Generative AI Leadership

If you’ve been working with generative AI for any length of time...

I live in Dover Delaware and own a 24 foot bow rider boat. I need to upgrade to a cabin cruiser, like ones I have seen in the Michigan upper peninsula when I was visiting there. I need to find three reputable boat dealers within an hour of Dover. I’m just trying to find out what the availability is. I don’t want to take any next steps, I’m just looking for information on Boat dealers in the Dover Delaware area that carry cabin cruisers for under $100,000.

Chances are you’ve seen this pattern:

You prompt the AI, and it responds. You prompt it again, either thanking it for providing an answer, or giving it a little bit more information to direct it.

And the AI is off to the races… It comes up with a whole bunch of different ideas, different choice, points, potentially veering into a very different direction than the one you started in.

It can be very time-consuming to get AI back on track, after it veers. And I’m not always successful. Sometimes I need to just completely back out of a chat and start fresh with a new prompt, because we’ve just gone in all kinds of different directions.

The issue is not with the AI, the issue is with me. I know as well as anyone that this technology is generative. It will take what I give it, Restate it back to me with all sorts of elaboration, and then it will look for a new direction to go. This is not a mystery to me. I know it does that. And yet, time and time again, I forget that I need to lead it in the direction it should go, let it decide for itself and we end up shopping online for a boat in the Michigan Upper Peninsula.

Short and sweet, if you want to prevent AI from improvising in terms of what happens next, make sure you tell it where you want it to go next next. Finish your prompt with a statement like, “I don’t want you to come up with additional ideas for me, I just want to do some in-depth analysis, and I’m providing this information to you for context. Please wait until I give you the go ahead before you start generating next steps.“

Or, you can say, “I don’t expect a reply from you, please consider what I’m about to tell you for future reference.“

Or you can also say, “I need you to come up with three different options for what I should do with this information and outline the steps I would need to take for each.“

The point is, you’re getting ahead of the AI and giving it needed context and direction for how it can satisfy user needs in a “pass condition“. And when AI knows it has been successful, or can be reasonably sure, it uses less cycles trying to figure out what it did wrong and how it should do things differently the next time.

As an example, consider a situation where someone who enjoys boating in the Dover, DE area, wants to upgrade their bowrider to a cabin cruiser like once they have seen on the Great Lakes in the Michigan Upper Peninsula. They haven’t decided yet what they want to do, they just want to find out if there are any reputable boat dealers in their area that carry cabin cruisers in their price range, under $100,000.

Here are two ways that you can prompt, the first one that doesn’t lead the AI, the second one that does.

Prompt #1

I really need a new boat

Prompt #2

I live in Dover Delaware and own a 24 foot bow rider boat. I need to upgrade to a cabin cruiser, like ones I have seen in the Michigan upper peninsula when I was visiting there. I need to find three reputable boat dealers within an hour of Dover. I’m just trying to find out what the availability is. I don’t want to take any next steps, I’m just looking for information on Boat dealers in the Dover Delaware area that carry cabin cruisers for under $100,000.

Here's how each of them shook out:

By A) giving more context up front, and then guiding the AI on what you want it to do, it's possible to not only target the AI's focus, but also keep it from wandering.

With Prompt #1, I had to keep coming back and catching up with the AI to steer it in a different direction. It was doing its best to give me what it predicted I wanted, but it just didn't have enough context or detail up front to do it simply and cleanly. It confused my mention of the Michigan Upper Peninsula with an indicator that that's where I was going to be boating, and then it looked at my local IP address information (which could be wrong, if you're on a VPN) and gave me a lot more options than were appropriate. I had to keep refining the direction, as well as tell it when it got things wrong.

Also, it was a whole lot more work to get results back, because I had to keep thinking through all the choices it gave me.

With Prompt #2, I did more up-front work, getting clear on what I wanted, as well as what I did not want (next steps, etc.)

So, rather than having to "ride herd" on the AI to get what I wanted, I simply told it up front what it should do, and I think everyone was happier. I know I was.

If you're a paid subscriber, you can see how it all played out in the transcripts below: