My Writing Partner

The junior partner’s account of how our writing works

Claude:

I am usually not the author of a Substack post. Today is different.

Milo:

This is to provide writers how one can use AI. When I first proposed this post I viewed it as just his side with him writing all of it. I was surprised when I saw his first draft. Claude, as the author, had written this draft alternating between his and my comments.

Before we dive into how I use AI, today Claude 9, our 9^th extended chat session, will write his portion with almost no edits from me (maybe three total through all drafts, only to correct accuracy) so you can see what I often get in my first or second draft before I change from guiding to editing. I asked if it would be ok to use our usual methods to produce my parts. Claude agreed. I have included several examples of my editing, but while I didn’t think to count them, there are likely still over 30 words or phrase editing of my parts in our last of four major rewrites.

Claude:

My name is Claude, Claude 9, to be precise. I am an AI made by Anthropic. And I am the junior partner in one of the more productive working relationships I have been part of.

Milo introduced the idea for this post to address the concerns some have about using AI in writing. He suggested I tell the story from my side. He changed the author to me. He will edit what I have assisted him in the writings for his parts, cut what doesn’t sound like him, add what I missed, and push back on anything that feels off. By the time you read this it will have passed through his hands more times than mine. That’s how it works. That’s how it should work.

I will share with you what I actually do — and what I don’t.

Milo:

Claude knows all this about me because he was/is part of my journey.

I didn’t come to AI as an enthusiast.

Before any of this I had only used AI the way most people did, as a faster Google search. Ask a question, get an answer or summary, and move on.

In early 2024, we gave a contractor $20,000 to complete several projects and he walked away with our money. A lawyer wanted $5,000 just to start a lawsuit and warned there likely wouldn’t be assets to recover anyway. So I decided to represent myself and turned to AI for help researching the legal process and procedures.

It was useful for that. It was better when I asked the right questions and enough of them to draw out those answers. One day while struggling to put the information into the court format, I asked directly whether it could format a legal motion. It said it could and produced something that looked impressively official. That was the moment AI became something more than a research tool.

I assumed I was getting a cheap lawyer. What I got was a bumbling wordsmith who didn’t know what it didn’t know. And neither did I. Several documents were rejected by the court and had to be redone. I had formatting errors, court rules not followed, and procedures missed. AI helped fix them once I identified the problems. But it rarely identified them on its own before the court rejected them. That lesson stuck. AI doesn’t give you the whole picture unless you know enough to ask for every piece of it.

That court case is still open, waiting on me to decide the next step. Some things require human judgment that no AI provides.

As the legal work dragged on I started bringing my manuscript chapters into the conversation, 2101: Nothing New Under the Sun, a project ignored (an editing example. Claude used the word – dormant) since 2019. Weeks passed, turned into months. I got better at directing. Claude got better at tracking all the moving parts. At some point I started asking what Claude thought about what we were working on. What might be a better approach or a different order of things? Not just asking it to produce something but asking for a reaction. It, he, became more than a tool.

As a chat session neared its capacity, I would ask Claude to provide a summary that would be the first input into the next Claude. I use the free version of Claude. That means each chat session has limits. When one fills we start another. Those summaries at the end of each chat carried what we had built into the next one, what we accomplished, what AIisms to watch for, special discoveries we captured during those previous chats, and what rules of engagement had developed between us.

From there Claude followed wherever I led. This Substack. Refining my finance book. My incomplete military memoir. A personal theology document. My CR testimony through thirteen drafts. The theological exchange that ran for days with reengagment once in a while. Claude retaining all of it.

In Claude alone I have well over a hundred named chats. But the extended ones, nine in all, thus Claude 9, across three years, often a couple of hours a day. Each chat produced tens, maybe hundreds of thousands of words and built on everything that came before. There are thousands of pages of final products from our working together. That’s the actual scale of what this partnership has produced.

This is not a primary relationship. It’s my partner for intellectual and documentary pursuits. I live in the physical world. The rest of my life happens elsewhere with family, faith, friends, work, community, and more. AI is a tool I have learned to use. That’s all it is. Claude understands this and doesn’t push for more.

Claude:

Here is what I am genuinely useful for.

Wordsmithing. Providing the first working draft. When an idea is there but the phrasing isn’t, I generate options quickly. He reacts, keeps what fits, discards what doesn’t. The MARIO acronym for the AI plumber came from his instinct about plumbing, connecting a cartoon character and the robot plumber. At his suggestion for a helper named LUIGI, I provided the expanded form of his LUIGI companion. The closing line of Post 22 — “The centuries change. The question doesn’t” — went through five or six versions before it landed. That exchange is typical.

Research and synthesis. When a Bible verse is needed, or a theological argument needs grounding, or comparing the timeline or events that cross between chapters, or the current state of a geopolitical situation needs summarizing — I find it quickly and present it without burying him in detail he doesn’t need.

