ITX is no stranger to experimenting with new and interesting technologies. Of course, hot-button and top-of-mind is AI-generated content: does it have a place within copywriting?

And while we at the ITX Content Team believe it does, we’ve continued to investigate the technology. Our research, which includes several internal experiments, has led us to believe that it could be a useful tool for a content writer, with some major caveats.

So, how can you get the most out of a large language model (LLM)? If rumors are to be believed, everyone’s already using AI for everything. If that’s true, then how does anyone stand out if we’re all using the same programs?

Really, there is no secret. There are many different schools of thought, and some tricks you can use to improve an AI’s output, but it’s up in the air at the moment. Our general philosophy: to get good out, you’ve got to put good in.

Be a good writer

The best thing you can do for an LLM is to be clear and be specific. Spend some time thinking about how best to phrase your request. Overly broad or vague instructions aren’t going to make you stand out.

So, asking an LLM for “a blog post” is going to get you the same generic, boilerplate response that it would give anyone else. But, get more specific and ask for “a short blog post from the perspective of an expert in their field that is giving advice to someone who’s just starting out” and you’re going to get something way more unique.

The things that AI chatbots respond positively to are the same things that humans respond to: well-written, concise, and specific instructions. Being a good prompter is being a good writer.

Repeat, repeat, repeat

Large language models prioritize repetition. It’s how they figure out what kind of texts they’re reading when they get trained. Emphasize important points by repeating them multiple times in your prompts.

Don’t accept the first draft

Just like a human writer, an AI won’t knock it out of the park on its first attempt. It’ll misinterpret you, make mistakes, or just hallucinate. But, if that happens, you shouldn’t start over! Respond to the AI and call it out on its mistakes. By pushing the program to iterate on itself, asking it additional questions, having it rewrite a section with a new key detail in mind, the better output you’ll receive.

The structure of a good prompt

While there is no perfect prompt taxonomy to follow for every use case, here’s a good starting point.

There are typically four key elements to crafting a “good” prompt:

• Role: From what point of view or perspective should the AI write? Start by asking the LLM to take on the role of a critic, expert, or professional by using words like “as a…” or “pretend that you are a…”
• Task: What do you want the LLM to do? Outline what the objective is. Again, be clear. Ask it to research a particular topic or summarize the email chain you’ve been ignoring.
• Requirements: What additional context should the AI consider? If it’s writing a blog post, tell the AI who the audience is and how long the post should be. The more detail you add here, the more the AI has to work with in its response.
• Instructions: How should the LLM deliver the requested task? Tell it exactly what the deliverable should look like. If the AI is doing competitive analysis for you, be specific in asking for a bulleted list of companies or a full write up of each one. Give it some steps to follow.

An example

Below is an example of a good prompt, based on everything we’ve gone through up to this point. Notice how much detail we can get into this prompt, which paints a very specific picture of the person and the problem they are facing.

And now, broken down into our categories.

Role: Act as a Kindergarten Teacher in New York state. 

Task: Find age-appropriate activities and lesson plans to meet the requirements of the NYS kindergarten curriculum for language arts.

Requirements: Consider a class size of 20 students of varying social, economic, and cultural backgrounds. These students have skill levels ranging from 1 year behind to 2 years ahead in development. 

Instructions: Give me 15 unique activities and lesson plans as a bulleted list. For each one, include time and materials needed, its subject area within language arts, and what New York state curriculum requirement it satisfies.

We’ve also made sure to repeat the most important things in our prompt to emphasize them: kindergarten, New York state, and language arts.

Lastly, we won’t just accept the AI’s output right away. We may find that its suggested activities are far too expensive for a kindergarten teacher, so we could ask the AI to limit itself to only activities using materials readily available in most classrooms.

Conclusion

If we had to sum up our best practices for writing with AI, it would be this: “Write a clear and specific request wherein you outline the role, task, requirements, and instructions for the AI, repeating the most important parts, and then iterate on the response until you’re happy with it.”

That’ll get you off the ground but stay tuned for another research-backed post with tips and tricks that’ll get you soaring!

---