This content is locked. Please login or become a member.
Prompting AI is super weird, weirder than you could possibly imagine. There is a bunch of research, for example, that suggests that emotionally manipulating the AI gets you better answers. If you tell it my job depends on it, you get nine percent more accurate math results than if you don’t. If you tell it’s really good at something, often it becomes really good. There is some evidence that it performs less well in December because it knows about winter break. And if you’re convinced that it’s May, you get better results. We don’t know why.
Use a chain-of-thought approach
There are a couple of techniques that work very, very well when doing prompts. One of those is called chain-of-thought prompting. And there’s actually a lot of definitions of this, but the most simple way is to give the AI step-by-step instructions on what you want it to do. First, create an outline of this document. Second, fill in the outline with more details. Third, go back and check the results to make sure they’re good. Fourth, produce the final document for me.
The reason why this works is not 100% clear, but it seems to be that the AI doesn’t think. There’s no internal monologue. When we ask ourselves a question, stuff happens in our brain, then we sort of think things through. The AI doesn’t think things through. It thinks things through by creating the next word. If I give it a chance to create some words already and to outline its presented logic, it then will follow that logic forward and produce better results.
The other advantage of a chain-of-thought approach from a human perspective is you can actually then look at all of these points that the AI is making and see where it went wrong and see where you could add something to it. So it helps you become the human in the loop because you can better diagnose problems, understand where the AI’s logic is faulty, and you can insert your own.
Iterate together
There’s a few reasons you might not wanna copy, and paste the first things you see with AI. It’s not just because there’s gonna be errors and you need to check them, but also the worst thing you could do in producing sort of AI work is to just ask for something and then use it. The way these systems work best is actually interactively. I would actually go back and say, oh, can you punch this up a little bit more? Or can you break this down into parts for me? Can you make the second paragraph more interesting? Can you stop using the word rich tapestry, which is an AI favorite?
Choose when to bring in AI
Whenever you consult anyone about anything, you immediately go through the mental process of anchoring on that idea or concept. If you’re doing idea generation, there’s a reason we tell people not to do brainstorming anymore in a group to start off with because as soon as somebody suggests an idea, you mentally anchor on that idea. So the same danger happens with AI. The AI does a piece of writing for you. You might say I’m gonna edit it, but it’s going to be very different than if you started on your own. You’re gonna have to make choices. When do I wanna be original and start with my own work and get AI help afterwards? And when do I want to do the reverse where I’m willing to get AI help from the beginning?