AI Prompting for Bloggers: My Trial-and-Error Discoveries

Six months ago I set out to see if artificial intelligence (AI) could help me be a better blogger. In this post, I am sharing what I learned and providing tips to fellow bloggers.

I want to thank the many trailblazers in business development, program management, and content development who helped push me along with their presentations, workshops, and webinars. I have absorbed their guidance and made it my own.

My journey took me from a basic understanding of AI—through experimentation—and, finally, to a state of cautious optimism about its benefits and potential pitfalls, even dangers. I experimented with Poe, Grammarly, Claude, and ChatGPT (mostly the latter). I also tried various prompting techniques and patterns (primarily by accident). I had some successes and some failures.  Here’s what I learned along the way.

Categorizing LLMs

One of the biggest challenges for me—certainly, when writing this blog post—was how to think about large language models (LLMs) like Poe, Claude, and ChatGPT. I wasn’t so much concerned with how they work (more on that later) as I was concerned about how they fit into the world of content development. (I am at heart a pattern person, after all.)

I asked two questions: Where do LLMs fit in the potentiality of AI? What are the best uses of AI for writers?

Seven Use Cases for AI

Obviously, the world of AI includes more than just Alexa and Siri. But it also includes more than LLMs. According to Kathleen Walch of Cognilytica (now owned by the Project Management Institute [PMI]), AI currently has seven use cases in the world:

• Hyper-personalization (targeting a specific human)
• Recognition (think facial recognition)
• Conversation and Human Interaction (think chatbots and voice assistants)
• Predictive Analytics and Decisions (think medical applications)
• Goal-Driven Systems (think gaming)
• Autonomous Systems (think self-driving cars and robots)
• Patterns and Anomalies (recognition or detection)

According to Walch’s presentation “Mastering the 7 Critical Patterns for Success” (at the PMI Mile Hi’s Women in Project Management Leadership Summit on Sept. 17, 2024), LLMs like ChatGPT cross at least two of these use cases: Conversation and Human Interaction plus Patterns and Anomalies. From what I’ve seen business development professionals do, I would argue LLMs can also perform Hyper-personalization well. Thus, you can have conversations with LLMs (just as you do with Siri and Alexa), but you can also ask them to do work—including personalizing content—for you.

Not work instead of you.

They are especially adept at patterns (like me!). In other words, LLMs can assist you because they are good at synthesizing information, categorizing and defining it, uncovering what you missed, organizing content, and rewording it. And they can do it rapidly—more rapidly than a human assistant could.

However, you must remember that LLMs can do all that because they have at their well-programmed fingertips all publicly available data and content—including copyrighted material—on the internet. (Ah, the rub! More on that later.) Well, all data and content within a specific date range (for now).

Generative and Transformative Use

You can treat an LLM as an assistant, a marketing consultant, a strategist, an editor, and even a content designer because LLMs work well in two arenas specific to the work of writers: generating content and transforming content.

I understood the distinction between these two arenas better after recently listening to Jayme Perlman, a senior tech writer at GitHub. Perlman spoke about “Using Publicly Available AI Interfaces as Editorial Tools” to a group of technical editors through a Society for Technical Communication mini-conference (Nov. 19, 2024). The generative uses of LLMs include everything we’ve feared: creating arguably “original” content based on a prompt. The transformative use of LLMs refers to using prompts (and uploads) to manipulate existing content while preserving its original meaning.

As bloggers experimenting with AI, we should be conducting trials in both arenas so that we can find:

• A general comfort zone with AI assistance
• The right tool or set of tools
• Prompt patterns and techniques that achieve desired outcomes

Regarding the second point, note that the tools you use for generative work might differ from those for transformative work. For generative work, I use ChatGPT, but I have also tried Poe and Claude. You can also use these tools for transformative work. But I’ve relied more on Grammarly for transformative work. Some of my editing peers like Hemingway.com and QuillBot for improving or transforming content.

Your prompt patterns and techniques will also likely differ between your generative and transformative work. More on that later.

I’m sure you’re aware that LLM prompting has rapidly become a cause célèbre. Many articles, courses, webinars, and social media posts are available to educate us on the best way to prompt an LLM. My advice is to glean what you can from these and then experiment on your own. You don’t have to spend a lot of money. Many information sources are free or low-cost. Basic accounts with most LLMs are also free (for now).   

Question to Strategy Transformations

To start creating effective prompts, take a step back and imagine that your LLM of choice is your peer writer or a sympathetic editor. Consider the types of questions you would ask a person in that role if you were looking to start, improve, or expand a piece of writing.

I should note that I don’t rely on LLMs to generate an entire blog post. For me, doing so would not be serving my readership. But that’s a personal choice.

Start with Questions

I use generative prompts to complete, define, break down, outline, and create headings for content. Thus, I look for answers to questions such as:

• What is missing?
• How can I best define this term?
• What are the steps for accomplishing this outcome?
• What are some companion concepts to consider? Follow-on blog post topics to consider?
• What is the best sequence of concepts for this topic?
• What’s a good heading for these paragraphs?

