First Draft Mode: What writing with ChatGPT taught me about writing without it

A retrospective on using the chatbot to write a play

A bird taking flight after taking a dump on drama. I took this photo outside the Sydney Opera House.

I love starting projects.

There are few things in life that give me the same high as that initial whisper of an idea or spark of inspiration. The number of story ideas in my Notes app that has been steadily growing since high school will attest to that. Some of these ideas have made it past the brainstorming phase, and they now sit multiple pages into a first draft on my desktop. Others are not story ideas at all but just two lines of dialogue with no context whatsoever: “You’re out of your right mind! / Yes but I’m steadfastly in my left.” Most ideas sit somewhere in between, with several lines of notes on plot or characters, maybe even the odd mention of an overarching theme.

Like I said, I love starting projects. It’s finishing them that I struggle with.

I didn’t always have this problem. As a child, I could easily fill entire notebooks with stories. By my late teens, however, I was finding it harder to see a writing project through to its end. Of course, I always had a perfectly reasonable excuse. I should be studying, I should be working, Gossip Girl is on. You understand.

But at some point I needed to acknowledge that with each reasonable excuse for abandoning a project I was building a pattern. And once that had become clear, I did what every reasonable human being would do. I stopped writing entirely. My rationale was simple: If I don’t start a project in the first place, I couldn’t fail to finish it. Duh. This deception I had tricked myself into believing kept me well-fed for a long time.

Until in late 2022, when two things happened that would help me finally break that pattern. First, my friends and I started a new theater company and were looking for a first play to perform. And second, OpenAI released ChatGPT.

I think you can see where this is going. So before I lose some of you, a bit of expectation management.

I think it’s fair to say that many of us have grown allergic to the telltale signs of ChatGPT et al. in texts. So far-reaching is the influence of LLM-infused language that some of us might find ourselves saying delve, boast, and swift (not Taylor) more often than we did three years ago.

I’m not here to declare that “ChatGPT optimized my writing routine—here’s how!” I value human writing both as a creative outlet and an important tool that helps me think, so I don’t advocate offloading it to a chatbot.

But I also live with the conflicting knowledge that if it wasn’t for the timely release of ChatGPT, I might still be stuck in that chronic pattern of start-and-abandon. Because although ChatGPT didn’t “fix” anything—I still struggle to see projects through—I’m writing a lot more now. Without ChatGPT.

So this is my attempt to piece together how I think it helped me while I was working on that play. And why I don’t plan on using it that extensively ever again.

Finally, given the topic, I want to emphasize that no part of this article was created by or with the help of LLMs. As one renowned mathematician’s recent conference slides state: “All em-dashes in [this article] were human-generated.”

---

Break the habit, break a leg

In that first week after ChatGPT’s release, it was already being heralded—and cursed—as a tool that would help people do more things better and faster, including but not limited to creative endeavors. I know it feels like a lifetime ago, but in ye olde days of late 2022, ChatGPT was still a shiny new toy, “AI slop” hadn’t entered our collective vocabulary, and we had yet to find out what these LLMs were and were not capable of.

So in December 2022, I started working on a play with ChatGPT to answer the question “Can an AI write a play?” as much as “Can Olivia finally finish a damn project?”

To test ChatGPT’s limits, I wanted it to generate as much of the actual script as possible. Ideally, everything. I would guide it and ultimately be responsible for every word in the script, yes, but I was planning on using it for every step of the way: Outlining, plot development, actual drafting etc.

The seed for the play came from French playwright Yasmina Reza’s 1994 masterpiece Art. It centers around three long-time friends, Serge, Marc, and Yvan. When Serge buys an expensive, modern painting—a white canvas with white diagonal lines—Marc openly disapproves of the purchase: “You paid two hundred thousand francs for this shit?” (pardon his French). Yvan is caught in the middle, trying to mediate between the two. In hilarious ways, the play demonstrates how fragile the foundations of relationships can be.

The play has stayed with me ever since reading it in high school. I saw a place for this trio of characters in the context of a technology like ChatGPT that elicited such mixed responses from the public: enthusiastic, skeptical but tempted, and horrified. In fact, I felt that the only story that would make sense given my comprehensive use of the chatbot was one in which human characters struggle to embrace such a technology. I wanted to see how ChatGPT would depict an AI character based on itself and how it would suggest navigating a world in which it is threatening to encroach on a uniquely human activity: creativity.

From December 2022 to June 2023, I worked on this play with ChatGPT. I won’t be detailing the writing process and its many hiccups, but here’s a brief summary.

I completed a first, horrifyingly bad draft in three weeks, by Christmas 2022. Every word was generated in some way by ChatGPT, though I did at times have to force it to parrot what I was inputting when it wouldn’t give me what I wanted.

A shift in strategy came during the editing phase. I was finding it less useful to rely uniquely on ChatGPT for rewrites after I had received human feedback and had a better idea of what the story should actually be. I still used the chatbot, but I abandoned the idea that it had to generate every word.

