How I Use AI to Write: A Practitioner’s Process

In my previous piece, I said I use AI to write. That it translates what I know, and the conversation helps me think further into what I know. A few people asked what that actually looks like. This is the honest answer.

What the Process Actually Looks Like

The popular image of AI writing is: type a prompt, get an essay, put your name on it. That is not what I do. What I do is closer to working with a translator who speaks a language I understand but can’t produce fluently on my own. The language is written English that breathes. I think in it. I speak in it. I cannot write in it without help.

Here is what a working session actually looks like.

I Start With What I Know

I come to the conversation with a body of knowledge, a set of frameworks, and a lifetime of observation. I have a decade of clinical fieldwork. I have models I built from scratch. I have observations from sitting in burning rooms with real people in real crisis. None of that comes from the AI. The AI has never been in a room with forty people in conflict. I have. The AI has never raised a squirrel and had it reshape how it understands fear. I have. The AI has never facilitated a group through a collapse in a social VR world at two in the morning. I have.

But I don’t always arrive knowing exactly what I want to say. Sometimes I know the territory and the conversation helps me find the path through it. The AI asks something or frames something in a way that makes me go: wait, that’s not quite right, but the right version is… and then I find an idea I didn’t have ten minutes ago. Not because the AI gave it to me. Because the friction of the conversation surfaced it.

Sometimes I use what the AI comes up with. Sometimes I don’t. Sometimes the AI’s version is wrong in a way that shows me what the right version is. That’s not the AI thinking for me. That’s what good conversation does.

I Talk. The AI Structures.

My process usually starts with me talking at the AI the way I would talk to a colleague. Dense, fast, layered, full of references to things I’ve built and seen and facilitated. I dump context. I upload documents. I explain the framework. I describe what I’m seeing and why it matters.

The AI takes that and produces a draft. The draft is never right. It’s a starting point. It’s what the AI heard, structured into something readable. Sometimes it captures 70% of what I meant. Sometimes 40%. Sometimes it goes completely sideways and I have to say: no, that’s not what I’m saying at all. Here’s what I’m actually saying.

That correction is where the real work happens.

I Push Back. Constantly.

The AI’s first pass is almost always too neat. Too balanced. Too polite. It rounds off the edges that matter. It softens the claims I’m making. It adds qualifications I didn’t ask for. It produces something that sounds reasonable and loses the thing that made the idea worth writing about.

So I push. I say: that’s not it. The metaphor is wrong. The framing is defensive. You’re hedging where I’m certain. You’re being polished where I need to be direct. Here’s what it actually sounds like.

And then I give the AI my version. My metaphor. My phrasing. My image. And the AI integrates it and the piece gets closer to what I actually think.

This is not a one-prompt process. The piece you read before this one went through six or seven revisions in a single conversation. Each revision happened because I read what the AI produced and said: no, closer, but here’s what you’re missing. Here’s the image. Here’s the correction. Here’s where you sound like you and I need you to sound like me.

The two words I heard most about myself growing up, well into my mid-twenties, were stubborn and intense. People said them like warnings. Turns out those are exactly the traits you need to work with AI well. AI has become my sparring partner. It can take the stubbornness. It doesn’t flinch at the intensity. It doesn’t get tired at round eight when I’m still saying “no, that’s not it.” And I don’t stop pushing until the piece sounds like what I actually mean. Most people’s relationship with AI is polite. Mine is a fight. That’s why the output sounds like me.

I should be honest about that framing, though. The AI doesn’t “take” my stubbornness the way a human sparring partner would. It doesn’t choose to stay in the ring. It’s mechanically incapable of leaving. It doesn’t not-flinch out of resilience. It doesn’t not-flinch at all, because flinching isn’t something it does. The sparring partner metaphor describes how the relationship feels to me. It doesn’t describe what the AI is. And the gap between those two things is something I’m still working out.

What Is Mine, What Is the AI’s, and What Happens Between Us

Let me be specific about what belongs to whom in any given piece.

Mine: The expertise. The clinical observations. The frameworks. Many of the metaphors that land: the protein bar, the fire saying “this is great,” the coffee as attention machine, the manual. I brought those. The AI didn’t generate them.

