The Vocabulary of Prompt Engineering

Before we dive into the specific guidelines for crafting prompts, let’s first examine some key definitions commonly used in the realm of prompt engineering and large language models.

Large Language Model (LLM): The core principle of LLMs lies in their ability to generate responses one word at a time. Each word is predicted based on the preceding text, creating a chain of dependencies. This mechanism allows the model to tailor its output according to the context of the conversation, facilitating a more nuanced understanding and producing contextually relevant responses.

The Prompt: A prompt is the initial input provided by the user to request a response from the model. It may take the form of a question, statement, or any text that directs the model’s output.

Prompt Engineering: This term describes the practice of designing and refining prompts for language models. By framing questions effectively, users can steer the model’s responses, ensuring they are precise, informative, and relevant to the context.

The Importance of Precision

When an AI system encounters vague or generalized input, the resulting output tends to be similarly unclear. For instance, if you ask a language model a broad question like “Tell me about animals,” it might produce a generic overview, listing various categories without providing in-depth details.

Conversely, as the input becomes more specific, the output improves in accuracy. For example, asking, “Can you share 5 facts about cats?” prompts the AI to deliver a focused and detailed response.

It is crucial to structure your input thoughtfully and provide any necessary context, as the quality of the response is heavily influenced by the input specifics. Generic inquiries yield generic results.

Instead of saying:

“Write a short story for kids,”

Consider providing more context:

“Write a humorous soccer story for children that teaches the importance of persistence, inspired by the style of J.K. Rowling.”

Types of Prompts

There are three main types of prompts:

1. Zero-Shot Prompt: A straightforward question without any examples. For instance, “What is the capital of France?”

2. One-Shot Prompt: Here, the user offers a specific question along with an example to guide the AI. For example, “Translate to Spanish: ‘Where is the nearest bookstore?’”

3. Many-Shot Prompt: In this case, the user provides multiple sentences as context, asking the AI to generate a summary based on that information. For example:

   “Here are three historical events:

   1. The signing of the Declaration of Independence in 1776.
   2. The fall of the Berlin Wall in 1989.
   3. The Apollo 11 moon landing in 1969.

   Write a narrative that connects these events and discusses their historical significance.”

General Methodology for Crafting Prompts

In prompt engineering, various parameters can be adjusted to influence the model's behavior and output quality. Here are a few key parameters:

• Temperature: This controls the randomness (or creativity) of the output. Values range from 0 to 1, where higher values yield more creative responses while lower values produce more deterministic outputs.
• Max Tokens: This parameter sets a cap on the length of the generated response.
• Special Tokens: Placeholders can be used in prompts to guide the model. For example, [USER] might indicate where the user’s input should go.
• Language and Tone: The choice of language and tone—whether formal or informal—can significantly affect the style and feel of the generated output.
• Task-Specific Instructions: Providing explicit instructions regarding the desired format or constraints can help the model produce more targeted and relevant responses.

Techniques in Prompt Engineering

If you want to apply a consistent format across all outputs, consider using the following approach:

“Whenever you generate output, ensure it is clear and concise. From now on, present information in bullet points and include relevant examples for better understanding.”

To revert back to normal responses, simply state:

“Go back to standard ChatGPT behavior.”

Persona Pattern: One of ChatGPT's impressive features is its ability to adopt various roles, from a friendly customer service agent to a knowledgeable expert in a specific area. To assume a particular role, use the following template:

“Act as Persona X and perform task Y.” For instance, “Act as a skeptical computer scientist and provide a detailed critique of my statement.”

Chain-of-Thought Prompting: Introduced in 2022, this technique enhances LLMs' reasoning capabilities by guiding them through intermediate steps. When paired with few-shot prompting, it can yield better results for complex tasks requiring logical reasoning.

If the input includes logically structured sentences, the output will follow the same logic. For example:

• Input: Brick → Output: Hard
• Input: Pillow → Output: Soft

A recent innovation is the concept of zero-shot Chain-of-Thought (CoT), which involves adding the phrase “Let’s think step by step” to the original prompt.

This allows language models to generate a chain of reasoning that leads to more accurate answers. The “Cognitive Verifier” pattern encourages the LLM to produce a series of follow-up questions that refine the original query.

To implement the Cognitive Verifier Pattern, your prompt should state:

“When asked a question, follow these rules: Generate additional questions that will help answer the main question and combine their answers to formulate a comprehensive response.”

For example:

“When planning a trip, generate questions about my budget, preferred activities, and transportation options, then use those answers to create an optimal itinerary.”

If you're unsure about what additional details to include in your prompt, ask the user clarifying questions until you have enough information to craft an optimal prompt.