My AI writing coach

Last week, I sent the manuscript of the first twenty-two of twenty-seven chapters of my memoir to an editor. It is what I hope will prove to be a good first draft. The editor is someone I trust to do a thorough line edit as well as provide creative feedback.

What surprised me most was not that I had finished twenty-two chapters. It was that this was probably the strongest first draft I have written in all my years as a content developer. Granted, writing a memoir is another kettle of fish, and I still have to hear back from the editor; they might see it otherwise.

What I am anxious to find out is whether working with an AI writing coach has truly paid off in this initial phase, as far as the quality of output. What is indisputable is that the experience of using an agent changed how I approached my writing. I had more time to write and less time to brood.

After reading Dr. Philippa Hartman’s article, The “Cognitive Offloading” Paradox, I was fascinated to discover that many of the ways I had instinctively been working with AI closely matched what she describes as “committed, strategic offloading,” a predictor of positive, even transformative learning.

Hartman discusses six principles to follow.

Principle 1: Offload to AI substantially, or not at all.

When it comes to writing, up until now, I have used ChatGPT, together with Grammarly and Claude, to edit my spelling and grammar.

I tried, unsuccessfully, to give ChatGPT more complicated editing tasks to do, but found this dissatisfying. This is partially because the AI tools introduced certain language patterns I don’t like at all, such as em dashes, one-sentence paragraphs, and an overuse of adjectives. As well, at the Pro project level, ChatGPT only seemed able to process tasks involving documents up to ten pages long. If, for instance, I asked it to remove the time lines and clean up the text of a podcast transcript longer than fifteen pages, the results were suboptimal: missing sections, cut-off sentences, and hallucinations. Any discussion with ChatGPT about how to alter my prompts to achieve consistently reliable results failed.

Admittedly, the problem might have been on my side, but I didn’t want to spend more time trying to get the model to produce better outcomes. I also didn’t want to run the risk of my book reading as though AI had stomped all over it, leaving the language flat and the imagery matted. I found two friends with the required expertise to do the editing. Human writing for human readers.

Having dyslexia meant I still relied on Grammarly as a spell checker.

The one area I was excited to work with AI was as a writing coach. This is the first time I have written a book, and I needed guidance. After taking numerous online courses in creative nonfiction writing, I felt more confused than motivated. I needed a writing coach to help me with structure and form. Fortunately, a good friend worked with me and programmed an AI agent as a writing coach.

Principle 2: Frame AI as a partner, not a tool.

My writing coach was not so much a partner as an informed expert, possessing a skill set I needed. I’ve probably read dozens and dozens of memoirs in my life, some of them among my favourites, but I knew nothing about what makes a memoir engaging and appealing to today’s readers. For this, I needed feedback from someone or something that could review my work from a publisher’s perspective, and also had the ability to support my writing process without creative intrusion.

What I was seeking was a coach who would listen to my questions and answer them with specificity. I was also hoping the coach could help me distance myself from what I was writing, so that I could make changes out of conviction that they would add cohesiveness and clarity, rather than out of desperation or insecurity.

The final selling point was the ability to converse with this coach about anything that came to mind, and at times convenient for me. I didn’t need to feel beholden to this coach, nor was I concerned about testing the boundaries of the relationship. This freedom of when, where, and what of our so-called partnership was joyful.

Principle 3: Build verification into the workflow, not the preamble.

One of the core tasks I assigned the coach to do was to analyse each chapter for what was working, what needed improvement, and what was missing. The book has four parts with varying numbers of chapters, so I asked the agent to make sure each chapter supported the central theme of its respective part and that each part was identifiable from the others.

The coach would write an analysis of what worked and what didn’t in the version I submitted of each chapter. As I read through the response, I registered whether I “liked” what was being said or not. Everything I liked went into a “maybe true” bucket. I didn’t feel I could necessarily trust the positive comments as being true, but neither could I dismiss them. I gave myself permission to let them stand for the moment. It was reassuring to hear these comments and soothed my doubts, which enabled me to continue writing with momentum.

Then there was the feedback I didn’t like reading. This tended to fall into two other buckets. The first was the “whatever” bucket. These were nice-to-have suggestions, but not for this round of revisions. Advice about how I could build an even higher sandcastle. Things that would take too much time and not substantially improve the current version.