Then, in June 2023, with a wonderful team of creatives, we put on a small showcase and performed our play, Assistant. It was magical and fun. And I was so satisfied. I’d done it. I’d seen something through.

---

Productivity = learning?

For nearly two years after we’d wrapped up production, that was enough. I’d broken the curse. I’d written a play!

It wasn’t until I was sitting in the audience of a 2025 conference about AI’s impact on society that I revisited my conclusion. I was listening to a talk given by Manu Kapur, Professor for Learning Sciences and Higher Education at ETH Zurich. In this talk, he made the distinction between productivity and learning when using tools like ChatGPT. Essentially, his point was, yes, LLMs can help us increase our output volume, but potentially at the cost of learning.

It seems obvious to think of it in those terms now. But before that day, I’d only ever focused on the first part—“Wow, I actually finished a play”—and not the second—“But did it make me a better writer?” I’d always just assumed that it had. After all, I’m a firm believer in learning by doing. I was doing the doing, so wouldn’t the learning automatically follow? What I believe I’d missed was the more subtle question of what exactly I was doing.

---

Quiet in the peanut gallery: First Draft Mode

Without a doubt, my biggest struggle when writing is completing a first draft. Attempting to fill an empty page always fills me with dread. Somewhere along the way from being an unburdened ten-year-old writing stories about snakes with glasses and enchanted cherry trees, I’d lost the ability to write anything without the immediate urge to edit. Editing is slow, slow is frustrating, and frustrated me doesn’t want to write. Especially not when I’m stuck in the senseless activity of revising the single page of text I’ve managed to conjure.

I think this was—and to some extent still is—the crux of my dysfunctional writing pattern. I get so focused on the details and wanting the text to sound good that I get paralyzed. I want nothing more than for that inner critic to shut up, but she’s annoyingly persistent.

In short, what I’d been unsuccessfully chasing in those years since my childhood was a way back into First Draft Mode.

The best way I can describe First Draft Mode is as a playful and curious state of mind. A bit like brainstorming in prose. The goal of First Draft Mode is not high-quality writing but exploration. It’s about forward motion. Many children excel at this. And while as a child, I probably never managed to go beyond First Draft Mode, as an adult I rarely manage to access it.

But as I was developing the first draft of the play with ChatGPT, a feeling began to sneak up on me. I was genuinely enjoying the writing process (if you can call it that), embracing the discovery without second-guessing at every turn. And it was a joy I recognized. I’d found my way back.

So what happened? What had ChatGPT “unlocked” that had remained inaccessible to me for so long?

I think a key aspect of getting into First Draft Mode—at least for me—is admitting to yourself that what you’re writing now is not going to end up in the final text. Typically, that’s easier said than done. But when I was working with ChatGPT, it happened very naturally.

As much as people criticize those who use ChatGPT for “writing”, we seem to forget that prompting is also writing. Don’t get me wrong, I don’t want to confuse drafting something yourself with asking an LLM to “write the next scene.”

But I found that prompting ChatGPT is what helped me get into First Draft Mode. I didn’t put much thought into how I described what I wanted. I knew that I wasn’t doing the writing; ChatGPT would translate it. But in instructing it, I was often inadvertently doing the very thing I was asking of ChatGPT. So when I compare the first draft to my prompts, I find quite a few instances where my prompt ended up in the draft verbatim.

I think it’s the same mechanism that’s at play when you tell your friend about what you’re working on, and something about the casual nature of the exchange, or the enthusiasm they mirror back at you keeps you talking, and at some point they tell you, “You should write this down.” It’s that.

“Well,” you might ask, “if you were able to put into words what you wanted to say on your own, why not skip the intermediary?” That’s the silly bit, isn’t it? Whether I’m writing the words for ChatGPT or directly putting them in the script, it’s ultimately the same thing, no? No. I don’t think so.

Apart from the fact that the whole point of the project was to have ChatGPT write the script, I believe the underlying psychological processes giving rise to what looks like the same behavior—typing (almost identical) words on a keyboard—are distinct.

I apologize to any memory researchers if this is a poor analogy, but hear me out. When I write, it often feels like I’m trying really hard to remember the name of a movie. I can name words that sound similar—Iridescent? Inflection? Indecent?—maybe even recall the intonation—“da-DEE-da”—but not the actual name. The harder I try, the more the name seems to elude my mnemonic grasp.

But the instant I give up on ever hoping to remember the name and turn my attention to something else... Inception!

That’s what working on a first draft feels like on my own vs. how it felt like with ChatGPT. It’s not that prompting ChatGPT required no attention or thought whatsoever; I guess that’s where my analogy falls short. But my focus was on something else. Instead of trying to sound good, my goal was to convey my intention in an understandable way. Maybe that’s all that prompting ChatGPT was, and maybe that’s why it feels like explaining to someone, “Look, here’s what I’m trying to say.”

