Most businesses know they could be getting more from AI, they just don't know where to start. So I like to share one common example.
Yesterday, I was on a discovery call with a business that is interested in hiring me to help them with their SEO and AI optimization. Naturally, they wanted to know what I would be doing to help them get more traffic from search and AI results. And I explained that about half of the work we'd be doing would be content creation.
They immediately jumped in and said, we actually don't need help with content creation. We have a team of writers. And I said, oh, that's great.
Are these writers experts in your industry? And they said, yes, we've trained these writers to understand everything that's important to know about our industry. We can't use AI to write the content because it constantly gets things wrong about our industry.
Now, first let me say I think there still is a place for prolific expert writers. If you have extensive firsthand experience or are truly a prolific writer, then AI cannot replace you. However, that was not the case with this business.
These writers had merely been trained on what was important to understand about this business. And the business was under the common misunderstanding that AI could not be similarly trained.
And this is the common issue I see over and over again. Businesses are willing to spend weeks or months training new hires before they're able to produce reliable output, but often not willing to spend more than 15 minutes with ChatGPT or Claude. And to be fair to them, I think it's just because they don't realise it's possible or they don't know how to get started.
So here are the basics. You need to set up some kind of project. They're called custom GPTs in ChatGPT, they're called gems in Gemini, or in my favourite, Claude, they're just called projects.
Projects have two parts to them: the instructions and the knowledge. The knowledge, you're going to place everything that is important for this project to be able to reference. You wanna document with all the basic information about your company.
You wanna document about all your products and services. You wanna document about your ideal customer profile. And then in the instructions, you want to give very clear instructions about how to do the task, and you also want to explain all the documents it has in its knowledge and, most importantly, specific instructions on when to use each document.
Think of setting up your project like the onboarding process of a new hire. You're not yet training it on specific tasks, you're just giving it the basic information it's going to need to start learning those tasks. For each specific task, you're gonna need a prompt template.
So for example, you're gonna have one prompt template for writing blog posts, one template for creating web pages, one template for writing social media posts. And these prompts are gonna be quite long. They're not gonna be a few sentences.
Imagine if you're training a new hire on how to write blog posts for your company. You wouldn't just give them a few sentences and send them on their way, right? You need to provide very clear instructions on what they need to do and what success looks like.
So here's the structure of each of my prompt templates. First, to find the role. This is gonna be something like you are an expert content writer for this brand.
Next, to find the goal. This is gonna be something like your goal is to produce an optimized blog post for a given search term. Then you're gonna define the input.
This might be the target keyword or the topic, along with any information you want it to reference while it's writing this piece of content. Next, you'll define the process. These should be very specific step by step instructions you want to take as it performs this test.
Next are the guidelines. These are any rules you want it to follow as it goes through the process. Again, be very specific.
Next are examples. You want to give it both positive and negative examples of what success looks like. You don't need to create the negative examples right away.
As you get bad outputs from the prompt, you can include those as negatives. Next, to find the format that you wanted to output the task in. And then finally, you're going to repeat the most important instructions.
And the reason you do this is because with long prompts, large language models often forget things you've said. And they pay special attention to the beginning and the end of each prompt. So including the important instructions at the end makes it more likely they'll adhere to them.
So with all those parts in place, you have the first draft of a prompt. The key to making this work is iteration. Every time you get a bad output, include the bad output as a negative example and update the important instructions.
And in case it wasn't clear, you're running this prompt in the project. So you create the project first, and then you paste the prompt into that project and run it this way. In addition to the prompt, it has access to everything that you've put into that project.
And the project also needs to be maintained. Anytime there's new information about your company, add it to the project, update the instructions.
And you might be thinking, this sounds like a ton of work. And if you're creating a single blog post, sure. But if you want to create 100 blog posts or 1,000, it's an incredibly efficient use of your time.
