Mind Readings: Let Generative AI Write Its Own Prompts

In today's episode, let's talk about prompt writing. Stop writing prompts for language models. Okay, that's a bit facetious. Stop trying to do it all by yourself. I was talking to a friend the other day about a prompt that they were working on.

It was just not delivering results. It was not getting good results. They're used to writing prompts for open AI smalls for GPT 4.0 and stuff like that, but the model they were using was a very different model. They were using Mistral Nemo. Prompts that work well in one system do not necessarily work well in another.

Especially if there's a big difference in the underlying model. Even among foundation models, the frontier models, the prompts and the internal model architecture differs. Gemini had, for example, has very different internal censorship to Chat GPT. You can't talk about politics with Gemini at all. It just gives you a flat out refusal.

There's some things ChatGPT won't talk to you about the Gemini will. Um Claude has a very different style of coding than Gemini. Claude is very different style of coding than ChatGPT. The two can actually learn from each other. You can pit them against each other.

But what that means is you can't just copy and paste a prompt to another model and expect the same results or give it a prompt and expect it to work out of the gate the first time around any more than you can take a librarian from the New York Public Library, put them in the Boston Public Library and expect them to know where everything is immediately. That's that's not how that works. So what do you do? How do you do this? Ask the model to write a prompt for you, to rewrite a prompt for you.

You tell it something, you give it something, and you say, hey, please rewrite this in the way that's gonna accomplish something that you know the the accomplish something to the best possible way. So let's do this. Let's do a very quick example here. I'm gonna go into Google's Gemini here, and I'm gonna say first, let's let's build ourselves a prompt that maybe isn't great. Say, today, let's write some let's write a sonnet in Shakespeare style.

I want the sonnet to be about the difficulties of changing the oil on my Toyota Prius. It should have the flavor and language of a Shakespearean sonnet. Right, the sonnet. So this is a very straightforward. This is a not a great prompt to begin.

Let's be clear about that. If I go and I put this into Gemini, I get one result. Let's go ahead and fire up chat GPT and do the exact same prompt. And this is okay, but not great. Suppose I wanted to improve this.

Well, how would I do that? I would take my prompt that I wrote that wasn't optimized. I'm going to put some spaces around it. And now I'm going to say, please rewrite this prompt to be as effective as possible in Chat GPT. The purpose of this prompt is to build a compelling sonnet in Shakespearean style.

Rewrite the prompt to accomplish this purpose. Write the prompt in the second person imperative. Now, what this will do is in chat GPT. This will gener they will have Chat GPT build the prompt the way it thinks you should do it for this system. So there's what ChatGPT came up with, right?

It said compose a sonnet in the style of Shakespeare, capturing the essence of his language and tone. The subject of the sonnet should be the challenges of changing the oil under Showday Prius. Infuse the sonnet with the intricate language, rhyme scheme, and structure for Shakespearean sonnet, making sure to evoke both the humor and frustration inherent in the task. Write the sonnet as though Shakespeare himself were reflecting on this modern dilemma. That's a much more thorough prompt than my original, right?

Let's go back to Gemini. I'm gonna paste in the same instructions here. Let's see what Gemini's internal architecture thinks. Instead, it tried to just go ahead and write the actual sonnet instead of following the directions. Let me try this in AI Studio.

We should use Gemini Pro, Zero Zero 2. Try following directions. Craft a Shakespearean sonnet detailing the tribulations of a previous oil change. Deliver the completed sonnet. That is a very different prompt than chat GPT.

Let's do Anthropics Claude. Let's see what Anthropix Claude said does to generate the prompt. See if it follows directions. Here's a rewritten prompt in the second person imperative. Channel the bar's eloquence, wit, and meter to elevate this mundane task.

Employ aambic pentameter and the rhymes, this rhyme scheme. Infuse diverse archaic diction, rich imagery, and poetic devices. Contrast the modern prehistor of Elizabethan sensibilities for humorous effect. Dramatize the struggle between twixt man and machine as one might face a tempest, and so on and so forth. You see what's happened here.

We've taken the same prompt that everyone would have written, and we've asked the systems to write it in their own language in the way that would work best for evoking and accomplishing this task. Look how different every system operates. Chat GPT is pretty compact, Gemini's is super compact, and Claude's is fairly verbose. Now, if you want to kick it up a notch, you could even use the Trust Insights Pair framework to preload the model with best practices about writing sonnets too. But you can see the internal model architecture so different from model to model that what we think is a simple prompt, if we ask the tool to write the prompt for us or to rewrite it, trying to understand our intent, we're gonna get very different results.

That's why, for example, in my prompt library that I have that maintain for myself, I specify which model a prompt is for. I don't want to use a Gemini prompt in Chat GPT because they have very different internal architectures. So if you follow this process, you will get optimum performance from the language models that you're working with, because the model itself has told you best how it works. Model has told you here's how I want you to prompt me. And by using that meta prompting framework, you will get great results from these things.

That's going to do it for this episode. Thanks for tuning in. 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.