I wish to start with a confession: When ChatGPT emerged into our lives about a year ago, I immediately developed what could be best described as an immune response. I love new tools and gadgets, and I have a background in technology, so I was curious about what this tool (or, as some people will claim, entity) could do, but I wasn’t positively curious. Instead of thinking about how it could help me, I was focused on finding arguments for why using AI for creative activities such as writing was a terrible idea.
In retrospect, it wasn’t a response to the tool or the technology but rather to the hype around it. In no time, my social feeds were full of people who did whatever they could to demonstrate just how good ChatGPT is for both writing and reading and just how useless it would be to compete with it. It seemed (and maybe still seems) that people are just waiting for an opportunity to write and read less and delegate these “nuisances” to their new best artificial friend. And I couldn’t accept that.
What many people considered a utopia, I saw as a dystopian future where we delegate all communication and all original thought to the machines: We let AI write for us, we use AI to listen and read for us, to summarize everything for us, and then we instruct it to respond on our behalf. In that future vision, we delegate our voice to an alien entity. We lose our voice. As someone who writes a lot and wishes to help people communicate better, I felt the urge to fight this not-unlikely future. I still do. I believe writing and reading are essential parts of what makes us human. Writing is a way to process information and think about new ideas. Reading is a way to understand what is on the minds of the people we interact with. It is a way to collect raw materials, which are then required to generate new ideas. I believe that if we delegate these functions to machines, we are doomed not because they will trigger some apocalyptic event but because we will gradually become less human. If we let machines read and write for us, our thoughts will become less valuable; we, humans, will lose our value.
I believed that, and I still do.
But all this hype and the instinctive backlash response made me blindsided. I failed to distinguish between the technology and how it is used. Delegating all writing to Large Language Models is disastrous, but who said this is the only way to use AI? I failed to consider the option that some parts of my creative process could benefit from an AI sidekick. Maybe this is an opportunity to become better communicators instead of lazier ones.
With that overdue realization, I started to explore subtler ways to integrate ChatGPT into my workflow — ways that will not come at the expense of my writing. I wasn’t looking for ways to shorten my process; I wanted to find ways to enhance it. I wasn’t looking to generate more words at a given time; I wanted AI to help me write better.
The article you are now reading was written entirely by me. I didn’t use ChatGPT to generate any part of it. I didn’t even use ChatGPT to brainstorm some ideas before writing. It is based entirely on my thoughts and insights; every word and sentence is mine. And yet, I used ChatGPT somewhere along the process. Somewhere between thinking of an idea and having a final draft ready to be published, there were some things ChatGPT could do that I couldn’t. I used ChatGPT where it has value that doesn’t come at the expense of my creativity.
Many people propose to “consult” ChatGPT before starting to write to come up with ideas or even skeletons of content. I did the complete opposite. I used ChatGPT for the first time only after completing the first draft. And still, it made this article better and possibly helped me improve my writing.
Here’s what I did.
My Writing Workflow
To understand how I use ChatGPT to enhance my writing without outsourcing it, you must first know some things about my process. Before I write even a single word in the first draft, I do two things that help me think about what I wish to say and the best way to express it: I collect material and model the text.
As you can guess, the collection phase is about collecting raw material that I will later use in the text. I capture any random insight or idea I have in the context of the topic; I look for resources and reference material; when relevant, I look for pieces of data; quotes, diagrams, other views, examples, and open questions can all potentially contribute to the piece I am writing, and so I create a collection of Content Bits that I can play with later. This phase is pretty unstructured; it is more associative and exploratory. I might end up using 50% of these building blocks, 90% of them, or just 20%. At this stage, I am just looking around (and thinking) about things that can help me build my case.
The next phase is where I think about what I am going to say more deeply. I am thinking about the best way to take all this raw material and create the skeleton of the text I am about to write. This is where I decide which Content Bits to use and which to omit; I decide how to arrange them for maximum impact; I consider if something is missing; and I refine each Bit to capture precisely what I want it to. None of the Bits is in its final form at this stage. In fact, they are captured as nothing more than bullet points. I create the logical skeleton of the text without adding the layers that make it readable and engaging.
When I am happy with the model I’ve created, I let it sink in for a while, and then (and only then) I open a blank page and start writing the first draft. As I write, I follow the logical flow of the model. I don’t have to think about what should come next, where each piece of the argument fits, or any other structural question. I can be fully invested in crafting the best text that captures the logic I’ve modeled.
