Improve your success with AI by building a prompt library

Garrett says: “Build a prompt library for AI content.

The reality is that these LLMs (ChatGPT, BARD, etc.) are here to stay. They’re incredibly valuable for being able to generate content, depending on your resources, but you want to use it as efficiently as possible. Learning these tools now is going to future-proof you for SEO and in any other marketing capacity. You absolutely need to learn that skill set.”

Where does a prompt library live?

“It depends on your use case. ChatGPT has recently released an enterprise version of their software where you can save prompt templates. However, you could use something as simple as a Google Sheets spreadsheet and copy and paste all of your prompts in there.

You can also use tools like AIPRM, which is a prompt management tool. It allows you to build your prompt libraries and organise them. The whole purpose of this is to be more efficient, and that’s a great way to make efficiencies within the efficiencies.”

What kinds of prompts should SEOs be using?

“There are so many and, again, it depends on your use case. You might be using it for ideation or building out topic clusters, though I’m not saying that you should use AI to identify keywords and search volume because it can’t do that. You have to understand the limitations of these tools. It’s best for coming up with ideas, idea-building, and writing content.

You need to make sure that you’re checking your content so that it’s factually correct and accurate. These AI prompts can be really efficient for getting that first draft out on paper, paragraph by paragraph and section by section. For SEO, the most important areas are headlines, meta descriptions, and regex formulas for technical SEO and research.

When you’re building your prompts, you need to understand your constraints, understand the tone and your brand voice, and have prebuilt prompts that you can add to other prompts. That way, you’re saving time and maintaining a consistent output. Having brand tone instructions for your prompts will give you great output.”

How do you phrase a prompt to ensure that the output is consistent?

“When I’m using ChatGPT for SEO, I scaffold my prompts and I stack different components. Start with the role you want your LLM to have, whether it is acting as an SEO or as a subject matter expert in the home services industry. Then, the task. What type of blog post do you want? Do you just want a few short paragraphs, or do you want an entire blog post? I would recommend working piece by piece rather than creating a whole blog post from scratch.

Then, provide context and constraints. That’s where the tone and voice come in. You may want it to use a certain style, use a certain number of bullet points, or avoid using emojis. Finally, look for an opportunity to give examples of the type of output you want to see. The more specific the prompt the better the output, then you can fine-tune it with different variables to make it reusable across other tasks.”

How did you establish this formula?

“There are a few different large language models out there – OpenAI has ChatGPT, Google has BARD (which is based on PaLM), Meta has Llama, and Anthropic has Claude. They all have their own best practices for prompt engineering. As you read the documentation and listen to other people (such as myself) who spend hours and hours testing through trial-and-error experimentation, you start to get an intuition as to what works and what doesn’t work.

I can now generate a much more effective prompt much quicker than I did six months ago. When you’re getting started, you have to get a sense of the nuances. Equally, even when you are more experienced, there’s no such thing as a perfect prompt. There is always trial and error, and you can never completely predict the outcome. You still have to be prepared to tweak it so that you can get what you’re ultimately looking for.”

If an AI model starts hallucinating, can you start a new conversation without having to re-teach everything that you’ve been working on?

“You can pre-train them. For instance, OpenAI introduced custom instructions so that, if you have the tone really well mapped out, you can pre-train the model by simply installing those custom instructions. AIPRM does the same thing with custom profiles, where you can have a profile pre-trained so you don’t have to do it from scratch.

If you’re in a conversation, the model will remember what you’ve already talked about in the conversation. However, that degrades over time, the more that you write. At some point, it is going to forget the instructions that you gave. Therefore, there will always be an element of retraining.

I do want to emphasise that we are at the very beginning of this technology. Things that we’re dealing with now are probably going to look vastly different in six months, a year, or two years. Take everything that you’re learning with a grain of salt. It’s more about learning how to use the tools and being able to adapt as they get better and better.