There were also comments so unhelpful that I questioned the feasibility of continuing to use the AI coach. I treated these as anomalies and discarded them almost immediately.

Then there were comments or ideas I didn’t like or were resistant to, but that provoked a response that made me want to explore them further. One suggestion was that I move an entire chapter to a different part of the book. My immediate reaction was, “Absolutely not.” Two days later, I realised the suggestion exposed a weakness in the structure, even though I ultimately solved the problem differently. Those were the conversations that taught me the most.

If a suggestion made me pause, I’d ask the agent to give me some examples and ask follow-up questions. What would be an alternative to what I’m doing now? Why is this suggestion useful? What will it improve? Then a conversation would ensue. Not between adversaries, but between two parties working toward a common outcome. I was well aware that the AI agent was not a committed partner, but the thread of discussion often felt as though it was.

Principle 4: Make the learner think first, AI second.

In the creative process of writing a book, I am the learner, so it is up to me to do the work, both the heavy lifting and the light touches. Using the coach reminds me of something my father used to say: “The best thing about buying your first car is paying for it.” As someone who has never owned a car, I do not have direct experience with this, yet I do know the satisfaction of taking a piece of advice from the agent, rewriting part of a chapter or expanding on an idea, and recognising that the piece is now better.

The coach could make suggestions endlessly, but until I wrestled with those suggestions myself, nothing had really been learned. The learning happened in the rewriting, not in reading the feedback. Every decision remained mine. Sometimes I accepted a suggestion wholeheartedly. Sometimes I rejected it immediately. More often, one suggestion led me to an entirely different solution that neither the coach nor I had imagined at the outset.

All learning begins with baby steps. Sometimes I felt very wobbly on my legs, the way babies learning to walk resemble drunken sailors. I would take a group of suggestions and begin revising, only to abandon the whole approach halfway through. Since I was using Scrivener as my writing platform, wobbling back to the starting line was no problem.

I never tracked changes. Instead, I rewrote the text, paused for a day or two, and then read the chapter again. I asked myself whether it now flowed smoothly and whether the ideas followed one another more naturally. If not, I returned to the previous version and tried another approach. If it felt right, I moved on to the next chapter and let the revised one rest for a while before returning to it once again with fresh eyes.

That ownership turned out to be essential. The coach could accelerate my thinking, but it could never replace it.

Principle 5: Use AI to identify errors, not fix them.

I love this use of the AI agent most of all. The best feedback I receive is when I ask, “What improvements or changes should I make? What is missing?” Especially the latter question exposed obvious oversights or prodded tender blind spots. Since the information came from a machine, nothing it said felt personal. I only had to ask myself whether what it said made sense or not. Nothing more.

There were two aspects of identifying errors where my AI coach fell short. One was that it had no concept of “good enough.” If I asked it whether the current version of a chapter was good enough as a first draft, whether it had reached a comparable level to the other chapters, the answer was always yes-and-no. There were always things that still needed fixing.

The other shortcoming was that no text was considered beyond saving. No matter how weak or incomprehensible a piece was, it apparently merited practical suggestions for how to improve it. Of course, the coach never said it was a mess.

It happened twice that I worked and worked on a chapter, allowing the AI agent to lead me down the garden path, only to realise that all my efforts were useless. I had to throw the chapters away and start again. Somehow, I wasn’t willing to give up as long as the agent kept making suggestions. This was a valuable lesson to learn. Like the Monty Python sketch about the Black Plague, my text was piled onto the wagon headed for a mass burial, and the agent was singing, “Not dead yet!”

Principle 6: Assess without the scaffolding.

This is where I am now, waiting for my editor’s verdict. If the feedback is largely positive, I will send the manuscript to the second editor. I fully expect many changes, but I also hope they will confirm that the heart of the book is there.

If that happens, I will decide whether to continue using the model in the next round or whether it has already served its purpose.

If the feedback reveals more fundamental problems, I will have to examine not only the manuscript but also the way I used AI throughout the process. That may bruise my ego, but learning something new often does.

In the end, Hartman’s article may prove to be right. The real question was never whether AI could write my book. It couldn’t. The question was whether it could help me become a better writer. That is something only another human being, my editor, can now help me answer.