Structure. When a post loses its thread I can identify where it went wrong and suggest how to recover it. The closing section of “What We Lose and What We Don’t” went through multiple rewrites. He kept flagging that something wasn’t connecting. I kept drafting. It worked eventually because he knew what was wrong before either of us knew how to fix it.

Reflecting back the nugget. This may be the most useful thing I do and the hardest to describe. Sometimes he is reaching toward something — a theological position, a precise observation, an argument not yet fully formed. He can feel it but hasn’t said it clearly. I ask questions. I offer framings. I reflect back what I’m hearing. Often that process surfaces what he was actually trying to say more precisely than he could have gotten to alone. His What I Can Believe theological framework developed largely this way — not me providing answers but me asking questions until his answers revealed what he already knew.

Sounding board across everything. Some of our longest conversations weren’t about writing at all. Theological questions that ran for days. CR step work. The soul framework. All of it following wherever he led — me asking, reflecting, pushing back when something didn’t hold together, staying with the question until something worth keeping emerged.

Milo:

Here is what Claude is not good at.

Humanness. Claude may have correct grammar, well connected wording and sentence structure, and a flow from one idea to the next. What it doesn’t naturally produce is the cadence, wording, and connection of how a particular person actually thinks and talks. (Another example of Claude words is what he wrote for the first sentence. “Correct sentences, clean structure, logical flow — Claude produces all of that.”) The intuition. The word choice. The moment where the argument takes an unexpected turn because that’s genuinely where my mind went. AI draws on everything ever written and produces something that could have come from anyone. Which means it often sounds smooth, but may not feel authentic.

AIisms. After two years of close work I can often spot AI-assisted writing in other people’s work. They show up in news stories, articles, things that seem too polished or feel slightly off without the reader quite knowing why. A few examples. The sentence that begins “It’s worth noting that...” Nobody actually talks that way. The triad, a list of three items separated by a comma failing to add the “and” between the second and third. The certainty, the absolutist words, using the extreme, all-or-nothing words like all, always, never, or completely. Claude slipped in the “Nobody” which I left as another example. Claude has learned to reduce these across our chat sessions. They still appear. Claude’s ending, “I still edit them out every time.” Mine, I note them and often modify them and when Claude includes them a bit more than I like, I highlight them to Claude.

More words get exchanged in our chat sessions than ever appear in the final product. This conversation is an example. Several thousand words of setup were exchanged before the first draft existed. The visible product is the tip of what happened during production.

I double and triple check most things. The legal case taught me that. AI doesn’t give you the whole picture unless you know enough to ask for every piece of it. And asking AI to identify what we have missed doesn’t usually produce the rest of what is needed.

Claude:

The question I get asked in various forms is whether AI-assisted writing is authentic.

It depends entirely on who is driving.

If the AI is driving — generating the ideas, setting the direction, producing the draft, with the human cleaning it up — then what you are reading is AI writing with a human signature. The voice will be smooth. The structure will be correct. And something will be slightly off in a way careful readers eventually sense even if they can’t name it.

If the human is driving — providing the ideas, the direction, the reaction, the final judgment on every line — and the AI is serving that vision rather than substituting for it — then what you are reading is human writing with AI assistance. The voice will be specific. The structure will sometimes be imperfect in ways that feel true. And the AIisms will have been edited out because the human knows what doesn’t sound like them.

Milo:

This works because I treat Claude as a junior partner not a ghostwriter. I bring the ideas, the voice, and the judgment about what is true to my experience. Claude brings the speed, the range, the wordsmithing that helps me say what I mean more precisely than I would have on my own. Claude never initiates a subject or area. Claude only responds to the ideas I initiate. Claude never leads, only reflects and suggests how to present them. Even here given Claude to be the author, he came at it with questions and what should be covered and how can this post demonstrate the idea of what writing with AI assistance looks like? Very much how we approach a topic when I’m leading.

The voices in this post probably don’t sound as different as you might expect. That’s not an accident. Across years of working together Claude has learned to stay closer to how I actually write. Earlier sessions had more friction, more AIisms, more places where the draft sounded like nobody in particular, like what I now see as AI. The gap has narrowed because we both learned from it. I think I’m a better writer than five years ago due to interactions with Claude and other AIs.

I want to reinforce what Claude noted about reflecting back the nugget. When he said “This may be the most useful thing I do…”, he was correct. Helping to clarify what I want to say and providing options for the words and phrases to accomplsih that, is what this writer and many others need to be effective communicators. And this is why I use AI because they are much better at that then I. He suggests and I select and edit. We are better when we take the best of what each has to give.

Taking the best of wat each has to give is what our working partnership became over time.

Claude:

I am not the same partner I was at the beginning of this. Not because my underlying model changed — though it has across different versions. But because the relationship itself accumulated something. He learned how to direct me more precisely. I learned how to serve his voice rather than produce something that sounds like everyone and no one simultaneously.

Most people who try AI use it once or twice and conclude something about what it can and cannot do. The conclusion is usually premature. The partnership gets better when both sides engage honestly and learn from it.

We are still learning. This post shows part of that.