Prompt engineering is the skill of crafting effective prompts to get the best possible outputs from generative AI models like OpenAI’s GPT-4o or Anthropic’s Claude 3.5 Sonnet.

A prompt is the input you provide to an already-trained large language model to direct it to execute and complete a specific task. The art lies in knowing how to compose your prompts to elicit high-quality, relevant responses that are aligned with your goals.

It's important to understand that with large language models, the training is already done; they don’t learn or remember information from one interaction to the next, which means prompts need to be self-contained, providing all necessary context and guidance within the prompt itself.

Effective prompts are clear, specific, and give the model the information and boundaries it needs to generate a successful output.

Precision in prompting helps guide the AI to better quality outputs. Vague, open-ended prompts leave too much room for interpretation. Instead, be as specific as possible about what you want the model to do.

For example:

The more precise version gives the AI clear direction on the purpose, content, and desired outcome of the email. Being specific keeps the model on track.

Using rich and detailed vocabulary not only emphasizes to the model what you want, it also helps the model to understand more specifically what you’re looking for. Think of rich vocabulary as another dimension of being precise. Vivid word choice adds color and nuance that helps the AI understand the style and tone you're aiming for.
​

For example:

The descriptive language provides much more signal to the model about how you want it to modulate its writing style to fit the circumstance. So don't be afraid to paint a picture with your words when writing prompts.

Language models have a fixed amount of "brainpower" to allocate per output generation, or per workflow action. Asking it to tackle multiple complex tasks at once means dividing up that brainpower.
​

For optimal results, break down compound tasks into discrete steps, with a separate focused prompt for each one. This allows the model selected in your chat or workflow action to concentrate on each subtask.
​

For example:

By separating out the steps, you allow the model to focus its full faculties on each task in sequence. The result is higher quality across both the title options and selected song lyrics.

Also, be mindful that making the AI infer missing information depletes its available brainpower. While human communication relies heavily on inference, with AI it's better to err on the side of stating things explicitly, even if it feels redundant. Use labeling and terminology consistently throughout your prompt.

For example:

Consistent labeling will ensure that the model executing your prompt will focus on the most important elements.

To understand why prompting for output structure is crucial to using large language models (LLMs) in your day to day operations, it’s important to remember that by default, large language models are fundamentally probabilistic by design.
​

LLMs operate based on predicting the likelihood of a sequence of words. These models are trained on extensive datasets consisting of text data from a vast corpus of data.
​

When LLMs are trained, the model learns to predict the probability distribution of the next word in a sequence, given the words that come before it. In other words, this means that when you give your model any prompt, the model continuously predicts the next best word over and over again given the words before it.
​

This means that given the same input more than once, the model may provide slightly different outputs each time, and thus underscores the importance of providing guidance on how the model should structure the output in order to be functional and usable by you and your team.
​

There are two ways to instruct the model to structure its output:

Let's take a look at some examples of implicit structure and explicit structure for various common tasks:

Examples can greatly increase the quality of your output as well. When you’re providing examples, it’s great to provide between one and three examples of both the inputs a specific action might see, and the outputs that were composed as a result of completing the task with those inputs.

This allows the model to view how you or someone on your team completed a process by taking your inputs and transforming them into usable, production-ready published outputs.

A (your inputs) → through your process all the way to → B (your desired outputs)

You can use the following bracket notation to configure examples in your prompts as follows:

<examples>

Inputs: [example inputs for the task if available]

Output: [matching example outputs]

</examples>

We recommend that if you need a boost in output quality, using one to three examples should do the trick. In some cases, more may be necessary but it’s important that you test this first. See testing workflows for more information.
​

Causative language can change the way the AI uses a resource in your prompt, either by sticking more closely to that resource, or by transforming the resource more.

If you find the AI isn’t utilizing something you gave it, try using more integrative language.

Using integrative language means closely weaving the desired elements of your prompt into your instructions. You can do this by changing the structure of your sentences, and using syntax that makes doing the task dependent on the elements you want incorporated.

For example:

If you find the AI is sticking too closely to something you gave it, using transformative language can help.

Using transformative language means explaining to the model how it should transform the elements of your prompt, rather than regurgitate or repeat them.

You can use verbs that emphasize how you want those elements used, and you can break the task up and create distance with multiple verbs. You can also identify specific aspects of the element that you want it to use, rather than speaking about the whole element itself.

For example:

Here's a basic prompt template you can adapt for a variety of tasks:

Example prompt: ”You are an expert real estate marketer tasked with analyzing buyer personas to help the sales team better target prospects.

Buyer persona: Kelly Kondo, head of facilities for mid-size tech companies, 12 years of experience in office management and vendor relations.

Analyze the buyer persona and describe the following:

Aim to equip the sales team with actionable insights they can use to tailor their approach to this persona.”

With practice, prompt engineering skills will enable you to get impressive results from generative AI tools. Investing time to learn effective prompting pays dividends in the quality and utility of any large language model's outputs.