I am essentially using the LLM as a partner in the writing process by asking for assistance with specific writing tasks. I have preferred to use it when I am stumped or believe I have overlooked something, but I plan to use it more as a paired writing partner eventually.

I use transformative prompts to improve my writing or change it in some way. So, I look for answers to questions such as:

• What is a better way to say this?
• How can I shorten or break up this sentence? Heading? Section?
• How can I make this sound less formal? More empathetic?
• How can I expand on this concept?
• How can I change this description to a set of steps? Or a bulleted list?

Note that I have Grammarly installed in my browser and in my copy of MS Word. So, I consider its wording suggestions as I go. But I am not a slave to its tips. I hit “Dismiss” or “Reject” as often as I hit “Accept.”

Add Some Strategy

Once you uncover the questions you want to ask about your content, you can begin to put some strategy around them to create prompts. For your prompts, consider how to provide context to the question(s) you want to ask the LLM and decide whether you wish to have a conversation or direct some output.

When LLM experts like Ethan Mollick refer to providing context to a prompt, they are referring to methods to refine the scope of your inquiry so that you can get “more unique answers that better fit your questions.” By giving your prompt context, you are narrowing, containing, and instructing. After I had already been doing some experimentation, I found two mnemonics that helped me to better form a context for my LLM prompts: TRACI and CREATE.  

The TRACI acronym, first mentioned, I believe, by Steve Brand of structuredprompt.com, refers to providing the LLM with a Task (T), a Role (R), an Audience description (A), some instructions and parameters for the output you wish the LLM to Create (C), and an Intent (I) for that output (for example, persuasion). Including each element in a prompt should get you close to what you want the LLM to accomplish for you. But to hit the right voice, tone, and style, you might have to also provide examples.

That is where the CREATE approach comes in. Taught by Dave Birss in his online LinkedIn course, the CREATE approach to prompting encourages you to provide examples of your tone of voice, even previous blog post headlines. The CREATE acronym stands for Character (C), which is the role you want the LLM to play, your Request (R), Examples (E) you want the LLM to follow, any Adjustments (A) you want to include to refine your request (such as “include subheads”), the Type (T) of output you are seeking, and any Extras (E) you want the LLM to consider (for example, an overriding instruction such as “ignore everything before this prompt”).

These acronyms both represent strategies for structured prompts, a methodology driving the launch of many subscription tools in the past year. While structured prompting might be overkill for simple requests, building a strategy for prompting should be part of your experimentation. The key is to be as specific as possible. You should also be aware of prompt techniques that can broaden your approach and strengthen your interactions.

Birss recommends building a library of prompts in that use various patterns and techniques. But what is the difference between a pattern and a technique?

Patterns and Techniques

As you develop a skill for working with LLMs, you can collect prompting patterns and techniques that work best for you. A prompt pattern is a fill-in-the blank logical sequence that you can use over and over to streamline your AI work. Many experts have provided prompt patterns and examples in various formats over the past two years; use your imagination to adjust them for your circumstances.

A prompt technique, a term from prompt engineering, describes the amount of data you provide the LLM when asking for a result. Prompt engineers speak of zero-shot prompts (direct, unadulterated questions like the ones I’ve included above), one-shot prompts (include one example, scenario, or model), and few-shot techniques (which include multiple examples, scenarios, or models) as well as complex reasoning prompts (such as compare and contrast). (Structured prompts often combine prompt techniques.) For more on prompting techniques, including prompt chaining, I suggest reading Michael Watkins’ LinkedIn article “Twelve Advanced Prompting Techniques for Large Language Models.”

But let’s get back to simpler territory. Here are some example prompt patterns to help you start using LLMs as writing assistants.

Here is a PDF of this same table for your convenience.

This list is not exhaustive. I am still learning! Add your pattern tips in the comments below.

Takeaways and Tips

The most significant caveat in working with LLMs is, of course, that current versions are far from perfect. Because they are programmed to give you an answer every time, they still hallucinate (provide false information). They can also use language that is repetitive and vague. Here are some tips for counterbalancing those imperfections:

• Never lift text directly from an LLM response. That content is probably too obviously LLM generated. Update it with your own words and ideas.
• Do a factual check of the assertions in the LLM response. Check the sources cited (and I always ask the LLM to provide its sources).
• Check the clarity and conciseness of the LLM response. Edit the language accordingly. Check for relevance to your audience.

On the ethical side of things, remember that public LLMs are open, self-adjusting systems. Your data and your content are not secure once you upload it. LLMs are, to some extent, content thieves. So, if you don’t want your content to become fodder for the next LLM synthesis, follow these tips (with thanks to Jayme Perlman):

• If possible, use a closed, customized LLM (if you are lucky enough to work for an organization that has one.)
• If your organization does not have its own LLM, use a reputable public LLM and pay for a subscription. Check the subscription terms of service.
• Work with small pieces of content only. Never upload your entire article or blog post to a public LLM.
• Follow discussions about protecting your copyrighted content from public LLMs. Some movement is underway to license original content explicitly.  

I know that the resistance to LLMs is real. I see you. I sympathize with you. I want to hear from you. And I want you to know that I wrote this blog post without the assistance of an LLM. But will you read it?

Image by Riekus from Pixabay.