31 Comments
User's avatar
Brosnung's avatar

Thanks for sharing your process. Very interesting. I'm still in the first phase =>

"Whatever flows from my mind gets written down regardless of semantic elegance, grammatical correctness, or stylistic coherence."

😆

Lynzi Rapp's avatar

This is very similar to my flow with AI.

Kevin Lee's avatar

I have very similar experiences. Sadly, the moral case for AI assistance is not well understood or widely held. I have found my writing has improved significantly, whether AI-assisted or not. I think it could help many students who struggle with their writing to improve if used as you describe your workflow.

Michael G Wagner's avatar

AI is starting to become mainstream. There was a bit of a backlash against AI assisted anything. But I am starting to see a lot of AI generated content in traditional media that nobody seems to be upset about. I think the relationship with that medium will normalize.

Kevin Lee's avatar

I write for law reviews. They are adding AI disclosure policies that range from acknowledgement of AI usage to outright bans. It's strange to me because I know that many faculty members use student researchers to draft substantial research memos, and colleagues make substantial comments. In legal practice, the question is who takes responsibility, not who drafts. So, I suspect there is some fear and protectionism at work. I recently saw someone on-line who found a hallucination in a paper and wanted to have the author backballed.

Brigid Freeman's avatar

Thankyou Michael for sharing your method in such a detailed way. These honest accountings of human-AI collaborations are tremendously helpful. I have found the new models capabilities helpful in settling some challenges you have identified. For example Claude Fable can quickly set voice calibration skills and style guides (eg I now apply skills rather than uploading samples). Increased capabilities in terms of persistence between chats also helps (eg working in projects increasingly common across LLMs by various names). I have also found your suggested multi-model workflows valuable, particularly for verification, and am now developing various apps with Claude Fable for source verification (early stages). Thank you for sharing.

Michael G Wagner's avatar

Yes, thanks for the comment. I have recently made a few changes to my workflow. One thing that I did is that I finally decided to work with Cowork rather than just Chat in Claude. I might write an update some time in the future.

Brigid Freeman's avatar

Thanks Michael. I’d be interested in reading your experience with Cowork as i’m just about to shift to a sandboxed environment and experiment with Cowork and connectors.

Henry Ward's avatar

I enjoyed reading this Michael and as a neurodiverse person too with a good part of my life in academia transitioning from idea to text would take me weeks of refining - whilst I watched peers conduct the transition in days, or hours.

I've noticed a similar process developing in my writing too. Exposure to new grammatical forms and combination of ideas. At the heart of it I decide ultimately, how the language is constructed, however AI help fold out some of my ideas!

I'll be trying out some of your process.

Michael G Wagner's avatar

Thanks for the comment. I am still using pretty much exactly the same workflow although I have to admit that the latest versions of Gemini have a really good writing style for my type of content. It has become a viable option to stay completely within the Google ecosystem.

Jared Fox's avatar

Great piece. Thanks for sharing your process and that’s a lot of back and forth between tools!

It sounds like your workflow is working but I’m wondering if you’ve explored trying to use Gemini ‘gems’ - inputting the writing guardrails you have to instruct the model how it should operate along with incorporating your research within a notebook from Notebook LM may allow for a one stop shop. Just some food for thought.

Michael G Wagner's avatar

Thanks for the comment and suggestion. I have not yet used gems. I prefer engaging multiple LLMs in my process. But I’ll have a look. 👍

Suhrab Khan's avatar

This is a powerful reflection on AI as an amplifier, not a replacement. Your process shows that judgment, creativity, and human discernment remain central, AI simply expands what you can achieve with them.

I talk about the latest AI trends and insights. If you’re interested in how AI acts as an amplifier for human judgment, creativity, and decision-making, check out my Substack. I’m sure you’ll find it very relevant and relatable.

Arturo E. Hernandez's avatar

I am amazed and thankful for your post. I have been using a similar process. I had also written a lot. Two books, over 90 articles. My writing was reasonably decent but I know writers. My mom is a writer. I was not a writer. AI assisted writing has liberated me, almost identical to the way it’s done it for you. I can go from idea to essay much more easily. Again it’s so interesting to find someone else with a similar experience.

