Stop getting garbage content from AI! In this episode of Whiteboard Friday, Chima Mmeje shares 7 actionable tips for writing great content with LLMs by leveraging training documents, guardrails, and storytelling.

Click on the whiteboard image above to open a high-resolution version!

So everybody is trying to write content with AI these days. You have your popular tools, like ChatGPT, Gemini, and Claude. And then when you put your content in there, you give it a prompt, right, and the content that ends up coming out is just garbage, trash, absolutely unusable.

And what I see most people doing is that they complain. "Why can't I get great content when I try to write with LLMs?"

But what if there was a way? What if you could write great content when using an LLM?

Hi. My name is Chima Mmeje, and I am the Senior Content Marketing Manager at Moz. Today I'm going to be showing you how to use LLMs, like ChatGPT and Gemini, to write content, basically, seven tips for writing great content.

And this is all based on stuff that I am doing. So it's not generic advice you can get on the internet. 

Now, before we get into these tips, I want to just put a caveat. These tools, by themselves, won't make you a great writer. But they can help you get more writing work done.

I say this because LLMs are garbage in and garbage out. They cannot create authority. They cannot create expertise. They swallow up the whole internet and are only capable of giving out the quality that you put in.

So the only way for you to get good quality from them is for you to be an expert writer who understands storytelling. Got it? 

All right, let's go. The first thing you want to do is to create a training document that you will feed it as an example.

So let's say I'm trying to write a blog post, right? If I'm trying to write a totally dashing blog post that has a strong editorial voice, I'm going to pick one of the articles I've done in the past that was a very similar style, and I'm going to use that as my training document. And that is what I'm going to feed into the LLM every time I'm trying to write that type of blog content.

If I'm trying to do a webinar landing page, the same thing applies. I will take a previous webinar landing page that I have written, and I will give it to read as an example.

Asking these LLMs to just write content for you, with no training or context, is only going to lead to bad results.

Now, once you give it that content, the next thing you want to do is provide context. I keep saying this. You can never, ever give these LLMs too much context. There's no such thing as too much context.

Now, what are you trying to achieve with the content that you're writing right now? In the training document that you've given it, what's worked well? What did you like about the training document? How did you go from point A to point B to the end of the training document?

You're giving it all this context so that it understands how it's going to structure the content, how it's going to mirror your tone of voice, and how it's going to give you a similar output.

Now, once you provide all of that context, the next thing you want to do, you want to create a project. You can even ask LLMs to help you create the instructions for the project.

And then you create the project. If you're using ChatGPT, that's going to be a GPT. If you're using Claude, that's going to be a Project. And if you're using Gemini, that is going to be a Gem.

Then you do two things. You provide guidelines for writing. I have a super-detailed list of instructions for my GPT that I use when I'm writing. It's really long. It has like I think maybe up to 100 bullet list of things that I want it to do. In the same way, I also have guardrails to avoid, certain words that I don't like. I've noticed, for example, it likes to use the word "shape" every other sentence. Or maybe it puts in an em dash. Or maybe it writes like really short, abrupt sentences that are like two or three words.

All those things that I don't like, that look very AI-ish, that is going to be the guardrails that I'm going to ask it to avoid.

This is very important because, otherwise, it's just going to give you literal garbage.

Now, the next thing you want to do is to add a document detailing your core offerings and values.

This is so important because you're not just writing content, right? I consider myself a product-led marketer. So in that context, every time I'm creating a blog post, webinar content, any type of content at all, I am looking for opportunities to plug in products.

So let's use Moz as an example. We've got offerings like Moz Pro. We've got AI Visibility tool. We've got MozBar. We've got Moz Local and a bunch of other offerings. I'm going to put all of this into a document, and I'm going to explain: What does this tool do? What is the value? Who is this for? I'm going to put in the positioning statement.

I take that document, and I add it to the project as part of the training document or as part of the instructions.

That way, when it's writing or generating content and there's an opportunity to naturally, this is so important, to naturally include that value in the content, it's going to put it in there, and I don't have to do it every single time.

When it misses the context, I can provide that guidance and say, "Hey, you missed this. You are supposed to add X, Y, Z in here." Then the next time it has an opportunity, it understands what the right context is to plug in value, product-led value naturally.

A big issue I keep seeing people make when writing with LLMs is trying to write the whole thing at once. Give me a 2,000-word blog post on X, Y, Z. Help me write a product-led content that does X, Y, Z. You're not going to get great output when you ask the LLM to write super-long content because it's going to use minimal effort. The best way is to write in small sections.

You have your introduction. I want to write an introduction. I want it to be three paragraphs. I want the first paragraph to say X. I want the second paragraph to say Y. I want the third paragraph to say Z. That allows me to control the output. It gives me that small section, and then I can provide feedback at a granular level.

"Here's what you missed. Here's what I changed. Compare the two of them." You're able to give it feedback and to control the output.

The same thing with my next H3. A H3, I'm able to tell it this is what I want to cover inside this H3. Then, when it gives me the output, I'm able to give it feedback. You can't give feedback if you're writing the whole thing at once. Where are you going to start from? 

Now, when you're done getting the output from the LLM for that small section of content, you want to personalize it. For me, this is the thing that makes your content stand out from that sea of sameness that we currently see on the SERPs.

When you personalize it with storytelling, it just makes it stand out. It makes it look like something that you wrote versus something that an LLM wrote. When you put in expert-led insights, it just gives it that personal touch. It's like I've had this experience. I've made this mistake. I have done this thing, and I got this result. 

It now makes it feel like you wrote it. Even better, nobody else is going to have a similar experience to you. That means that when you publish this content, people are not going to think, "Oh, an LLM wrote this." They are going to think, "Oh, you authored this post based on your experience." That is what makes you a trusted source. 

Then, finally, you want to give it feedback.

Every time it gives you output and you've added context, you've added storytelling, you've added expert-led insights, you've added anything, you've made changes, you take the final version of what you've written, for that section, and you give it back to the LLM. You say, "This is the final version I'm going with. I have made X, Y, Z changes because you did not include this, or you left this out, or you did this. You've gone against our guardrails. You have not followed our instructions." You're able to give it that feedback.

The more you give it that feedback, the fewer mistakes it's going to make. The goal ideally, for me, is to get to 70% quality output. I'm not aiming for 100%. I'm not crazy. Seventy percent, that's the goal. We want to get to 70% of great output so that we're making fewer and fewer changes to the output.

Again, stop trying to automate content. I don't think that's really viable if you want to maintain quality. The trick is to write in batches, personalize the output, and then make it feel like you. But remember, you have to keep training these LLMs because they are machines at the end of the day.

And that, my friends, is how you write great content with LLMs. Thank you.