One morning, a few weeks ago, I found myself in a position that I hadn’t been in for a long time.
About a month prior, I had offered to write a blog about how our team was using ChatGPT inside and outside of work. As a former reporter, the idea brought on a bit of nostalgia. The approach I had in mind would take some time to execute, but my projects were moving smoothly and I had a rare bit of extra time to put to good use. I said I would have it done by May 4th.
When Monday came, I asked for an extension to the following Friday. It was granted. Soon it was next Thursday, and I found myself in the same position I mentioned at the start. Having a past due deadline and nothing but a blank page in front of me evoked some other kind of much less pleasant nostalgia.
I didn’t have time to do interviews with the team, so I decided to reduce the scope of my story. Maybe I could even reduce it to something that ChatGPT could do for me. After all, wasn’t one of the initial concerns when ChatGPT was released that students writing papers would become a thing of the past? I decided to test the merit of those concerns.
My new topic: Is It Time-Efficient to Have ChatGPT Do Your Writing for You?
I began documenting the process of getting ChatGPT to write an original article about how to use ChatGPT to write an original article for you. In order to evaluate whether the approach was time-efficient, I started a timer.
Throughout the process, the timer was running anytime I was directly working with ChatGPT or its responses, and stopped anytime I turned my attention back to this higher-level writing assignment. The timer in my project management system (Wrike) ran the whole time I was working.
Beginning a Slow (and Painful) Journey
I took a moment to think through the best process for doing this, based on some of my knowledge around ChatGPT’s limitations. Knowing that it would time out after a certain point, so it wouldn’t be able to write the full article based on one prompt, I asked it to write an outline for the article to get me started.
FourFront:
Provide me with an outline for writing an article on how to use ChatGPT to write an original article, given its limitations.
ChatGPT:
I. Introduction
A. Briefly introduce ChatGPT as a language model developed by OpenAI
B. Explain the purpose of the article: to guide users on effectively using ChatGPT to write original articles while understanding its limitations
II. Understanding ChatGPT’s Limitations
A. Explain that ChatGPT is a language model and not a human with real-world knowledge
B. Discuss limitations such as:
Lack of up-to-date information beyond its knowledge cutoff date
Inability to fact-check or verify information
Read the full outline: ChatGPT Outline
As a next step, I asked ChatGPT to write each individual section of the article. I allotted 100 words to the introduction, 250 to the tips, 100 to ethical considerations, and 100 to the conclusion. To ensure a smooth flow between outputs, I instructed it to write each section as though it were writing the next section of an article from which the previous prompt was an excerpt.
This took about 35 minutes, and what came out did not sound anything like something I would write. To find the best way to fix this, I turned back to ChatGPT.
Read where we’re at here: ChatGPT Blog Version 1
FourFront:
What is the best process for using ChatGPT to write an article that looks and sounds like something that I would write?
ChatGPT:
- Provide initial instructions: Begin by providing clear instructions to ChatGPT about the topic, tone, and style you want the article to have. You can specify the level of formality, preferred vocabulary, and any other relevant details that reflect your writing style.
- Provide writing samples: Share some of your previous articles or writing samples with ChatGPT. By analyzing your writing style and language patterns, ChatGPT can better understand how you write and attempt to mimic your style more effectively.
- Engage in a conversation: Instead of requesting the entire article in one prompt, engage in a conversation with ChatGPT. Gradually build up the article sentence by sentence or paragraph by paragraph. This allows you to guide ChatGPT’s responses, review and provide feedback, and iteratively refine the output to match your style.
- Review and provide feedback: Regularly review the generated text and provide feedback to ChatGPT. Identify areas where the output matches your style and where adjustments are needed. You can ask ChatGPT to revise specific sentences or paragraphs to align them more closely with your writing style.
- Refine and iterate: Continue the conversation with ChatGPT, refining the text with each iteration. Give clear feedback when the output does not match your style and provide specific instructions or corrections to guide ChatGPT towards the desired outcome.
- Incorporate personal touch: To further enhance the resemblance to your writing style, consider adding personal anecdotes, examples, or unique perspectives that reflect your own experiences and voice. ChatGPT can then incorporate these elements into the article.
Remember that ChatGPT is an AI language model, and while it can attempt to emulate your writing style, it may not perfectly replicate it. Regularly reviewing and providing feedback is essential for achieving the desired result.
Well, I’d already blown my opportunity to provide clear instructions on tone and style from the start, and the goal was to get this done quickly, so I decided to go with the quickest sounding option from the list above as a starting point: Provide writing samples.
FourFront:
Rewrite the first half of the following article in the style and tone of the author of this piece: https://www.fourfront.us/blog/is-seo-a-good-career/
I read the output. No, no, no, no, no! This still sounded like it was written by a robot. Did I write like a robot? I didn’t think so and I had to jump into a meeting. This was going to have to wait until later.
Ok, what did ChatGPT even do to make this sound version some more like me? I asked it.
FourFront:
Tell me what considerations were used when fulfilling the previous prompt in regards to updating the article to follow the style of the author of this piece: https://www.fourfront.us/blog/is-seo-a-good-career/
ChatGPT:
When fulfilling the previous prompt to update the article in the style of the author of the piece from fourfront.us, I considered several key elements:
- Tone and Voice: I aimed to adopt a similar tone and voice as the author by maintaining a professional and informative style. The writing style reflects a conversational tone while providing valuable insights and advice.
- Structure and Organization: I followed the structure and organization of the original article, ensuring that the content flowed smoothly and logically. I maintained the subheadings and bullet points to enhance readability and comprehension.
- Vocabulary and Sentence Structure: I used vocabulary and sentence structures similar to those employed by the author. The language is concise, clear, and easily understandable, avoiding excessive jargon or complex terminology.
- Content and Information: I retained the core content and information from the original article while consolidating it into approximately 550 words. The essential tips, considerations, and insights were preserved to provide a comprehensive overview of using ChatGPT for writing original articles.
By considering these elements and aligning the writing style with the author’s piece on fourfront.us, I aimed to create an updated version that follows a similar style and tone while delivering the requested content in a concise manner.
That was interesting, but I didn’t feel that it provided me with any insight to help guide the style and tone of the new piece. Since that didn’t work, I decided to try some reverse engineering by asking ChatGPT to deconstruct the style and tone of my other article:
FourFront:
Describe the style and tone employed by the author of this article: https://www.fourfront.us/blog/is-seo-a-good-career/
ChatGPT:
I apologize, but as an AI language model, I am unable to directly access or browse the internet, including specific articles. I can generate text based on my training on a diverse range of data, but I cannot provide real-time analysis or descriptions of specific articles. If you have any questions or need assistance with a particular topic related to SEO or careers, feel free to ask, and I’ll be glad to help to the best of my abilities.
*Gulp* Since when? How did it review my article before? That didn’t seem right. I did a quick Google to see if anything had changed in the last few minutes. Nothing came up indicating this was true, in fact there were many articles saying the opposite. I decided to try asking directly.
FourFront:
Can ChatGPT browse the internet?
ChatGPT:
No, ChatGPT, including the version I am based on (GPT-3.5), cannot browse the internet or access real-time information. My responses are generated based on the knowledge I acquired during my training, which goes up until September 2021. I do not have access to current or specific internet content, including articles or websites. I can provide general information and engage in discussions on various topics, drawing from my pre-existing knowledge.
FourFront:
Please summarize this article: https://techcrunch.com/2023/03/23/openai-connects-chatgpt-to-the-internet/
ChatGPT:
I apologize for any confusion, but as an AI language model, I don’t have direct access to the internet or specific articles like the one you mentioned on TechCrunch. Therefore, I cannot browse the article and provide a summary of its content. My responses are generated based on the training data I have received, which goes up until September 2021. If you have any specific questions or need assistance with a particular topic, feel free to ask, and I’ll be happy to help to the best of my abilities with the information I have.
Well, either ChatGPT was lying to get out of doing my work for me or the internet was collectively mistaken. I decided I would come back to this later.
Navigating the Final Stretch
I was two hours in, and still didn’t have an article that I thought would pass as something that I wrote. When I tried to get ChatGPT to run the same prompt as the day before, it still didn’t work. As a workaround, I pasted the entire article into the prompt box and asked ChatGPT to describe the style and tone of the piece.
While I couldn’t help but think that something would be lost with this approach, as I wasn’t sure it would understand what was a heading, a link, etc (in retrospect I probably could have tried using the page source…), it did return a good response:
ChatGPT:
The style employed by the author of this article can be described as informative, persuasive, and conversational. The tone is generally positive, enthusiastic, and confident.
The author begins by stating a clear and straightforward answer to the question at hand, presenting SEO as a good career choice. They highlight the potential benefits of pursuing a career in SEO, such as good income, work/life balance, job opportunities, and creative challenges.
The author then adds a personal touch by introducing Jared Groff, the Search Engine Optimization Manager at FourFront, who shares his perspective on why SEO is a great career choice. This inclusion of a real-life example adds credibility and authority to the article.
Throughout the article, the author uses a conversational tone to engage the reader. They pose questions, provide answers, and offer insights in a friendly and relatable manner. The use of casual language, such as “take it from,” “brainstorming ways,” and “yep… SEO is pretty safe,” contributes to the approachable tone.
The article also employs bullet points to organize information and make it easily scannable for readers. This formatting choice helps to convey the information in a clear and concise manner.
Overall, the author aims to inform and persuade readers about the merits of pursuing a career in SEO while maintaining an engaging and positive tone throughout the article.
