The Ultimate Guide to Prompt Engineering

You've read about how AI can help you articulate your taste, its transformation of the knowledge economy, its potential for creating billion-dollar companies, and practical applications to help you increase your productivity. It's time to get down to brass tacks with a tactical guide to get the most out of ChatGPT—starting with how to prompt, or ask, it. Good prompting is the difference between AI that feels like magic and AI that feels mundane. Michael Taylor, who's writing a book about prompt engineering, explains exactly how to question ChatGPT to get the answers you want. We'll resume regular publishing on Monday with a new installment of Gareth Edwards's column, The Crazy Ones.—Kate Lee

---

ChatGPT’s output is the average of the internet. It has seen the best and worst of human work, from angsty teenage fan fiction to the collected works of Ernest Hemingway, and everything in between. But because it is the average, the default response you get is often undifferentiated and bland. ChatGPT is capable of doing almost anything—you just need to ask in the right way.Finding the right way to ask, or prompt, the model is known as prompt engineering. With the GPT-3 beta in early 2020, you had to hack the prompt to find the right combination of magic words or phrases to trick the model into giving you what you wanted.

As OpenAI released smarter, more sophisticated models like GPT-3.5 and GPT-4, many of these old tricks became unnecessary. As someone who freelances as a prompt engineer and created a popular course on the topic (we just passed 50,000 students!), many people ask me if prompt engineering will even be needed in the future when GPT-5 or GPT-6 comes out.

Sam Altman, the founder of OpenAI, certainly doesn’t think so. “I don’t think we’ll still be doing prompt engineering in five years,” he said in October 2022. OpenAI’s stated goal is to build AGI, or artificial general intelligence: a computer that performs at a human level across every task. When it reaches that point, we can ask the computer to do whatever we want in natural language, and it’ll be sophisticated enough to anticipate our needs.

Yet I don’t believe prompt engineering will disappear. It’s an important skill. Look around at your coworkers (or if you’re working from home, take a look at your Slack or Teams messages). Everybody you work with is, by definition, already at AGI—but they still need to be prompted.

You’re communicating a prompt every time you brief your designer, explain how to do something in Excel to an intern, or give a presentation to your manager. Every manager is a prompt engineer, using their communication and data analysis skills to align the team toward a common set of goals. Even your employment contract is a prompt: a standardized template of language designed to align your behavior with the commercial goals of the organization.

While we might not call it “prompt engineering” in five years, we’ll always need mechanisms to give direction to our AI—and human—coworkers. Those who get good at this set of skills will have an unfair advantage over those who don’t. Let’s dive into how to engineer prompts for text using five simple principles, as well as learn the basics of image generation.

The five pillars of prompting

As AI models get better, a consistent set of useful principles have emerged. It’s no coincidence that these principles are useful for both working with humans and AI. As these models approach human-level intelligence, the techniques that work for them will converge with what works for humans, too.

I first put these prompt engineering principles together in July 2022 and was relieved to see they mapped closely to OpenAI’s prompt engineering guide, which came out a year later. The principles are as follows: