You Ask, I Answer: Why Roles Work in Generative AI Prompts?

In today’s episode, we’re tackling the challenges of language models and how to get the best results. You’ll learn how to craft prompts that unlock more reliable and helpful responses. Discover the secret of “roles” and how it dramatically improves output quality. If you want less frustration and more value from AI tools, this is for you!

You Ask, I Answer: Why Roles Work in Generative AI Prompts?

Watch this video on YouTube.

Can’t see anything? Watch it on YouTube here.

Listen to the audio here:

Download the MP3 audio here.

Machine-Generated Transcript

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for watching the video.

In today’s episode, let’s tackle critics and undermining their ability to be critical of you.

This is especially true of the worst critic of all, our internal critic.

Have you ever ever posted something online, personally or professionally and just just gotten flamed for it? Just had your butt handed to you? Worse, was it was the criticism at least partially valid, ignoring, you know, all the typical invalid critical tactics like, you know, personal insults, slurs, etc.

Was there some truth to what was said? Wouldn’t it be great to reduce the likelihood of that happening of that unnerving sense of dread when you check your notifications after you post something where you’re getting negative feedback? For high stakes content, we will often or should get a second opinion.

We’ll ask a co worker or friend to give a read to something and it after doing so they will hopefully offer some constructive feedback.

But we can’t do that for every little thing.

I mean, I guess it depends on your friends, but sometimes things do blow up even with a second opinion.

So how do we further reduce the likelihood of this happening? How do we make ourselves feel a little more reassured? No surprise, we ask a machine, we ask a machine to behave like a critic, we ask it to take us to task on something.

This is especially valuable because we know ironically, that a machine is in some ways more trustworthy in some ways.

Here’s why.

It’s not our friend.

Right? A generative AI model is not our friend.

So it’s not going to hold back out of fear of hurting our feelings.

If we ask it to be critical and says, Wow, that sucked.

It’s because it actually sucked.

Right? It’s also not our enemy.

Right? So it’s it’s not like the human critics online, it’s not going to find fault with something, even if it’s spurious, just to make us feel bad.

While AI models do have biases, with careful prompting, we can have them give us some honest feedback that doesn’t have doesn’t have the emotional weight of human feedback.

Let’s, um, let’s look at how we might deploy this, how you might bring this to life.

First, we should define who the critic is that we want the machines to roleplay, we go ahead and pull up Google’s Gemini here.

And what I’m going to do is I’m going to say, we’re gonna create an artificial persona, I’m going to name mine Thomas, and you should use the largest language model you have access to.

Now, that might be Gemini, that might be chat GPT, that might be Claude doesn’t really matter as long as they’re one of the big models, I would recommend that it’d be one of the paid versions of these, because they’re, they’re the best at doing this sort of thing.

So I’m going to give it a character card.

So a character card is a concept of borrow from role playing, right, where you describe who the person’s going to be.

And we’re going to tune it to what we specifically want the the critic to look out for.

So let me show you my character card for Thomas.

So Thomas, the critic role, critical thinker, contrarian core characteristics, these are the personality insights, you can see here, there is these biases, I gave Thomas doubting Thomas here a lot of biases to look for, hey, look for these things in perception and memory, social cognition and influence, learning and decision making belief perception.

Now, this is a fairly long prompt, I wrote this because I specifically want to, I want to evaluate my own content based on these criteria, Thomas is going to be critical.

So I’m gonna go ahead and paste this and say, today, we’re going to do some critical analysis, you will be role playing the part of Thomas, a critic, here is Thomas’s character card.

Now, this is just what we just showed on screen, right? This is nothing new.

Gemini is going to think this through, it’s going to say, Okay, I’m ready to be behave like Thomas.

And this is very thoroughly designed because I want to be able to beat up my stuff pretty well, and highlight flaws in my thinking.

Christopher Penn: You should tune this to however you want it to be, however you want Thomas to respond, or whatever you choose to name your your AI.

So let’s take a LinkedIn post, I was thinking about writing, I was thinking about publishing, I wrote this LinkedIn post, I was like, I don’t know.

So here’s the post.

It’s all about, you know, we’re being clear about AI, it’s not a good thing that AI is going to take jobs, you know, what’s the solution, etc, etc.

I’m going to take this, I’m going to give this to Thomas.

Yes, please critique this LinkedIn post.

I’m gonna put in my post.

And now I’m going to do one other thing.

I want Thomas to look for five specific things.

And the five specific things I want Thomas to look for our list up to five things this piece gets mostly or entirely right, there may not be any.

That’s an important consideration in this prompt.

List up to five things this piece gets mostly or entirely wrong, and there may not be any list up to five things this author has overlooked, forgotten about or doesn’t know and they may not be any list of the five things that five ways people might criticize this piece, take it out of context or willfully misinterpret it, and there may not be any list up to five ways to improve the piece and there may not be any.

You’ll see that there may not be any a lot in this prompt.

And the reason for that is, if you don’t give it that out, it will come up with things that may not be real, right? It will hallucinate because you gave it a very specific thing saying give me five things for this.

So I say give me up to five things and if there aren’t any don’t make them up.

So that’s an important part of this.

So let’s go ahead and see.

Thomas is going to critique this and I’ve designed this prompt to be clear and bold in its feedback.

You do not have to do that.

So let’s see what Thomas has to say.

Things get post gets mostly entirely right.

Ethical concerns, socioeconomic impact, oligarchic control, the risk of suffering, empathy is key.

Things the post gets mostly or entirely wrong.

Difficult to identify is where the post is significantly incorrect.

Most of the arguments are based on potential consequences, not definitive wrong statements of fact.

Yeah.

Right.

So Thomas is not like, dummy.

Things the author overlooked the benefits of AI and the post narrowly focuses on the negatives has the enormous potential to improve healthcare, education, etc.

AI for labor augmentation, the post paints a purely replacement based picture of AI and jobs overlooks ways AI can be used to improve efficiency, ongoing AI ethics work, historical adaptability.

So it acknowledges, hey, these are the things that as an author, you should have done better on ways this thing could be misinterpreted or a critic, criticism is an anti technologies stance could be misconstrued as entirely anti AI, despite the author Fear mongering with strong emphasis on negatives, especially without balanced potential benefits, could be labeled as spreading fear and certainty of doubt.

Socialist leaning, some might interpret it as the desire for technology and profit controls advocating for heavily centralized economic control.

Lack of concrete proposals, yep, that one for sure.

Naive on oligarchy, while valid to raise concerns, calling it oligarchy could be deemed an oversimplification of real world power structures.

Ways to improve the post, acknowledge the positives, offer partial solutions, tone down the negativity, cite sources, and call to action.

So that’s a pretty balanced set of criticism, and there’s stuff in there that is 100% valid, which is why I didn’t post it on LinkedIn.

I had a feeling, I know what I wanted to say, but I couldn’t get enough of it out, so I just put it in the draft folder and parked it for later.

Now, I’ve got this analysis.

And is it correct? Yeah, it’s coherent.

It highlights the good and the bad, and it gives me some suggestions for how I might want to edit if I move forward with it.

Also, pretty clearly pointing out what the risks would be if I went forward with this and I didn’t make those changes.

And Thomas is pretty well said.

Yeah, this is kind of half-baked.

Here’s the thing.

You can customize this to however you want.

If you are uncomfortable receiving criticism or confrontation in a certain way, this is the best way to get.

And ahead of the problem, as a person, you give it to a machine.

Thomas isn’t human.

It’s a model.

It will literally forget everything that’s happened the moment I click new chat.

Thomas machine holds no grudges.

And by anticipating the very real ways that real people could take something and make something from it that I didn’t intend, I can protect that as best as I can.

So if you feel maybe insecure about a piece, maybe you don’t feel like.

It’s good enough.

Have a machine critique it.

Have a machine critique it and anticipate trouble.

Anticipate what could go wrong.

When the finished piece goes live, if you choose to take it live, the criticism it does receive will either be something that you know about, because you’ve evaluated it, or it has a higher probability of just being baseless.

A human hallucination that’s not rooted in the actual content itself.

Someone has an ax to grind with you, the person, and not you, your content.

