Non-obvious prompting advice for writers

I’ve been using AI to help me write since 2020. After three years of bashing my head against my keyboard, I've figured out a few prompting techniques that help these models write and think better.

I haven't read much about these techniques elsewhere, so I figured I'd share them here.

The key thing to grasp is that language models are like smart interns who lack common sense. They can be so bright, yet so dumb. If you get good at giving instructions, you'll tip the balance towards "bright."

Here's some non-obvious prompting advice that may help you get more from your most junior employee:

• Use words like "insightful", "surprising," "interesting," or "useful" in your prompts. Language models have a nasty habit (really, it's a feature) to drift towards the mean. They're trained to give you the most likely answer to a question. That's fine for most tasks, but great writing, by definition, is not average. Thankfully, you can just ask language models to avoid what's predictable. For example, you can get better results if you say something like "write an About Me page for my lawn care company that's interesting and insightful" instead of the same prompt without "interesting and insightful." This seems to break the more natural drift towards what's predictable.
• Have the AI ask *you* questions. Language models can only be trained on what's already been said. But great writing often says something new, or at least says something old in a new way. Chances are, you have an idea or experience that hasn't been expressed before. So, tell the AI: "I want to write about x. Ask me some questions to get the writing process started."
• Say who your audience is and what they care about. Great writing often answers a specific question that a specific type of person has been struggling with. Instead of telling the AI to write tips for business owners, get more specific. Ask it for tips that small law firms can use to acquire more customers. You'll often get more interesting ideas – or you'll get more actionable examples you can use to illustrate an idea. Or both!
• State the obvious, clearly and often. It might seem obvious to you that a landing page for a dentist shouldn't include a bunch of emojis. But that's not always obvious to our AI intern, since many of the landing pages they've seen do have emojis or annoying, over-the-top language. State these obvious things directly in your instructions. For example, "write me a landing page for x and don't use cheesy verbs like *unleash*" or "...make sure to include a call to action to schedule an appointment." And so on.
• Include examples of what "good" looks like. Where instructions fail, provide examples. You'll be surprised at how far this can get you. You could tell the AI to write in "a straightforward, professional tone," but the output can still feel sterile. Instead, try something like this: "write in a straightforward, professional tone. For example, say 'we grew 15% last month' instead of 'we saw our sales increase by fifteen percentage points.'" Better yet, include an example of a finished piece of work you'd like to emulate alongside your instructions.

If I had to boil all of this advice into a single sound bite, it'd be this: "tell the AI the same stuff you'd tell an intern doing this task for the first time."

If you're still getting mixed results, OpenAI published a more advanced prompt engineering guide you may find helpful.

AGI is coming, but for now we live in the Age of Interns. And since they can't fetch us coffee (aka the ultimate writing aid), we might as well train them to collaborate well.

Happy writing! – er, prompting.