Of course, a first draft is just what it sounds like: a first draft. So, once I finish writing it, I start the self-editing process. Typically, I will read the text I wrote out loud, make some corrections along the way, pass it through Grammarly, and finally, reread the result to get a sense of the (hopefully) better revision.
My writing process has many benefits (if I might say so), but at least one major drawback: for the most part, I do my own editing, especially when writing blog posts and my newsletter. Generally, it works, but it could work better because the power of editing — deep editing that examines the effectiveness of the content and not just technical aspects — lies in having someone else interact with your text. Editing is not about grading the text or just finding errors. Good editing is built on a dialogue in which the editor brings a perspective potentially different than yours.
Surprisingly (or not), this is where I found ChatGPT helpful.
Co-Editing with ChatGPT
When ChatGPT emerged, everyone was fascinated with its ability to generate text. And not just grammatically correct text, but content that actually sounds like it was created by a human. Sure, it was hallucinating quite a bit, but who doesn’t? With some double-checking and later with the release of ChatGPT 4, it seemed pretty plausible to use ChatGPT as a replacement for just about anything you need to write.
But I wasn’t interested in text generation. I didn’t want a machine to replace my writing. I wanted to publish my authentic text, regardless of how good of a job ChatGPT can do.
Luckily, Large Language Models like ChatGPT also excel in summarizing text. They seem to be able to capture the essence of the text quite accurately. They do so without knowing what I originally wanted to say. ChatGPT summarizes the text as it “understands” it. And this is a perfect trait that can help me edit my content.
Method 1: Making Sure the Text Follows the Model
So, I started with modeling the piece I was going to write. I arranged the Content Bits I had collected to form an effective logical structure. I wrote a draft based on this model, trying to follow the model’s design.
The problem is that often, in the flow of writing, I can unintentionally deviate from the model. I might allude to other arguments, omit some critical points, or fail to articulate something I had in mind. The model of the draft I wrote might be different from the model I had planned to follow. Sometimes, such deviations are acceptable, but for the most part, I invest in the modeling phase with the intention of following it rigorously when I write.
I try to identify these structural gaps when I read my draft as part of the editing process. The problem is that I am biased. Being the person who wrote the text, I am more attached to it and naturally think it is coherent and logical. Simply put, I understand it because I conceived it. ChatGPT, however, is unbiased. I can ask it to read my text and verify it understands it the way I meant it to be read. I don’t want ChatGPT to give me just the bottom line, though. It is essential to ask ChatGPT to create a skeleton of the text and maintain the order of ideas presented in it. I want it to show me the underlying model of the text I’ve written so I can compare it with my original design.
I use the following prompt:
Generate the skeleton of the following text. Summarize the key ideas in the order they are presented.
Here’s a fragment of what ChatGPT generated when I asked it to do so with a recent article I wrote:
I should note that this skeleton is shorter than the original model, but I decided to settle with this summary in this case. Being generated by AI, I regard it as representing how someone else might understand my article. I can read this AI-generated skeleton and see if it follows what I aimed to say (as captured in my original model). If I alluded to things I hadn’t meant to cover or failed to capture the logical order of my arguments, I could quickly identify such gaps by comparing the two skeletons.
Taking this idea a step further, I can ask ChatGPT to compare my original model with the actual skeleton of the text. To do that, I use the following prompt:
Compare this skeleton with the following:
Once fed with my original model, this simple prompt resulted in the following analysis:
This comparison enabled me to reach a greater resolution and identify finer-grained points I missed (or added) when writing the draft. It is not unlikely that some of these omissions and additions were intentional because I felt they were needed to make the text more fluent and engaging. But this kind of comparison allows me to reevaluate my former decisions, some of which are unconscious decisions I might now decide to change.
It’s important to emphasize that when applying this method, I don’t rely on ChatGPT to tell me whether the structure of the text is good or bad or even if it can be improved. This method assumes that I had a good model before I started writing, and I use my co-editor to verify that I followed it and nothing got lost in translation.
Method 2: Paragraph-Level Evaluation
We can do something similar at the paragraph level, although it will not be based on comparison. A good paragraph should be focused on one idea. We don’t want to develop more than one main idea in a single paragraph, as this may confuse and defocus the readers. One idea per paragraph helps us create structure and build (and later understand) the logical flow of the text. If my text is effective, the reader will be able to understand the fundamental idea of each paragraph, even if unconsciously, and how these ideas connect and evolve to create my argument.