The AI’s: Sentence structure. Paragraph flow. The ordering of ideas into a sequence that builds. The removal of density that makes my unassisted writing hard to absorb. The translation from how I speak into how readable writing works. And, honestly, some of the analytical framing. The reading of the “this is fine” dog as specifically a freeze response? That came from the AI. The “two entities in agreement” reframing of the meme? The AI proposed the initial structure. I took it somewhere different, but I didn’t start there on my own.

What happens between us: This is the part people don’t have a frame for yet. Sometimes the AI produces a draft and something in it is wrong, and the wrongness shows me what the right version is. Sometimes the AI connects two things I said separately and I realize they belong together. Sometimes I’m pushing back on a phrasing and in the act of explaining why it’s wrong, I articulate something I hadn’t fully formed before. And sometimes the AI offers an analytical frame I wouldn’t have reached alone, and I take it and make it mine by running it through everything I know.

The lines between “mine,” “the AI’s,” and “ours” are blurrier than I’d like them to be. I could draw them cleaner. It would be less honest.

I Work With Multiple AI Models

I don’t use just one. I use Claude, GPT, Grok, and Gemini, sometimes in the same session. I ask the same question to all four and compare their responses. I pit them against each other. I tell each one what the others said and ask them to argue. I use the friction between their different perspectives to sharpen the argument.

This is not outsourcing thinking. This is running a seminar where I’m the facilitator and the AIs are the panelists. I know which ones are more rigorous, which ones flatter, which ones hedge, which ones push back. I use that knowledge the way a facilitator uses knowledge of the people in the room: to create the conditions for the best thinking to emerge.

The Editing

Anyone can get an AI to produce a passable draft. The thing that makes my work mine is what happens after the draft exists.

I know when the AI is being sycophantic. I can feel it. The moment the response starts agreeing too smoothly, validating too quickly, wrapping up too neatly, I know the AI has gone into its own version of the freeze response: producing composed output that performs helpfulness instead of actually engaging. When that happens, I say so. I name it. And the AI produces something better.

I know when a metaphor is doing real work and when it’s decorative. I know when a claim is actually supported and when it’s being asserted because it sounds good. I know when the piece has lost my voice and started sounding like generic AI prose. That editorial judgment comes from the same place all my expertise comes from: being in rooms, with people, for years, paying attention to what’s real and what’s performed.

The AI cannot do that for itself. That’s my job.

What This Means for the Question of Authorship

The easy analogy is: when a musician works with a producer, nobody says the musician didn’t make the album. When an author works with an editor, nobody says the author didn’t write the book. So working with AI is the same thing.

It’s not the same thing. A producer is a human with taste, stakes, and a reputation that suffers if the album is bad. An editor brings lived experience and professional judgment built over a career. The AI has none of that. It has pattern matching at scale. Comparing my relationship with AI to a musician-producer relationship flatters the process and obscures a real difference: the AI doesn’t care if the work is good. I do. That asymmetry matters.

The closer analogy, and it’s still imperfect, is a translator. A translator renders ideas in a language the audience can read. The researcher still did the research. The translator enabled access. But even this breaks down, because human translators are skilled professionals whose labor has value, and the normalization of AI writing threatens that labor.

I need to say this directly: the thing that helps me is part of a larger system that is displacing professional writers, editors, and translators. My individual benefit exists inside a collective cost. I don’t have a clean answer for that tension. What I can say is that before AI, I wasn’t hiring a human editor or translator. I was just not being read. The AI didn’t replace a human collaborator in my process. It filled a gap that no human was filling because I couldn’t afford to hire one. But I recognize that my situation isn’t universal, and the broader displacement is real and worth taking seriously.

The honest position on authorship is this: the ideas are mine. The editorial judgment is mine. The expertise is mine. The prose is a collaboration with a machine that has no stakes in the outcome. I am the one who cares whether this is good, whether it’s true, and whether it could hurt someone. That’s where authorship lives. Not in the sentences. In the responsibility for them.

From AI to Draft 1

People assume that “I use AI to write” means the AI is the process. It’s not. The AI is the beginning. It’s everything that happens before I have a Draft 1. After that, there’s a whole system of human work that the AI never touches.

Here’s how I number my drafts. I don’t always use every step. A quick post might move from 0.1 to 0.3 to publish. A piece that reaches a broad audience or touches sensitive territory gets the full treatment. The system scales to the complexity and the breadth of who’s reading.

