI's strengths while avoiding areas where it might struggle. For instance, breaking down a complex task into smaller subtasks might help the AI produce more precise output without exhausting its limitations on how much information it can process at any given time.

Read the manual

AI vendors often have their own documentation on how to provide the best prompts for their models.

Looking to the future with retrieval-augmented generation (RAG)

Let’s explore the cooking analogy a little further with a relatively recent concept called “retrieval-augmented generation” or “RAG”, introduced by researchers from Meta, University College London, and New York University in a 2020 paper.

Imagine a chef that consults multiple cookbooks, colleagues, and the internet to refine a dish or incorporate the latest culinary trends. That is what RAG attempts to bring to the world of AI models. 

One of the main limitations of current AI models is that the output tends to be quite shallow, especially when working with more technical or very specific topics. 

With RAG, the AI attempts to first retrieve specific information on the relevant topic from external databases or documents as context to its response, and then integrates this data with the input prompt to generate a more informed and accurate output. 

Another potential benefit of RAG is enabling the AI model to cite external sources so that users can further verify the output.

Enhancing AI with domain-driven prompt engineering

While Uniscale doesn’t currently utilize RAG, Uniscale works in a similar fashion by enhancing responses with context from the functional specifications built using Uniscale Describe. 

Uniscale also guides AI code generators by providing precise domain information in a structured manner. This method ensures that the code generated is not only relevant but also maintainable and aligned with customer needs.

Domain-driven prompt engineering offers a powerful combination for enhancing interactions with AI systems. By fine-tuning how we interact with AI, we not only improve the quality of outputs but also pave the way for innovations that could redefine entire industries.