Prompt Engineering
The technique of prompt engineering, which entails the crafting of precise, task-specific instructions in natural language, either manually or through automated means, and the careful selection of representative examples for inclusion in the prompt, has become a central area of investigation for LLMs.
To effectively use Large Language Models (LLMs) and generative AI, you will need to construct a short textual description, called a prompt, that the LLM will use to generate text, images, or other media. Usually, you will need to iterate and change your prompt to improve the LLMs outputs a few times until you have a result that meets your needs.
How you construct your prompt will also impact aspects of the model's output, like accuracy, verbosity, or style customization. The set of techniques and approaches for constructing prompts is collectively called prompt engineering.
In Anthropic's Prompt Engineering Guide, they suggest that before you begin applying specific prompt engineering techniques, you have a clear definition of success criteria for your use case, ways to empirically test those criteria, and a first draft of your prompt.2
Popular Prompt Engineering Techniques
Let's explore some popular prompt engineering techniques in detail.
Zero Shot
In a Zero-Shot prompt, ask the LLM to perform a task that the LLM has not been explicitly trained on. The resulting output is dependent on the LLM general knowledge and understanding based on its training data.
Example: Please translate this sentence into Spanish, "The 2024 Summer Olympics are being held in France".
The LLM likely hasn't been trained on translating this exact sentence but uses its general understanding and translation patterns to provide a response.
Including Examples or Multi-shot Prompting
In this technique, you provide a few (1-5) examples of what you would like to see in the LLMs output as part of your prompt. Multi-shot prompting particularly helpful when you require the output to include structured data like FOLIO JSON documents.
Example: You are an expert cataloger, please return any records as FOLIO JSON, here is an example:
Q: Parable of the Sower by Octiva Butler, published in 1993 by Four Walls Eight Windows in New York
A: {"title": "Parable of the Sower", "source": "ChatGPT",
"contributors": [{"name": "Octiva Butler", "contributorTypeText": "Author"}],
"publication": [{"publisher": "Four Walls Eight Windows", "dateOfPublication": "1993", "place": "New York"}] }}
Chain-of-Thought
Another useful prompt technique is called chain-of-thought where you explicitly ask the LLM to provide the reasoning or steps the model went through to construct the output. This can be helpful if you're confused how or why a model returns the output to your prompt.
Example Create a circulation report template that includes all books that have been checked out in the past month. Please show your thinking step-by-step as you construct the report.
Remember you can combine multiple prompt techniques!
For more examples, please check-out Prompt Engineering Guide3.
Function calling or connecting Systems to External Tools
Related to prompt engineering is a powerful technique called function calling that allows you to provide a LLM with a function signature and expected parameters. LLMs response will then generate a function call and parameter values that can then be integrated into a Compound AI system. Specific information on this technique with various LLMs: