Transcription
I have a confession. Even though I build complex AI applications day in and day out, I've not written one single prompt from scratch in over a year. While everyone's posting on social media their 500-word miracle prompts or selling prompt engineering courses, I've been quietly using a simple trick that improves my prompts every single time.
And here's the thing, most of us don't need to become prompt engineers. Instead, we just need to take 15 minutes to learn how to utilize AI to improve our prompts for us. Let me show you my exact workflow that's replaced every prompt template that I've ever had in the past, and it'll probably do the same for you. Let's dive into it.
So, there are three different approaches I'm going to show you of how you can improve your prompts. Two of which are directly related to talking to AI, and one is dedicated to a tool that automatically improves your prompts. We'll start with AI, then we'll go to the tool.
So, with AI, there are two tactics I like to use. One is a simple method, and the other one is a more complex method. The way you would know which to use is depending on how high-stakes the task is and how vague the idea is in your head. If it's a vague idea and it's high-stakes, I recommend taking the complex path. If it's a simple task or a simple idea that you have some understanding of, you take the simple path.
Most of the time, I'd recommend starting with simple. If simple doesn't work out and you feel like you're not getting what you need, then you can go to the complex side. So, 90% of the time you're using simple. Sometimes you go to complex.
So, what is the difference between these? Well, the simple one is you're basically asking AI to write the prompt for you. And then the complex, you're doing an AI interview. We'll start with simple.
Now, here's the prompt that I recommend using for the simple approach. And this is the base prompt you'd give to an AI. In this case, we can give it a ChatGPT 5. In the prompt, we're stating, "I'd like you to create a system prompt for Chat 5, but I'm not sure where to start. The goal of that prompt would be to have the AI rewrite my email and newsletters to draft to improve the clarity and simplicity. Can you research best practices for prompting this specific model as of today's date? Once you have those best practices, then write a system prompt for me that meets that need."
Now, the things I'll call out here is first, we're being explicit about the specific model that we're using. So, it knows when it researches best practices for prompting this model, which is called out here. It's going to research best practices for this given model. But also, we're going to clarify even further that we're going to specify today's date. So, I just put this in brackets for now, but you can put in the actual date when you do the prompt. And then it's going to find best practices as of that date and or close to that date. So, you're getting the best practices that are modern and relevant today.
So, we're basically asking AI to research and write the prompt for us. And I want to show you some examples of what this looks like. So, here I have that exact prompt. So, I put the prompt in, I put the date in, and then I had it write out and it researched for me. So, it thought for a minute and a half, and during the re—during the thinking, you can see it did research. So, it went to a series of websites that had best practices for prompting. It consolidated all of its dots and then if I go through here, it kind of gives you some of the best practices. And then down here, it gives you the base prompt. So, here's our improved prompt or a prompt that we can start from that takes a lot of the best practices and puts it into a prompt for us.
So, this is a great starting point, but sometimes, as you can see, AI can be quite wordy, especially if you ask it in this way. So, if you're asking it in this way, I'd recommend if it's too wordy or the—the prompt is a bit too overwhelming for you, you can copy this out and put it into a tool, which we'll talk about later, which would optimize it further to cut out some of the fluff and get to the point.
So, that's our first tactic. And the meta-lesson I want you to take away from this is like I stated, the AI can be quite verbose. So, if you're using Claude, the prompt will be kind of big. If you're using GPT, sometimes it would be big as well. And if the base prompt it gives you is too big, I recommend cutting it down because less is more. When you have a starting prompt, you want to start small with any prompt, no matter what the task is. And you want to iteratively add complexity and different nuances and edge cases to the prompt over time. And the reason we're taking this approach is we need to truly understand what each part of the prompt does to the AI for the use case we care about. Because often times what you'll see is the end state. So, if somebody shares something on social media or gives you a prompt and it's massive and you're like, "Oh, this is super complicated. It must be amazing." The only reason it's complicated now is what they've done is they've iterated over time. So, they started with something small and they built complexity onto it so the AI performed in a specific way for a specific task or ask. So, meta-lesson: start small, always iterate over time so you can understand how it's impacting the AI over time.
Oh, hey, this video is brought to you by me. So, there's two things I want to share. First, below is a link to a free 30-day AI Insight Series. It's completely free. You'll get 30 insights in your inbox over 30 days of how you can apply AI to your business and your work. Second thing is, if you'd like to work with me, below are a variety of offerings that I have to see if there's a good fit between the two of us. That being said, let's dive back into it.
So, that's our first tactic. Our next tactic, the more complex approach, is going to be the reverse AI interview. And remember, we're only going to do this for high-stakes tasks that we're kind of vague on the need for. And here's the base prompt we'll start with.
So, "I want you to interview me asking one question at a time. Each answer I provide, you should inform the next question you ask." This is really important. So, I'm going to underline this and talk about it before we go anywhere else. We need to make sure that it's asking one question at a time and each answer informs the next question. Because if you just ask it to interview you, it's going to give you like 50 questions at once. And it's not going to tailor and guide the conversation in the direction that's relevant to your context, your scenario, and what you care about. If you structure it this way and you start the prompt off that way, it's going to better direct the conversation to the things you care about. And then I give some intent of what I need. So, the intent of you kind of interviewing me is for the end system prompt to be optimized for this given model. Again, we're specifying the model we're going to improve the prompt for. My goal is to create a system prompt for an AI that can effectively rewrite all my emails and newsletters in a more clear and simple way."
Since I have so many videos on this on my channel already, you can check out many of those and other videos. What I'll do is I'll show you kind of an intuitive understanding of what's going to happen here. So, it's going to ask you back and forth different questions. And here's an example of what it could look like. So, an example, the first question it asks is, "What types of emails are you writing?" And I give it some context of the emails and the newsletters. Then it says, "What specific clarity issues are you having with your current writing style?" And I can give some nuance to that. And it keeps asking you questions. And this will last for probably 20 to 25 back and forth. If it goes too long, you can just stop it. And the stopping prompt I usually give it is saying, basically, after we've done this, after we've finished this interview, can you now please create the system prompt based off of the context I've given previously?
It's a very straightforward approach. It takes a little bit more effort, but at the end of the—at the end of the day, there's two benefits. One is you've taken what's implicit in your head and you've made it more explicit in the world. So, now you understand what you want more, which is great. But also, you've given the AI enough context to then create a prompt that's hyper-tailored to your need and you're likely going to get a better AI output when you put this into an AI for the task you care about.
So, those are the two AI approaches. We have the simple ask method of asking the AI. Then we have the AI interview. Now, we're going to go on to the tools.
And what blows my mind is that a lot of people don't realize that two of the leading research labs in AI have prompt optimizers already. So, in both OpenAI and Anthropic, and their console tools, they have the ability to take in a basic prompt and then spit back out to you an improved prompt based off the best practices for prompting their models. And what's even crazier is that they're both basically free. So, OpenAI is free with an account, and then Anthropic, you can add some credits through their API setup, and it costs maybe like a dollar or $2, and that'll last you months.
So, we'll start with OpenAI. So, there's two different ways you can get to this uh specific optimizer. So, you can take the indirect path, and I'll show you both ways. This is kind of going through the platform, and this is the direct link to it. Both of which you need to be logged in for, obviously. But like I said, the benefits are is that it's free, that this optimizer is dedicated to optimizing for GPT-5 specifically, which is important, and also it shows you the rationale of what's changed, how it's changed, so you can learn along the way.
So, let me first show you how to navigate to these, and then we'll show you some examples. So, I'm going to go to this link here, and you're first going to land on something like this. Once you've landed on this, you'll probably have to log in if you're not logged in already. Once you're logged in, you'll go to the dashboard here, and you'll want to go to the chat symbol in the upper left-hand corner. Once you've gone to chat, you'll go to the create in the center. After you've selected create, you're going to have this inbox here for the developer message. This is where you're going to put your base prompt. So, that's one way of getting to this. Another way of getting to this is simply going to this direct URL. It'll ask you to log in. Obviously, once you've logged in, it's going to take you directly into the developer message here, and you can just put your base prompt here as well.
So, now that we know how to navigate there, let's take this sample prompt and improve it. But first, let's read the prompt. So, the prompt states that, "I want it to rewrite my email and newsletter drafts. I wanted to make sure that they're written in my writing style. I wanted to aim for simplicity and clarity. And then what I'm going to tell it what it'll receive. So, it's going to receive an email draft for either an email or a newsletter. And what I expect to give back to me is going to be a new and improved version." So, that's the basic prompt I'm going to give it.
So, let's copy this out and put it into our optimizer and see what happens. So, I'm going to paste it both into this section. So, once you've pasted it here, you can select optimize and it's going to optimize the prompt. And then if you select here and you've pasted it into this section for the developer message, you're going to go up to here to optimize in the upper right-hand corner. You're going to select optimize and it's going to do the same thing. So, once it's done this, there should be both optimizing. Actually, for this one, you're gonna have to select optimize once more. Now it's optimizing. So, they're both working and it usually takes between like 10 and 30 seconds.