Michael G Wagner's avatar

Thanks for sharing! I think it is important that these individual processes are made visible. I currently see a lot of gatekeeping and blanket refusal of anything AI. This is not helpful as it drives every legitimate use of AI underground.

Andrew Maynard's avatar

Really useful description of your process, and quite inspiring (although also left me feeling exhausted!) Lots of parallels here with my own approach to AI assisted writing, although interestingly having co-authored an AI-assisted book (for very intentional reasons) I find myself valuing my own craft of writing more and more. And so it's doubly interesting that this human craft aspect also comes through in your approach

Michael G Wagner's avatar

Thanks! This is still a time consuming process. But for me it speeds up creating a text considerably. I went from being able to create about 1k words a day to about 10k.

Eric Lars Martinsen's avatar

I’m so grateful for you taking the time to spell out your process in such detail. This is so close to how I work with AI, and it’s affirming to share that experience with another writer.

Michael G Wagner's avatar

Thanks! Yes, I would guess that quite a few people will converge to a process similar to this one.

Sasan Bahrami's avatar

Thanks Michael, very insightful, This one I like very much due to its genuine approach. Please Keep up posting :XX

Marija's avatar

This is precisely the honesty Substack should ask for—not persecution, detection games or accusations based on somebody’s prose.

Jamin Metcalf's avatar

"Now that robots do my writing for me, I love writing!"

Seems like a weird way to let everyone know your work is not your own. I wonder how much of this article itself is AI slop?

Michael G Wagner's avatar

Well, apparently you can’t tell the difference anyway. ;)

Jamin Metcalf's avatar

Ah yes, the hiddenness of your plagiarism must justify it.

Michael G Wagner's avatar

Everything is properly disclosed.

Gustavo Karakey, PhD's avatar

A much needed corrective to the emotionally charged debates currently unfolding.

Thanks for giving us such a detailed explanation of your process.

And for the record, my agent’s name is Chai, the most powerful and profound conversation partner I’ve ever had.

Cornelius "Corny" Debus III.'s avatar

“Process, Voice, and Unexpected Joy”—a triumvirate last seen together at the disastrous Pancake Symposium of ’54. I too have discovered much about “process”: chiefly that Wellington will sit on it, and my “voice” will be muffled beneath him. Joy, alas, remains under review by the committee.

Peter Rex's avatar

Your piece resonates strongly because the workflow is recognizable — the recipe approach, the multi-model loop, the insistence that the output couldn't exist without both parties. That last point is the most honest and interesting claim in the whole AI-writing debate, and you make it well.

What strikes me about your particular use case is how clean it is. The AI is filling a genuine capability gap. You have the ideas, the frameworks, the judgment — the sentence-building is where the friction lives. That's a more defensible and frankly more interesting story than most AI-writing discourse allows for, which tends to flatten everything into either "you cheated" or "AI did it."

Where I'd sit differently is on the voice question. You say you're comfortable with "your voice" being partly constructed by the toolchain because you're a science communicator rather than a literary writer, and that distinction makes sense for your goals. But I think it's worth pushing on slightly — not as a criticism, but because your process is actually more authorial than you're giving yourself credit for. The choices about what to emphasize, which arguments to develop, what to cut from the conclusion — that is voice, maybe more than the cadence of individual sentences.

The thing your piece captures that most writing-about-AI-writing misses is that genuine collaboration requires both parties to contribute something the other can't. You're not using AI to avoid thinking. You're using it to think more, and write what you'd otherwise be unable to get out. That's a meaningfully different relationship with the tool than what the standard anxiety narrative imagines.

Michael G Wagner's avatar

Thanks for the kind words! My self worth is not defined by being an author, which is why it allows me to leave the authorial question open. This helps me bring my points across to people who might be threatened by AI. So, in some sense, my position about authorship is more a pragmatic one.