Tips for generative AI large language model (LLM) prompt patterns

A generative AI (gen AI) model, such as ChatGPT, is trained using a variety of text from the internet. However, artificial intelligence (AI) doesn't have specific information about particular documents or sources used during its training. Instead of focusing on specifics, the model is trained to be general, allowing it to come up with creative answers, engage in complex conversations and even display a sense of humor. AI doesn't possess comprehension, understanding or belief, however. Its responses are generated based on patterns learned from training data.

AI systems like ChatGPT or any large language model (LLM) are reflections of humanity's collective knowledge in a single interface. They reorganize existing content from the internet, but do not "think", are not "intelligent" in the human sense, have no "general intelligence" as general problem solvers and are not "conscious” of what they find.

How generative AI works: What are tokens?

The way these models operate is based on the concept of tokens, which are discrete units of language ranging from individual characters to whole words. These models process a specific number of tokens at a time using complex mathematical calculations to predict the most likely next token in a sequence.

Models like generative pre-trained transformers (GPTs) generate text one token at a time. After producing each token, the model reviews the entire sequence it has generated so far and processes it again to generate the next token. This recursive process continues until the final token completes the generated text.

This means that the quality of the AI's response depends on the prompt or instruction that a user provides. In other words, the way we interact with and instruct AI significantly influences the quality of the answers it produces.

What is prompt engineering?

Prompt engineering refers to the practice of designing and crafting effective prompts or instructions for AI models to achieve desired outputs. In the context of language models like GPT-3, prompt engineering involves formulating input text that leads the model to generate accurate, relevant and contextually appropriate responses.

Effective prompt engineering is crucial because language models like GPT-3 don't possess true understanding or common sense reasoning. They generate responses based on patterns learned from training data. Crafting well-designed prompts can help guide the model to produce more accurate and meaningful outputs, while poorly formulated prompts might lead to incorrect or nonsensical results.

What is prompt design?

Prompt design is the systematic crafting of well-suited instructions for an LLM like ChatGPT, with the aim of achieving a specific and well-defined objective. This practice combines both artistic and scientific elements and includes:

• Understanding the LLM: Different LLMs respond differently to the same prompt. Additionally, certain language models might have distinct keywords or cues that trigger specific interpretations in their responses
• Domain expertise: Proficiency in the relevant field is crucial while formulating prompts. For example, creating a prompt to deduce a medical diagnosis requires medical knowledge
• Iterative process and evaluating quality: Devising the perfect prompt often involves trial and refinement. Having a method to assess the quality of the generated output that goes beyond subjective judgment is essential

Prompt size limitations

Recognizing the significance of an LLM's size constraint is crucial, because it directly influences the quantity and nature of information we can provide. Language models aren't designed to handle an infinite amount of data all at once. Instead, there's an inherent restriction on the size of the prompt you can construct and input. This limitation has profound implications for how you formulate and utilize prompts effectively.

An LLM has a maximum token capacity encompassing both the prompt and the ensuing response. Consequently, longer prompts could curtail the length of the generated response. It's important to craft prompts that are concise, but that convey the necessary information.

In practical scenarios, you must adopt the role of an editor, carefully selecting pertinent details for a task. This process mirrors the way you approach writing a paper or article within a specific word or page limit. In such cases, you can't simply dump random facts. Instead, you must thoughtfully choose and organize information that's directly relevant to the subject matter.

Prompt design is a human skill that helps ensure accurate and well-structured content. Tools can help a writer be more productive, but they're no substitute for a writer. Gen AI requires the assistance of a knowledgeable and experienced writer, researcher and editor. To be a good prompt designer, you need the same skills that make a good content writer.

Techniques for prompt design and prompt engineering

What prompt you decide to use in which situation is up to you. For instance, if you’re writing a long article, chain-of-thought prompting is useful for generating information one section at a time. The following is not an exhaustive list of all prompt design strategies, but are some of the top patterns in use today.

Persona Pattern

The Persona Pattern is a technique that directs the AI to operate from a specific standpoint, like adopting a professional role or embodying a fictional character. This approach empowers the model to generate outputs aligned with the responsibilities associated with that particular persona. For instance, you might instruct the AI to take on the role of a cybersecurity expert while evaluating code. This strategy becomes particularly valuable when you're seeking guidance, but lack precise specifications for the desired output. For example:

In this scenario, it's anticipated that the AI will provide a comprehensive examination of the Peloponnesian War with a concentration on its causes, as you'd expect from a historian's viewpoint. Note the elevated, academic writing style and vocabulary used in the prompt. The response will match that style. In other words, your writing style in the prompt determines the style of the generated content. This is true for all prompting approaches, so tailor your prompt writing for the context in which you’re publishing.

From a marketer’s perspective, you can use the Persona Pattern to get into the mind of your customers. Without revealing any confidential information, you can ask AI to assume the persona of your target audience, then ask it questions. The AI will respond from a theoretical customer's perspective. Of course, this isn’t foolproof and is in no way a replacement for good user research and marketing data. However, it can help you write different forms of copy and generate ideas.

New information pattern

AI has limitations and one very significant one is that it's not aware of information beyond the point where its training data ends. AI can't include recent events or consider updates that have happened after its training. However, you can incorporate new information into the model's knowledge by structuring your prompts effectively. For instance, consider the following interaction:

By providing a little more context in the prompt, you can create a more relatable explanation:

Refining questions pattern

AI possesses the capability to assist in improving the quality of your inquiries. For instance, if you pose a question, AI can propose a precisely formulated question that would likely yield more accurate results.

