Giving Instructions

Introduction

In this document, we cover our first prompting technique, instruction prompting. It consists in providing clear instructions or directives to the AI model.

We have already seen simple instructions used in the doc about prompt engineering like getting an AI to solve the math problem. However, we can instruct Gen AIs to do much more complex tasks.

You'll learn about:

1. What is Instruction Prompting?
2. Why Instruction Prompting is Important
3. Examples of Instruction Prompting
   • Name Parsing
   • Removing Personally Identifiable Information (PII)
   • Essay Evaluation and Feedback
4. Conclusion
5. FAQ

1. What is Instruction Prompting?

Instruction prompting centers around the ability of generative AI models, particularly large language models (LLMs), to follow natural language instructions. Instruction prompting aims to allow models to work with new, previously unseen tasks by understanding and executing instructions expressed in natural language, without requiring task-specific training data.

For example, instead of training a model on multiple examples of extracting the fourth word from a sentence, you can simply instruct it: "Write the fourth word of the sentence."

2. Why Instruction Prompting is Important

Generative AI models rely on well-crafted instructions to understand tasks and deliver accurate results. Instruction prompting enables users to guide AI through complex tasks—like text formatting, PII removal, and essay evaluation—without needing vast labeled datasets or pre-programmed rules.

This method is flexible and scalable, making it particularly useful for adapting AI to new tasks. It also helps mitigate the risks of incorrect outputs by enabling prompt iterations and adjustments.

3. Examples of Instruction Prompting

3.1 Name Parsing

One common task in data processing is formatting names correctly, especially when users input names in different orders (e.g., "John Smith" or "Smith John"). Instruction prompting allows you to automate this task effortlessly.

Try changing the name until you find one that the prompt does not properly handle. If you can do this, perhaps you will be able to improve to prompt as well! This is a fun exercise to improve your prompt engineering skills.

3.2 Removing Personally Identifiable Information (PII)

Organizations often need to redact personal information from documents. With instruction prompting, generative AI can automatically identify and replace PII, saving time and reducing manual effort .

Here is the AI's response:

It is very exciting to see how the model can extrapolate from the instructions. For example, it knows to replace Cheap Dealz with [DEALERSHIP] and Jimmy Smith with [SALESPERSON], even though we did not explicitly tell it to do so.

3.3 Essay Evaluation and Feedback

Instruction prompting can also be used to evaluate essays based on various criteria like grammar, coherence, and argument quality. The model can provide detailed feedback and scores, simulating the task of a human reviewer.

The AI's response:

4. Conclusion

Instruction prompting is a powerful technique that allows generative AI models to handle complex tasks by providing clear, concise directions. Through examples like name parsing, PII removal, and essay evaluation, we can see how flexible and scalable this method is. Instruction prompting will be a key tool as you continue exploring AI capabilities.

5. FAQ

Why is instruction prompting effective?

Although explicitly giving the model a specific task seems simple, LLMs are very good at following even complex instructions and applying them to user inputs.

What are some examples of instruction prompting?

Some examples shown in this doc are formatting name inputs, removing personally identifiable information from emails, and evaluating essays.

Sander Schulhoff

Sander Schulhoff is the Founder of Learn Prompting and an ML Researcher at the University of Maryland. He created the first open-source Prompt Engineering guide, reaching 3M+ people and teaching them to use tools like ChatGPT. Sander also led a team behind Prompt Report, the most comprehensive study of prompting ever done, co-authored with researchers from the University of Maryland, OpenAI, Microsoft, Google, Princeton, Stanford, and other leading institutions. This 76-page survey analyzed 1,500+ academic papers and covered 200+ prompting techniques.

Footnotes

1. Efrat, A., & Levy, O. (2020). The Turking Test: Can Language Models Understand Instructions? ↩

2. Mishra, S., Khashabi, D., Baral, C., Choi, Y., & Hajishirzi, H. (2022). Reframing Instructional Prompts to GPTk’s Language. Findings of the Association for Computational Linguistics: ACL 2022. https://doi.org/10.18653/v1/2022.findings-acl.50 ↩

3. Allowing AI to remove PII from text is a promising approach, but it should be used with extraordinary caution as it may make mistakes. ↩

4. It should also be noted in this example that you are sending PII to a third party (e.g. OpenAI if you are using ChatGPT), and these services may record and use your prompts for their own needs. ↩

5. We used a different GenAI, GPT-3, for this example. ↩