LLM prompt engineering is easy. The trick is remembering when to prompt at all.
Why is everyone trying to sell you on "99 Guaranteed Magical ChatGPT Prompts Without Which You Will Definitely Die"?
Since ChatGPT and other language models came out, a lot of people seem to be marketing courses and books about how to perfectly engineer prompts to get the best results from these tools, from “A Treasury of Talismanic Prompts for Summoning ChatGPT's Deepest, Darkest, Dankmemest Self” to “AI Giants Conspired to Hide These 99 Existential LLM Hacks From Ordinary Folk.”
But the truth is, I don't need any special words or formulae to talk to an LLM like ChatGPT. All I need is the skill of describing what I want in normal language.1
I don't need any special words or formulae to talk to an LLM like ChatGPT. All I need is the skill of describing what I want in normal language.
The hard part of using LLMs effectively isn't writing the perfect prompt. It's grasping all the different things they can do for me, and remembering to delegate accordingly.
Many knowledge workers aren't used to delegating complex, open-ended tasks. But LLMs aren't just digital interns, good for repetitive work and routines. They can do many things that most interns can’t do. They can do some things that probably no one you’ve ever worked with can do. And they can do things quickly and cheaply that we sometimes do ask of our colleagues or friends, but only when the stakes are high because it’s time-consuming, expensive, or difficult.
For example, some things I routinely ask ChatGPT or other LLMs to do:
Reformat the same content ad nauseum, in as many different ways as I can imagine until I decide what looks right or best matches my needs. A human would be so irritated!
Rewrite text in an entirely different style or tone on demand, from gentle to academic to sarcastic. An LLM can emulate essentially any voice or genre. For me as a human writer, this kind of flexibility is nearly impossible.
Explain almost any concept or topic at an expert level, drawing on a broad, deep knowledge of essentially any field. While it may hallucinate individual facts (which you need to check for vigilantly!), an advanced LLM has a grasp of an extraordinarily broad range of concepts and subjects that no single person can match. It can break down and teach complex topics at length without tiring or getting annoyed.
Redo a task from scratch, multiple times over, if my initial instructions weren't quite right. If I’m delegating to a person, I need to be careful to give them everything they need to know up front, or risk wasting their time and frustrating them. But an LLM doesn't get irritated at having to start over—it simply begins again with the new information. In fact, ironically, this fact means that prompt engineering is LESS important than with LLMs than with humans. If your first prompt didn’t get you what you wanted, no problem! All you have to do is name what needs to be different, and ask it to give you a new version. Sometimes I’ll do this 10 times in a row!
Emulate another person. This is something that people DO ask other people to do. A candidate prepping for a debate will bring in someone who can do a decent job of emulating their opponent. If I’m going to ask for a raise, I’ll practice the conversation with my coach or a friend. But we generally only do this when the stakes are high: Most people aren’t very good at it, plus it’s time-consuming and cognitively demanding. LLMs, meanwhile, are tireless and can do a pretty good job of emulating a broad range of people. Ask an LLM to act as your therapist. Ask it to pretend to be a particular profile of voter and give you feedback on policy ideas. Ask it to interview you for a job and give you feedback on your responses. The possibilities are endless!
Analyze how persuasive or compelling a piece of writing is, and provide concrete feedback on how to improve it. For most humans, a deep dive analysis of the rhetoric and psychology of a long or complex piece of writing is very time-consuming. And frankly a lot of humans aren’t good at giving this kind of feedback even if they have time. In just seconds, an LLM can evaluate the logic, emotion, style, and more—and tell me how to strengthen the overall impact.
Research and summarize multiple perspectives on a complex issue. For a human, gathering and synthesizing information from many sources with different viewpoints is extremely labor-intensive. An LLM can review vast amounts of material on any issue, then boil it down into a balanced, high-level summary.
The key is recognizing that an advanced LLM has some comparative advantages over humans that extend well beyond routine tasks. LLMs may lack human judgment and sometimes get facts wrong, but they also have superhuman flexibility, breadth of knowledge, and patience for repetition.
LLMs may lack human judgment and sometimes get facts wrong, but they also have superhuman flexibility, breadth of knowledge, and patience for repetition.
So back to my hook: All those "prompt engineering" tutorials. The authors of these want us to believe is that there are magic words to unlock an LLM's potential, because they can sell you a list of those and call it a day. But natural language is all it takes, as long as we know what's possible. The value in lists of example prompts lies not in the specific wording, but in the list of use cases that jog me to remember and think of ways in which I can use LLMs.
The hard part of using LLMs isn’t crafting the perfect prompt. It’s building the habit to delegate tasks you never would have delegated before.
But I guess that’s a lot harder to sell than “The Lost Scrolls of ChatGPT: Ancient Wisdom to Wrest Your Heart's Deepest Desires From This Digital Djinn”2
To be clear, this isn’t (yet) true for all generative AI systems. For example, prompt engineering for Midjourney and other image generators is a genuine skill, and people with a lot of practice at it will generate outputs much closer to what they want much faster than those who haven’t practiced/studied up. But a key distinction is that the core feature of Midjourney is generating high-quality imagery, not understanding natural language. Understanding language to some extent is a necessary secondary feature — but it’s not where all the skill of the Midjourney engineering team is being directed. Whereas, by definition, understanding natural language is the core feature of an advanced LLM.
Note: I came up with these funny titles by, you guessed it, giving an example to an LLM and asking it to brainstorm a bunch more.