This ability is encapsulated in the refining questions pattern. It proves especially valuable when you lack expertise in a particular field or aren't certain about the best way to phrase your question. Consider this sample prompt:

In this scenario, the AI is expected to enhance your general data science queries by narrowing them down to questions that specifically consider the aspects of statistical analysis. The AI will then inquire whether the user wishes to proceed with the refined question.

Cognitive verifier pattern

The cognitive verifier pattern involves a method of deconstructing intricate questions into smaller, more manageable sub-questions. The answers to these smaller components are subsequently integrated to generate a comprehensive response to the main question. This approach becomes particularly valuable when the initial questions are overly broad or when additional clarity is required to deliver a thorough answer. Consider this example:

In this scenario, the AI dissects the complex question concerning the search inference framework into three sub-questions. Once it has gathered answers for each sub-questions, it amalgamates them to produce a comprehensive and accurate answer.

This pattern entails the generation of multiple inquiries, leading to a more precise ultimate response. The process involves both verifying information and consolidating it to ensure a robust and thorough output. It's similar to the whole-part-whole method of learning and instruction in education strategy.

Chain-of-thought prompting

In a recent research paper titled Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, Jason Wei and his team aimed to mimic reasoning abilities within an LLM. Departing from the conventional practice of using direct prompts, the team employs a series of interconnected prompts that form what they call a chain-of-thought.

For instance, consider a user who wants to explore the potential impact of artificial intelligence on job markets:

The approach of chain-of-thought prompting entails a sequence of connected prompts that establish a "conversation" with the language model. This approach guides the model toward more precise and contextually fitting responses, fostering a progression of ideas and reasoning. It can also train the AI to mimic your specific writing style, so long as your prompts are written in the same tone, style, attitude and construction as your ideal answer.

Chain-of-thought is especially effective for article writing. If you have an article, an essay or an argument already outlined, you can structure your chain of prompts along that flow. Because an AI has a limit to how many characters it can give you, chain-of-thought can help you prompt information for one section of your piece at a time, rather than trying to get the AI to generate an entire article in one go. This allows you to have more control over the information and get more accurate, detailed responses. You control the structure, while AI generates chunks of the whole piece, one prompt at a time, under your guidance.

Research assistant pattern

In this prompt, you request assistance in finding sources for a research project. AI responds by recommending three relevant research articles. It provides the titles, authors and publication details of the articles, along with brief descriptions of their content. The AI also emphasizes the importance of cross-referencing and evaluating the credibility of the sources. The prompt design effectively leverages AI's capabilities to assist in research by suggesting reputable sources for further study. For instance:

Crucially, this prompt helps you avoid one of the major weaknesses of AI and LLMs: misinformation. By asking it for sources instead of content, AI can dramatically reduce the time required to research for your project, but you still evaluate the sources and pull what information from them you need.

Citation generator prompt enhancement

Similarly, when you ask AI to write content, you can request that it cite its sources and provide a references section. Declare which style of citation you prefer and the interface returns that format. For instance, consider this abbreviated example:

As always, you’ll want to act as editor and check the accuracy of the information. Examine each of the sources the AI provided and ensure that the claims that AI makes about a source's arguments are true. Additionally, verify that the references section itself is correct APA formatting.

Depending on the field you’re researching, you’ll likely be able to quickly recognize some of the big names in the references section, which is a shorthand way to check for accuracy. For instance, Alain Aspect was awarded the 2022 Nobel Prize in Physics for his decades-long work on quantum entanglement, so when his name tops the references list from a prompt about quantum entanglement is a promising sign that the source that the AI used is reliable.

This type of prompt is another example of why human writers, editors and researchers with expertise in a given subject matter are crucial to the content writing process. AI can speed up the process, but it's no substitute for the human writer.

Asking AI to cite its sources is especially powerful when paired with chain-of-thought prompting. To combine the two, follow the prompting strategy for chain-of-thought, as explained above. Before submitting each individual prompt in the chain, ask the AI to cite its sources in the format you require. Then submit the request.

Few-shot pattern

Few-shot prompting is a potent method used to train AI models to adhere to specific patterns or execute tasks. The fundamental idea involves providing the model with a handful of examples before posing the actual question. This practice assists in defining the desired behavior and guiding the model's subsequent responses.

Imagine a situation where you want to craft a succinct and captivating marketing message for a new tech gadget:

The remarkable aspect of few-shot learning is its reliance on demonstrations rather than detailed procedural instructions. This feature makes it adaptable for various applications. This technique goes beyond basic input-output matching. It enables the model to learn how to handle intermediary steps to achieve a particular goal. This method is ideal for helping marketing writers develop iterations of short-form copy, such as headlines, short body copy, subject lines, social media ads, calls-to-action (CTA) and so on.

Now that you know these especially helpful prompt patterns, give them a try in the LLM of your choosing. I used ChatGPT in my examples.

Artificial intelligence is powering new discoveries and experiences across fields and industries. If you’re ready to learn more, explore how to use Red Hat platforms to build, deploy, monitor and use AI models and applications, accelerated by the speed, trust and transparency of the open source community.

Learn about AI from Red Hat