I started to organize this into a prompt that I could use to ask ChatGPT to once again update the style and tone of my article, and then I realized that might be a good job for ChatGPT.
FourFront:
Consolidate your previous response into a prompt that will work well for instructing ChatGPT to rewrite a different article in the tone of the author of the ‘Is SEO a Good Career’ article that it just read.
ChatGPT:
Title: “The Benefits of Yoga for Mind and Body”
Introduction:
Yoga is not just a physical exercise; it’s a holistic practice that brings numerous benefits to both the mind and body. From increased flexibility and strength to reduced stress and improved mental clarity…
Well, that backfired. Back to Plan A. I started with the first half so that it wouldn’t time out.
FourFront:
Reword the below article so that style is informative, persuasive, and conversational, and the tone is generally positive, enthusiastic, and confident.
The first half sounded better. Not quite like me, but more workable than the original. I added the second.
Ugh, it rewrote it in a format completely different from the first half. I tried a different prompt:
FourFront:
Reword the second half of this article:
“Second half:
[pasted second half]”
So that it carries the same informative, persuasive, and conversational style as well as generally positive, enthusiastic, and confident tone of the first half of the article (below) throughout:
“First half:
[pasted first half]”
That prompt didn’t work either. It essentially turned the previous section into its own article, so it wouldn’t be compatible with the first half. I tried getting more specific:
FourFront:
Reword the below section of the above article using the same style and tone directions as with the previous prompt. Do not add any new information:
[pasted second half of original article]
It’s worth noting at this point, that both for the purpose of documenting this process and for trying to get the desired result without losing information, I had 5 saved word files and 3 different Google Docs files open, so navigating back and forth between them was causing some headaches.
That aside, I thought that what ChatGPT had put together so far was ready to move into a manual proofreading stage.
Read where we’re at here: ChatGPT Blog Version 2 – Tone Edits
Putting Together the Finishing Touches
Formatting the article was a pain. All of the copying and pasting, combined with responses that generated in various formats, left a mess and the process of cleaning it up was a time-consuming one.
When I was done, I decided to compare that approach with using ChatGPT to make the formatting uniform using an iterative process of pasting it in sections into the prompt box and employing the ‘format this as though it’s a continuation of the previous response’ prompt. Both methods took about 8 minutes.
Next, it was time to fine tune the article into language closer to my own. I read through and tweaked headings, changed adjectives, and restructured sentences as needed. This took another 20 minutes.
Finally, it was time to try finding some reliable sources for my article. These would need to be authoritative and offer good linking opportunities. Instead of identifying these claims and searching individually, I decided to see if ChatGPT could do this in a batch.
FourFront:
Identify authoritative sources on the internet that I can link to in order to support claims made in this article. Return these in the following format:
Source URL | Anchor Text from the below article that I can link from | Paragraph # of the article (to assist in finding the anchor text)
Source URL | Anchor Text | Paragraph #
- https://www.example.com/reliable-sources | reliable sources | Introduction paragraph
- https://www.example.com/fact-checking-guidelines | fact-checking and verifying the information generated by ChatGPT | Introduction paragraph
- https://www.example.com/writing-tips | valuable tips | Subsection “When getting started”
- https://www.example.com/clear-prompt | clear and specific prompt | Subsection “When prompting ChatGPT to Write”
Nope, that approach wasn’t going to work. I guess ChatGPT still wasn’t searching the internet. I consulted my timer:
While my article did not contain sources, I felt it was passable and decided to skip the step of manually searching them out. At this point, I’d found the answer to the question I’d set out to answer: Getting ChatGPT to write my article for me was not time-efficient.
But could it have been? Looking back over the process I followed, and comparing it with the recommended process that ChatGPT gave me, I knew there were a few ways I could have approached it better. My topic could have been clearer.
Also, I’d refined the process as I went along, learning via trial and error. A predefined list of effective prompts based on those learnings could help avoid some of the errors and rewrites I did, which could have shaved off an hour or so. Prompts of those types are probably available online now, and someone will probably create a program that combines code and large language models to make this a one-step process. Maybe they already have.
The Verdict
All that said, even if that program existed and my final article could be produced in one click, it would still require proofreading, research to identify the right sources, and then a bit of rewriting to work the sources in smoothly. That would save time, but essentially, you’d just be going through the same traditional processes of finding a topic, researching, writing, and proofreading in a different order.
After all this, I was inclined to agree with the conclusions of the article I had ChatGPT write:
While ChatGPT is a great brainstorming tool and writing partner, it can’t do your work for you… yet.
Final ChatGPT-written, human-edited article can be found here: ChatGPT Blog – Final Version
** On a side note, when I submitted this article for QA, I’d logged almost 7 hours to the process in our project management system. Probably could have just gone with my original idea. Sorry, Chris!