Again, when I read my draft, I am attentive to this, and I try to verify I’ve done a good job limiting the scope of each paragraph to a single idea. But as we already established, I am biased. Knowing what I meant to say can make it too easy to read the intended meaning into the text. ChatGPT, however, is unbiased, and when I ask it to summarize each paragraph in one sentence, I get to see how someone other than me understands it.
Here’s the prompt I am using for that and a snapshot of what ChatGPT generated when I gave it the same article:
Summarize each paragraph in the following text in one sentence.
When I read this list of sentences, I can verify two things in parallel. The first is that each paragraph says what I meant to say and that it didn’t lose its focus or get misunderstood. The second is that when read in sequence, these key ideas form a good, logical argument that flows naturally.
This summary is obviously not as deep and rich as the article I wrote, but it captures its logical skeleton, this time with a one-to-one mapping to paragraphs. If I feel this summary is good, effective, and reflects what I aimed to convey, I can assume my text is logically structured and easy to follow. I don’t expect ChatGPT to tell me that but to provide an “objective” view of how my text will likely be understood. Eventually, it is up to me to decide whether this is good enough.
These methods are not designed to replace human-based editing when the stakes are higher. They are not designed to evaluate the text but to offer a perspective other than my own on how the text could be read. A human editor does much more than that; the dialogue with an editor is more profound and provides more insights. But as a quick solution for verifying I have managed to translate my thoughts into words, sentences, and paragraphs, this outsider’s perspective is priceless. It provides me with something I cannot do myself as effectively when self-editing. Unlike using ChatGPT before and during writing, this approach helps me improve my writing without being intrusive, without losing my authenticity, my original ideas, and my voice. This process does not guarantee that my text is good and effective, but I know it captures what I wished to say, and that is not something anyone who writes can dismiss.
So, Why Bother Writing and Reading?
If AI does such a good job summarizing my content, and anyone can use it to summarize what I write, why bother writing? Won’t most people prefer to read the 500-word version of this article outlined eloquently by ChatGPT instead of investing the time in reading 3,000 words? And if that is the case, why not publish just the key ideas to begin with? Why bother writing long-form content instead of just cutting to the chase and sharing the bottom line? Some would argue that my co-editing process is the ultimate proof that we can reduce the “waste” and write so concisely that there will be nothing more to strip down.
Concise summaries will never replace authentic writing and deep reading. Even if they can technically, we must avoid using this ability. The tool I use to validate my writing does not replace my need to write the full text. Writing is not just a way to communicate my ideas — it is how I form ideas. It is a way to explore my ideas deeper, challenge them, evolve them, and address nuances that can never be captured in a summary. Similarly, if I have done a fine job, the article you are reading provides you with more value than a plain list of key ideas. There is obviously room for concise, catchy messages, but they will rarely be enough to take an idea, deeply explore it, apply it in the real world, and make use of it.
I know it is a slippery slope. What starts with a way to validate my text against a model could end up as the standard way to consume content (if it hasn’t already). I genuinely hope this is not the case. I won’t replace watching a good movie with reading its synopsis, and I’d never replace a good novel with the skeleton of its plot. I love the texture of the complete work, the atmosphere, the sentences that sink in even if they will never be included in any summary, and how all this makes me feel and think. Non-fiction is obviously a different kind of dish, but I still want to experience it fully as a writer and as a reader.
If you follow my social posts, you will undoubtedly find short posts with no more than 200 words. My social presence has to follow the rules of the land. But I almost always revisit a topic I am interested in with a longer text, hoping that my audience will take the time to go beyond the catchphrases and sound bites of the short version.
I started with having a strong opinion against using AI in my writing process. I ended up experimenting with using it as a co-editor: an “entity” that can give me a glimpse of how my audience might understand my content. My goal was to keep my voice and ideas authentic, so I limited the use of ChatGPT to after the first draft was ready. This is not to say that this is the only valid use of it.
Using an AI sidekick to refine the model of the text before starting to write the draft is an interesting option I will probably explore further, as is conversing with it to challenge my ideas. I am not trying to shorten the time it takes me to write or publish more content than I can write myself. I am not trying to delegate my writing but to enrich it.
It is an ongoing experiment with only one basic rule: whatever tools and methods you use, don’t stop writing.