But this technique is good if you have that sense of, I don’t know if this is good enough.

And I’m not saying that you feel that way all the time.

I’m saying you might feel that way about a piece.

Like, I feel this way about this particular piece.

It wasn’t strong.

It was not my best work.

And so Thomas Clued pointed it out.

This is not your best work.

Fix it up.

But it’s also super powerful if you have risk-averse stakeholders.

You can show them your customized version, your version of Thomas.

And how you’ve attempted to account for the most probable negative reactions.

Think of how many executive foot-in-mouth incidents we’ve had.

Imagine taking a Thomas to prepared remarks and saying, tell me the ways people are going to misinterpret this.

Tell me the ways, the things that I forgot.

Tell me the things that were insensitive.

You can proactively give your PR team a heads up like, hey, you’re going to publish something that Thomas said, was high risk and share with them the expected risks.

Like, hey, we’re pretty sure people are going to call this misogynistic.

We’re pretty sure people are going to call this whatever.

And obviously, if there are any decent PR team, they should be like, don’t publish it then.

But in some corporations and some power structures, you are given the materials, but you are not given the opportunity to change them.

But at least you can understand the expected risks up front.

If and when something does blow up, you’ve done as much as you can to plan for it in advance.

And if you advise your stakeholders that, hey, Thomas flag, this is potentially problematic.

When it does blow up, you have covered your butt, right? You have a record of the counsel that you’ve given that you said, hey, this is a bad idea.

Let’s not publish this.

Broadly, we’re using generative AI here as an editor, as a pair programmer, a partner, a discussion partner.

We’re not asking it to do the writing.

In fact, this technique is all about putting you, the human, and your human inspiration first, and then refining and editing with AI, not creating with AI.

We’re asking AI to evaluate our writing, to strengthen it, to make us better.

You could make a version of Thomas that specifically looks for biased or culturally insensitive text, a sensitivity reader, and warn you, you’re about to publish something that’s potentially offensive, right? If you write something that says, hey, this sentence here could be, interpret it as being cultural appropriation or culturally insensitive.

I was writing a post the other day, and one of them flagged me.

It’s like, hey, this word is potentially problematic.

Maybe change the word, and I did.

But you can use it for that purpose.

You can customize a version that looks for proprietary information, warns you, you’re about to give away some secrets.

Maybe it shouldn’t be, right? Maybe you put your prompt in your post, and it’s like, hey, you’re not supposed to be giving away prompts.

You could even…

If you have someone who is a pain in your butt, you could even model a version of this prompt after that specific critic, the person who gives you grief frequently, baseless or not, copy and paste their past criticism so you can anticipate what they’re likely to say for any future piece, and you can adapt it.

You can say, hey, I’m going to get ahead of this guy.

It’s almost always a guy.

I get ahead of this guy, and remove any possible afters for valid, critical criticism.

And of course, like I said, you can customize the outputs to match how you prefer to receive criticism, how you best process with minimal emotional harm.

My version of Thomas is blunt and to the point.

Dude, you screwed up, right? And here’s how.

That might not work for you.

That might be unpleasant for you.

So tailor your version of Thomas to how you process criticism best.

You could work like, hey, just a thought.

Maybe you don’t.

Do this as opposed to mine, which is that was wrong.

But customize it.

No matter how you customize this prompt, your Thomas has the potential to be a powerful partner to you to prevent outcomes you don’t want when you create and share content.

So take this idea and run with it.

In the blog post that goes with this video, I’ve put the entire copy of my Thomas character card.

So you can copy and paste it.

You’re welcome to.

It was created with AI, so there’s no copyright on it.

Play with it.

See how it goes.

That’s going to do it for today’s episode.

Thank you for tuning in.

I’ll talk to you on the next one.

If you enjoyed this video, please hit the like button.

Subscribe to my channel if you haven’t already.

And if you want to know when new videos are available, hit the bell button to be notified as soon as new content is live.

♪ ♪

You Ask, I Answer: Why Roles Work in Generative AI Prompts?

Machine-Generated Transcript

Comments

Leave a Reply Cancel reply

Pin It on Pinterest