And here's our output. So, you can see first off, it's a lot longer. So, that's—that's one obvious area where it's improved because length does equate to quality if it's not overwhelming length. So, obviously, if we're more than just two sentences and we have some more structure to the prompt, it's going to likely give us a better output. But we don't want to go too long, like I stated previously. Another thing is you'll see we have a series of hashtags. So, each of these sections are separated by a hashtag now. So, that gives it more of a uh separator between each section. So, the AI knows what each section represents. These are called delimiters. If you're interested, then you can see these little comments things on the right-hand side. So, each of these little comment bubbles, if you select this, it's going to tell you the rationale as to why it added this section. And here you can see it stated that it added this section to ensure that the assistant itself has a clear mission and a clear, consistent starting point as well. And I won't review everything here, but I'll call out a few things that I think are interesting for you to see and the improvements associated because I think it's important for you to read through this so you do learn a little bit about prompting, but note that you can just use these tools and don't have to learn it yourself.
So, the first thing here is it's giving itself a workflow checklist. So, what this is going to do is this checklist is going to be internal to the AI's—the AI's kind of brain, and it's not going to be explicit to you. And it's going to go through this checklist to ensure that it's writing in your style based off the email you've given it or the emails—the email newsletter draft you've given it. And it's going to ensure that it's focusing on clarity and simplicity. You can think of it as like a simple grading mechanism. Another really interesting addition that it's added here in this point here is at the very end you can see that if the validation fails, self-correct. So, what's happening here is it's saying that if you've written this once and the first draft doesn't meet the expectations of the initial checklist you've set previously that we just talked about, then I want you to redo it over and over and over again until it's good enough. And down here, it sets a verbosity metric saying that it wants it to be concise for verbosity. It gives it an output format and also a stopping condition. So, this stopping condition is at the end of most of the prompts that this tool will give you to ensure that the AI knows when to stop, when it's finished, and when it can give you back your output.
So, this is just one simple example of how this optimizer works and how it can dramatically improve a simple prompt that we started with here. Oh, not here. Sorry. Here we started with a simple prompt here. It was vague and basic, but now it's something that's very concise, has structure, and it's going to improve the output that you're going to get from the AI.
Okay, so now that we talked about OpenAI, let's talk about Anthropic. So, with Anthropic, this is the URL that you would go to. So, let's copy and paste this into the URL, and you're going to get something like this. So, you have to log in first off. Once you've logged in, you may need to add your credit card to give it maybe like $1 to $5, and that will last you likely a month to two months for prompt optimizations depending on how much you use it. And what you want to do here is you want to click the prompt optimizer here. So, generate prompt, and it's going to give you a pop-up. And we're going to take that same example that we ran OpenAI through. So, we'll copy this example over to that URL. And then down here, we have a few options. So, these are just starter prompts. You can ignore that for now. And then here it has a checkbox, says, "This prompt would be used for models that have thinking." So, we're going to say yes because any model that I use often with tasks, it has a thinking element to it. So, you can just check that box and you say generate. And then this one responds a lot faster than the OpenAI one. And it'll come back with a prompt that's optimized specifically for their models. So, we just optimized for GPT. Now, we're optimizing for Claude from Anthropic.
So, if I scroll up here, you can see a few things that are different. Right—right off, you can see there's this tag here for XML. And XML is just a form of structuring text that makes it easier for the AI to know what separate sections of the um prompt are for, similar to what OpenAI did with the hashtags that I showed you. So, this XML here is basically calling out the input that that's going to be put into the prompt. So, if you're using this as a system prompt inside of a Claude project, a GPT project, a custom GPT, something along those lines, when you add this, it'll just take what you've given it in the developer message and then insert it in here for the system prompt as the AI processes this information.
Let's actually go through this and see some of the differences between OpenAI and Anthropic when it optimizes for different types of models. So, let me zoom in here so you can actually see what we're talking about. Cool. And for simplicity and clarity. So, for simplicity, you can see that it's calling out the importance of shorter sentences, everyday language. For clarity, it talks about logical flow, and a variety of other things. And down here at the bottom, you can see that it's setting a reminder and a final response, saying that the final response should just be the response itself of the draft. There shouldn't be any additional explanations, no comparison between the original, no commentary, etc. So, it's saying, "I just want the output of that draft and nothing else." So, this is an improved prompt that takes a lot of the best practices for Anthropic's models and embeds into the prompt you've given it.
And those are the simple tactics. So, the true meta-skill that I want you to take away from this video and the importance of knowing how to use AI is knowing to ask the right question at the right time. Because if we know to ask AI to help us with improving the prompts and we know which tools to go to to do so, it's going to drastically improve our prompts without having to be a quote-unquote prompt engineer and have to memorize all these different frameworks and tactics for these models when we're just using these models to apply them to our businesses and our work. And we want to get the most out of these—these different models we have access to. And that's it.
So, if you enjoyed this, please share it with your friends. And also, as I mentioned previously, two things. One, below is a 30-day AI Insight Series. Completely free, 30 insights in your inbox of how you can apply AI to your business and your work. Also, if you'd like to work with me, there are links below. You can see if there's a good fit between the two of us. And with that being said, you should watch the next video that the YouTube gods think that's most suitable for you around.