I believe custom instructions are currently available for free on OpenAI. Obviously, they will probably offer extended capabilities in the professional version, but the most basic custom instructions are there. For some tools, like AIPRM and Writesonic, you typically have to pay for custom profile options.”

How do you establish the right tone that you should be using?

“You can combine your prompts with existing brand guidelines, but you still have to do a lot of the hard soul-searching work if you’re a business owner or a marketer. You have to understand your brand and you have to develop your brand identity.

These tools are effective in the same way that Photoshop is. If you’re already a great writer, it’s going to help you. However, it’s not going to turn you into a great writer overnight. You still need to be able to effectively communicate what you want. You’d be surprised by how many people don’t know what they want or how to articulate it. ChatGPT might help you ideate some of that, but it’s not going to solve that problem.

If you have brand guidelines, you can take all of those and use them to pre-train the model, then you’re good to go. That’s easy. However, if you don’t have clear, effective brand guidelines, you should consider building them out first.”

Does giving a novice a great tool just allow them to shoot themselves in the foot even more quickly?

“Absolutely. That’s why the education around these tools is critical going forward.

These tools hallucinate and they make up facts very confidently. If you’re not a subject matter expert on whatever content you’re producing, you have to assume that it’s not always going to be correct. You have to double-check, or get a subject matter expert to review it, to make sure the output is accurate.”

How is Google dealing with AI-generated content at the moment?

“They have publicly encouraged it, in the sense that you’re not going to get penalised strictly for using AI content. They’re currently skirting the issue by saying that, as long as the content is helpful, then you can potentially rank. If your content will help users, then that’s the type of content Google wants.

They’re not saying ‘yes’ or ‘no’ to AI content. They’re saying it’s how you use the tools, and they’re trying to avoid letting spam, unhelpful, factually inaccurate content rank the highest. Google’s under the gun anyway, when it comes to the performance of search results in general. They’ve done a lot of work over the last few years. One thing to pay attention to (that’s adjacent to all of this) is the search generative experience. That is Google’s version of integrating AI into search results and it will certainly impact SEO going forward, which is a whole other can of worms.”

When you’re using AI, what part of the process should humans keep for themselves?

“There are two major areas where humans are always going to be valuable in the content creation process. One is editing. Editing has already become so much more of an important skill than it was before. AI will continue to get better, but it doesn’t understand anything. It’s a very sophisticated piece of tech that can predict the next word based on its training data. Right now, it doesn’t actually think.

The second key area is subject matter expertise that is unique to people, which is based on EEAT – specifically experience and expertise. It’s about being able to infuse your content with real experience from real experts. AI doesn’t know how to actually fix a kitchen sink. If you’re a plumber, you want to talk to a plumber who deals with the latest tools, techniques, and technology. You want to get that specific expert to inform your content, and that’s what Google’s going to care about going forward.

As you see more and more AI-generated content, new ideas that are helpful for users are going to help content surface to the top. Supplementing your AI-generated content with expertise and experience is going to be the best combination for SEO going forward.

The technology is amazing, but it has a lot of problems. When e-commerce first came to the internet, none of us wanted to put a credit card in to buy something online. Yet here we are, a couple of decades later, and you have trillions of dollars online.

With any of this technology, it’s slow. However, this isn’t like Google Glass or even VR. AI-generated content is here, and it is going to disrupt a ton of industries, so don’t put your head in the sand. At the very least, learn about it and understand what it can do.”

If an SEO is struggling for time, what should they stop doing right now so they can spend more time doing what you suggest in 2024?

“Don’t get too sucked into the trends that are being parroted on social media. It’s very easy to chase a new trend, but SEO is still all about fundamentals. It’s all about being helpful. It’s all about internal linking. It’s all about making sure that your content is easily consumable and discoverable.

Ultimately, don’t spend too much time getting caught up in what people are saying about SEO and just learn for yourself. Every situation is unique; that’s why we always say, ‘It depends.’”

Garrett Sussman is Demand Generation Manager at iPullRank, and you can find him over at iPullRank.com.