Introduction to prompt engineering

GPT-3, GPT-3.5, and GPT-4 models from OpenAI are prompt-based. With prompt-based models, the user interacts with the model by entering a text prompt, to which the model responds with a text completion. This completion is the model’s continuation of the input text.

While these models are extremely powerful, their behavior is also very sensitive to the prompt. This makes prompt construction an important skill to develop.

Prompt construction can be difficult. In practice, the prompt acts to configure the model weights to complete the desired task, but it's more of an art than a science, often requiring experience and intuition to craft a successful prompt. The goal of this article is to help get you started with this learning process. It attempts to capture general concepts and patterns that apply to all GPT models. However it's important to understand that each model behaves differently, so the learnings may not apply equally to all models.

Basics

This section covers the basic concepts and elements of GPT prompts.

Text prompts are how users interact with GPT models. As with all generative language models, GPT models attempt to produce the next series of words that are most likely to follow from the previous text. It's as if we're saying What is the first thing that comes to your mind when I say <prompt>?. The examples below demonstrate this behavior. Given the first words of famous content, the model is able to accurately continue the text.

As you develop more complex prompts, it's helpful to keep this fundamental behavior in mind. Regardless of the prompt that is provided, the model is simply responding with what it determines is most likely (given its training data and training targets). If you ask a question in your prompt, the model isn't following a separate “Q&A” code path, but rather it appears to answer the question because an answer is the most likely sort of response for the given question as input.

Prompt components

When using the Completion API while there's no differentiation between different parts of the prompt, it can still be useful for learning and discussion to identify underlying prompt components. With the Chat Completion API there are distinct sections of the prompt that are sent to the API in the form of an array of dictionaries with associated roles: system, user, and assistant. This guidance focuses more generally on how to think about prompt construction rather than providing prescriptive guidance that is specific to one API over another.

It's also important to understand that while there could be other valid ways to dissect prompts, the goal of this breakdown is to provide a relatively simple way to think about prompt construction. With the Completion API, all the components are optional, but at least one must be present and most prompts include more than one component. There can be some grey area between components as well. The order presented below roughly corresponds to how commonly each component is used, from most to least.

Instructions

Instructions are likely the most commonly used prompt component. Instructions are simply that: instructions to the model on what to do. While simple in concept, they can be complex in practice. The following table illustrates this by showing both simple and more complex instructions, in two different forms.

Primary Content

Primary content refers to some sort of text that is being processed or transformed by the model. Primary content is typically used in conjunction with instructions. A simple example would be language translation. In the example below, the English text would be considered the primary content while "Translate to French:" is an instruction.

Primary content can also be much longer. In the following example, the primary content is the introduction section from the Wikipedia entry on time complexity, which is almost 300 words long. This content has been abbreviated in our table for ease of display.

GPT models can also handle primary content that is structured. In the example below, a TSV about beers in Belgium (abbreviated for display in this article) is passed to the model as part of the prompt. It is capable of properly interpreting the content and answering questions about the data.

Examples

Successful prompts often rely on the practice of “one-shot” or “few-shot” learning. This refers to the inclusion of one or more examples of the desired behavior of the model, typically by including input and output pairs. This is not learning in the sense that the model is permanently changed, but rather that the examples better condition the model to respond as desired for only the current inference. The use of prompts with no examples is sometimes referred to as “zero-shot” learning. Note that with the Chat Completion API few-shot learning examples are typically added to the messages array in the form of example user/assistant interactions after the initial system message.

The example above illustrates the utility of few-shot learning. Without the examples, the model seems to be guessing at the desired behavior, while the examples cleanly show the model how to operate. This also demonstrates the power of the model, it can infer the category of label that is wanted, even without a "basketball" label in the examples.

Cue

Cues act as the "jumpstart" for the output of the model, helping to direct the model to the desired output. It is often a prefix that the model can build onto. In the example below, we show the cue directing the story to include key elements. Cues are often used along with instructions, but not always. In the examples below, the model is being used to summarize a long announcement email. A cue is used in one case to help focus the output of the model and to suggest a specific output format (bullet points) in the other.

In the context of the Chat Completion API instructions take the form of the system message, and few-shot examples are indicated by providing a series of example user/assistant interactions to help prime the model.

Supporting content

Supporting content is information that the model can utilize to influence the output in some way. It differs from primary content in that it is not the main target of task, but it is typically used along with primary content. Common examples include contextual information such as the current date, the name of the user, user preferences, and so on. The example below uses supporting content to help organize a set of planned workshops for the user. Without the supporting (important topics) the model simply lists off the workshops (truncated for display), when informed of my important topics, the model is able to accurately group the meetings.

Best practices

• Be Specific. Leave as little to interpretation as possible. Restrict the operational space.
• Be Descriptive. Use analogies.
• Double Down. Sometimes you might need to repeat yourself to the model. Give instructions before and after your primary content, use an instruction and a cue, etc.
• Order Matters. The order in which you present information to the model might impact the output. Whether you put instructions before your content (“summarize the following…”) or after (“summarize the above…”) can make a difference in output. Even the order of few-shot examples can matter. This is referred to as recency bias.
• Give the model an “out”. It can sometimes be helpful to give the model an alternative path if it is unable to complete the assigned task. For example, when asking a question over a piece of text you might include something like "respond with "not found" if the answer is not present." This can help the model avoid generating false responses.

Space efficiency

While the input size increases with each new generation of GPT models, there will continue to be scenarios that provide more data than the model can handle. GPT models break words into "tokens." While common multi-syllable words are often a single token, less common words are broken in syllables. Tokens can sometimes be counter-intuitive, as shown by the example below which demonstrates token boundaries for different date formats. In this case, spelling out the entire month is more space efficient than a fully numeric date. The current range of token support goes from 2,000 tokens with earlier GPT-3 models to up to 32,768 tokens with the 32k version of the latest GPT-4 model.

Given this limited space, it is important to use it as efficiently as possible.

• Tables – As shown in the examples in the previous section, GPT models can understand tabular formatted data quite easily. This can be a space efficient way to include data, rather than preceding every field with name (such as with JSON).
• White Space – Consecutive whitespaces are treated as separate tokens which can be an easy way to waste space. Spaces preceding a word, on the other hand, are typically treated as part of the same token as the word. Carefully watch your usage of whitespace and don’t use punctuation when a space alone will do.

Next steps

Learn more about Azure OpenAI.