The problem is that for some reason when I write directly keyboard-to-page, I can’t pretend that I’m not trying to find the right words or that I’m simply talking to someone else. Ironically, trying to get into First Draft Mode can be the very thing that shuts me out.

---

FDM sans LLM

Now the lesson here decidedly is not “Write all your first drafts with ChatGPT.” It’s to determine what your version of First Draft Mode is, recognize when you’re in it, and recreate the circumstances that best let you access it and stay there.

For me, First Draft Mode remains an elusive creature to this day. More often than not, I find myself trying not to try too hard. Nonetheless, my experience with ChatGPT has taught me what it is that I seek when I’m drafting, and I’ve tried out some techniques that seem to help.

First, the example above of talking to someone was not just another allegory. I find it extremely useful. Simply verbalizing to someone the knot I’m trying to disentangle can bring clarity. The other person doesn’t even need to say much; it’s their presence that does the trick, the fact that I’m directing my speech at someone other than myself.

Other times, sticking to a strict writing time limit tempers that knee-jerk reaction to immediately rewrite. Parkinson’s law states that a task expands to fill the amount of time available, and writing is no exception. If I know I have all day to work on a text, you better believe I will take all day, with a potentially very unsatisfying amount of progress. But if I know I only have three hours, it’s a bit easier to refocus and prioritize when I find myself getting distracted by intriguing but irrelevant rabbit holes.

Most recently, I’ve noticed that writing in the form of comments also helps. I write in Markdown, so when I comment out text it appears light grey instead of black. The grey color makes my text seem impermanent, which oddly helps me jot down entire paragraphs without worrying too much about how nice the words look together. “I’m just a temporary comment,” the text seems to convey, “I don’t need to sound good, I just need to get the message across.”

If we’re being honest, all of these techniques are just little tricks of the mind. But if they work, why look a gift horse in the mouth? A part of me wonders if these strategies are really all that different from using ChatGPT… But then I recall the many times while working on the play that I felt tired or frustrated and how easy it was in those moments to just resignedly accept whatever ChatGPT suggested. It kept me going, but at the cost of surrendering my most valuable asset: my thinking. And it was only after the first draft had been completed and I got asked why I had written this or that scene in a particular way that I realized just how little I understood about the text that “I” had authored. It was then that I needed to acknowledge that this particular strategy of getting into First Draft Mode was perhaps not a gift horse but a horse of the Trojan sort.

---

The real lesson or the wrong lesson?

When I started this project in December 2022, I set out to answer two questions: Can an AI write a play? And can Olivia finish a project?

In my experience, I found that, sure, ChatGPT could write a play, but not one that I wanted to see on stage. And yes, I did indeed finish this project, but relying on ChatGPT to the extent that I did gave me a false sense of confidence in what I’d achieved.

So now, more than three years later, I believe my two original questions obscured a more important one. For every text that is written “by AI”, there is a human behind the scenes, guiding the LLM with every prompt. So perhaps the more pertinent question isn’t about the LLM’s ability to write or the human’s ability to see a project through but instead what exactly the human using the LLM is doing during the creation of the text. Are they simply producing or are they also learning, thinking?

While the lesson I detail above is real and the experience reignited my writing spark, I also found that I had indeed fallen into the “productivity = learning” trap that Manu Kapur was talking about.

Because had someone removed my access to ChatGPT while I was working on the play, I would’ve been just as paralyzed as if I was writing it on my own. Probably more so. And that tells me I wasn’t learning what I thought I was. Instead, what I was learning was to rely on ChatGPT too much. Yes, writing a play with an LLM was the point of the exercise, but it also left me instinctively wanting to reach for the immediate relief of ChatGPT’s answers for far longer than I care to admit.

I had to actively train myself to resist that urge again, to work through difficult text passages using only my own, frustratingly slow thought process. And I only bothered to do it because I care enough about writing to not want that cognitive muscle to atrophy. I don’t feel the same way about navigating without Google Maps, so you can imagine how bad my sense of direction is.

So while I relish the moments I can access First Draft Mode, I’ve also developed a new-found appreciation for friction. Writing with ease and flow is marvelous, but I’ve realized that I can’t complete a first draft in First Draft Mode alone. Often, it’s when I manage to push through a particularly frustrating passage that a text begins to take shape. You can’t really have a breakthrough if there’s nothing to break—no resistance, no doubt, no stagnancy.

In retrospect, perhaps the real “lesson” I mistakenly internalized while using ChatGPT was that writing should feel smooth all the time. That feeling stuck or frustrated is a sign that I’m doing something wrong. I don’t believe that anymore. Today, I’m still figuring out the right balance between flow and friction, trying to unlearn one aspect of my experience with ChatGPT while holding on to another.

My wishlist of writing projects is still growing at a faster rate than I can check off. I hope that never changes. What a luxury to have more ideas than I can tackle! But what has changed over the last months is having more faith that I’ll actually manage to finish some of them. This article is one example. More are in the works. And I hope some of you will stay along for the ride.

---