Overview

Prompt engineering is the strategic crafting of prompts to guide Large Language Models (LLMs) like GPT-4 to produce accurate and desired outputs.

Before prompt engineering

Before you follow along, we assume that you:

• have a first draft of your prompt
• know the audience that you are tailoring your prompt to
• have some benchmark to measure prompt improvements
• have some example inputs and desired outputs to test your prompts with

We recommend taking a moment to brainstorm on these points to make the most of the following suggestions.

How to prompt engineer

1. Be specific and clear
2. Use structured formats
3. Leverage role-playing
4. Implement few-shot learning
5. Use constrained outputs
6. Use Chain-of-Thought prompting

When to start prompt engineering?

• Start from the beginning. It’s never to early to think about how your prompt will affect the output.
• When you are refining model outputs to meet your expectation.
• When you are expanding features and need the model adapts to the new use cases.
• When you want to optimize cost and performance to reduce token usage, lower latency, and improve performance.

Why prompt engineer?

• Get more accurate and relevant responses.
• Get the response in a specific instructions, styles, or formats.
• Reduce costs by decreasing the number of tokens used, lowering API costs.
• Avoid inappropriate or biased outputs.
• Get consistent and reliable responses across different interactions.
• Improve user experience with more helpful and concise responses.

FAQ

How often should I update my prompts?

• Regularly: Especially if you notice changes in the model’s performance or after updates to the model.
• After use feedback: Incorporate feedback to improve prompt effectiveness.
• When introducing a new feature: Adjust the prompt to cover new functionalities or use cases.

What's the difference between prompt engineering vs. finetuning?

Prompt engineering is modifying the input prompts to guide the model’s responses without changing the model itself. Fine-tuning is training the model on additional data to adjust its internal parameters for specific tasks.

What are some common mistakes with prompt engineering?

• Vague instructions: Leads to unpredictable outputs.
• Overcomplicating prompts: Too much information can confuse the model.
• Ignoring model limitations: Expecting the model to perform tasks beyond its capabilities.
• Lack of testing: Not validating prompts with various inputs can result in inconsistent performance.

How does prompt length affect model responses?

• Short prompts can lead to ambiguous or generic answers because of a lack of context.
• Long prompts provides more detail but can increase token usage and overwhelm the model.

The optimal balance is aiming for concise prompts that include all necessary information without unnecessary verbosity.

Can prompt engineering remove biases?

Yes. When you carefully crafting prompts to avoid sensitive topics or by instructing the model to follow ethical guidelines, you can reduce the likelihood of biased or inappropriate responses.

Do I need to be technical to prompt engineer?

• For simple prompt adjustment, no extensive technical background is needed.
• For complex tasks or with the intention to optimize performance, some familiarity with AI concepts is helpful.

Can you cut cost with prompt engineering?

Yes. A well-written prompt can minimize the number of tokens required for both the input and output, thereby reducing API usage costs.