When I began my advertising career over 20 years ago, I was desperate to impress my Creative Director with my long-form writing skills for a pitch. I don’t even remember what exactly my idea was, but I remember writing the shit out of it, trying to prove how clever I was with every single line, using 50-dollar words and complex sentence structures everywhere. When I showed it to him, he only had one thing to say — “Read it aloud to me.”

Standing next to his desk and reading out my overengineered ad copy for The Economist, I knew by the third line that it didn’t feel right. Every sentence was a strain to get through, and none of it sounded like me. I went back to the drawing board, and that “long copy” print ad eventually became a single headline on that classic red background.

That was the first lesson I learned about writing. If it doesn’t feel right or feels unnatural or like a lot when you read your copy aloud, you need to change your copy. It’s something that I still do today, long after my copywriting days are behind me. Whether my writing has gotten any better is, of course, debatable.

Out with the old…

I used to think the trick to writing with AI was to prompt it so well that it could write the first draft for me, and then I could go in and refine that draft.

If you’re reading this in 2026, there’s a strong chance this is how you’re using AI to write today. I know the stigma around AI writing (or AI in creative work, in general) is still very strong. We don’t want to publicly admit we’re using AI even though we all are now. So in the shadows, we try to “de-AI” our work by removing the tells.

But that rarely works — because the tells aren’t what make or break the writing. It’s a sense you have when you read an AI-written piece, like it’s missing something that has nothing to do with em-dashes or contrast-framing.

Not so long ago, I’d begin my writing process by brainstorming with Claude. It was great at research; it seemed to know almost everything; it could search the web for things it might not have in its training data; and it would often push back if I was going down the wrong path. Brainstorm completed, I would then ask Claude to “generate a blog post about everything we discussed and include my (our?) perspective on the subject based on our conversation”.

The output would be meh, but full of what we all now know as “AI tells”. And so nervous was I about the judgment of the famously kind people on the internet that I would go through the draft and remove all the “AI-isms”. I’ve heard of folks doing this now — spending 20 minutes prompting Claude to write them something and then spending the next few days making that draft sound less AI-like.

For me, going in and removing every evidence of AI writing would only make the final piece sound even less like me — or human at all, frankly.

So I realised my old process wasn’t going to cut it anymore. I’ve always loved writing, and this shitty workflow was sucking the joy out of it for me.

…In with the new

Never one to dwell on shitty processes, I took off on a quest to find a better system that would take all the advantages that AI could offer my writing, without losing my voice. The process I follow now feels much more authentic to me, and AI is still very much a part of it.

I still begin with the brainstorm, because if anything, the models have only gotten smarter and more resourceful. To take the brainstorm to another level, though, I introduced an “interview” format. I begin the conversation by giving Claude the basic idea I want to explore, then ask it to interview me in depth about the idea.

I tell it to cover every possible angle in that interview, get deep into the weeds, push back against the assumptions I might be making, and call me out whenever I’m going down a path that doesn’t work. I end up learning so much more in this way. The goal, of course, is to get me to think critically — a lost art in the age of AI — and completely dissect and understand my idea, so I get to what I want to say, not how I should say it.

The act of writing down your thoughts — even the ones that never make it into the final piece — provides a clarity that you can’t get by simply ruminating on an idea (or prompting AI). For me, so much clarity comes from the back-and-forth of writing my responses to Claude. I come into those interviews with one seed of an idea and lots of thoughts, and I always leave with a pretty clear direction in my head about where I want that piece to go. Writing my answers to Claude’s interview questions is where my thinking gets focused and specific.

Once I’m satisfied with the interview process and itching to put it all down on paper, I ask Claude for one more thing — an outline. I have a tendency to ramble on and on in my excitement when I write. It’s a problem I’ve always had, and so a structure gives me guardrails that I can stay within. “I don’t need any written copy,” I always have to remind the ever-eager Claude. All I want is an overarching flow that will work for this piece.

And so Claude provides me with an outline that I can then use as a reference while I put my thoughts down in writing. For writing, I’ll use Mimir’s Notes app or iA Writer, which offers a perfect, minimalist writing UX with zero distractions.

Now that I have the outline and some structure, I can write my ass off and only refer to the outline every once in a while to make sure I don’t veer off course. I can stay in the flow as I write, and I can write like me, just like the good old days.

Of course, my curse is that I still have a tendency to ramble on. So once I’m done with my first draft, I pop back into the same AI conversation and ask Claude to review my piece and assign it a score out of 10. Then I ask it to explain why it chose to score the piece anything less than 8.5.

Again, I specifically tell it I don’t want it to rewrite anything for me, just give me pointers on how to tighten up some bits, or fix long-winded sentences, or re-arrange certain paragraphs for better flow. Like an editor’s first pass, basically (more on this in a minute). And none of the feedback has anything to do with “AI tells”, just good storytelling. The actual fixes I go back into the document and make on my own.

Wife with benefits

Once I’m more or less happy with my (now second) draft, I take it to the final review stage — my wife, Karanjeet, an editor with two decades of experience. She’s not ruthless, but she’ll call out all my unnecessary fluffiness. She’ll tell me when something is funny, but more importantly, when something isn’t as funny as I think it is. She knows my voice, and she doesn’t hold back when I come off sounding inauthentic. She has the kind of editorial taste that comes from two decades of reading, writing, editing, and murdering weak prose. Claude and I are not there yet.

Her feedback is worth everything, and I clinically implement it all before I consider the piece “done”.

AI is still a big part of my writing process, but the writing — the part I’ve always enjoyed — is untouched by it. What’s changed is the nature of the interaction itself — I’ve gone from prompting Claude to having Claude prompt me, so that I can think more clearly about what I want to say.

I know the joke is on me because I probably end up writing a lot more through my answers to Claude’s interview questions. But this process helps me get to the final draft, as messy as it may be.

And I’ll be damned if I don’t enjoy my slow, rambling, sometimes incoherent writing process. Isn’t that the whole point?

Read on Substack ↗