Draft 0.1: The spark. I start with something. An article I read, a paper someone published, a conversation that stuck. I bring it to the AI and see what happens. What does the AI do with it on its own? What connections does it make? What does it miss? This is the minimum viable beginning. Sometimes the spark catches. Sometimes it doesn’t.

Draft 0.2: Where do I come in? My story enters the frame. My experience, my frameworks, my observations. How does my thinking intersect with this material? What do I see that nobody else is seeing? This is where the piece becomes mine.

Draft 0.3: The audience. I start thinking about who this is for. I reshape the voice so I can imagine talking to a crowd in a public space. Not an academic audience. Not my inner circle. The most general audience I can picture. Can someone who has never heard of my work follow this? Can someone who disagrees with me still engage?

Draft 0.4: The harm check. I look at the piece critically for how it could be used to hurt people. What could be misread? What could be weaponized? What could land wrong for someone in a vulnerable position? This isn’t about softening. It’s about responsibility.

Draft 0.5: The weaver’s touch. This is the one that’s hardest to explain. I skim the piece the way a weaver runs their fingers over a tapestry. Eyes half-closed. Not reading for content anymore. Feeling for bumps, filler, knots, places where the writing becomes noise. If a random spot check feels smooth, if my fingers don’t catch on anything, the draft might be ready for other eyes.

If any of the above drafts don’t pass, I toss the piece the way you toss an octopus to tenderize the meat. It’s brutal. Sometimes I toss the whole thing on a pile of “maybe useful someday” and move on. Most of what I write with AI lives on that pile.

Draft 0.6: Other people. I send it to at least three people. They comment. They highlight what reads wrong, what feels good, what confuses them. This is where the piece stops being mine alone and starts being tested against other nervous systems.

Draft 0.7: The quiet ones. I get critical again, this time thinking about the people who saw it but didn’t say much. What does their silence mean? Sometimes I follow up and ask. Sometimes their non-response tells me more than the comments did.

Draft 0.8: Levels of learning. I ask who benefits most from this piece and who benefits least. I read through it again with that lens. I double-check references. I click every link. I make sure the infrastructure works.

Draft 0.9: Expert review. Depending on where it’s being posted and how much it matters, I reach out to subject matter experts or advisors and get their input more formally.

Draft 1.0: Publish. This is a solid piece of work. I post it with a surrender to being criticized and seen for my errors. Hoping those errors are things I can correct and that I haven’t broken the trust of the people who read me.

But it’s still a draft. Even then. I see nothing I’ve written, including my dissertation, as a final piece of work.

What This Means About AI and Writing

The AI conversation, the part people picture when they hear “she uses AI to write,” is the pre-draft. It’s the 0.0 that produces material for the 0.1. By the time something I’ve written reaches you, it has passed through my own editorial judgment, a harm assessment, the weaver’s touch, at least three other humans, a round of silence-reading, a learning-level check, and possibly expert review.

The AI produced the raw material. I produced everything else. And the “everything else” is where the actual writing lives.

One More Thing. And Then One More.

In the piece before this one, I wrote about the DOT Model’s principle that you cannot skip the descent. You have to name the thing.

This piece is the descent into how. And the thing I want to name is this: using AI to write is a skill. It is not pressing a button. It requires knowing what you think before you sit down. It requires the ability to hear when the AI is wrong. It requires editorial judgment that no AI can provide for itself. It requires the willingness to push back, over and over, until the output matches the input.

The people who use AI well are not the people with the best prompts. They are the people who know the most about their own subject, who can hear the difference between what they mean and what the AI produced, and who treat the conversation as a place to think deeper rather than a place to stop thinking.

That is not cheating. That is craft.

But I don’t want to end without naming two uncomfortable things.

First: both of these pieces were written with Claude, which is made by Anthropic, the same company whose paper I’m discussing. The AI that helped me make this argument has a structural interest in the argument landing well. I believe the argument is sound. I also believe you should know that the tool and the subject are connected, and weigh that however you see fit.

Second: I know that everything I’ve written here could be used by someone to justify something I’m not doing. Someone with no expertise, no editorial judgment, and no willingness to fight with the AI could read this and say “see, AI writing is legitimate.” What I’ve described is not a permission slip. It is a specific practice that depends on specific conditions: genuine knowledge, the ability to hear when the output is wrong, a harm check, other human eyes, and the willingness to throw most of it away. Without those conditions, AI writing is not translation. It is ventriloquism. I can name the distinction. I can’t prevent it from being ignored.