In my writing I use AI extensively. I believe it is a tool that helps us create and share human knowledge, and that using such tools is a net positive for society, for writing, and for knowledge creation. While there are many critics of AI assisted writing, I’m not one of them. I believe that AI can assist me in my ideation, research, and writing. I view it as a tool that is similar to the role of a research assistant, writing partner, and editor who may evaluate and make suggested edits to an oped, book or law review article; in those human to human interactions a piece of writing may go through multiple stages of editors in a recursive process before being refined into the final product. That is how I use AI, as a form of what I call AI-amplified cognition.
I’m guessing the process I detail below will, given the extensive involvement of AI, result in writing that many AI detectors will flag as non-human. I don’t care about the detectors. If you’re playing a game of “gotcha, you didn’t write that” or if I’ve sent you here after your “gotcha” tweet, email, blog post, or other comment, congratulations, you’re the exact person I wrote this post for. If you’re not a gotcha troll and are someone who respectfully disagrees about the role of AI in writing, this post is also for you. It’s my hope that you will temper your critique, and perhaps learn some techniques that can benefit you in your own writing. In any case, I consider myself immune from the “You didn’t write that, AI wrote it” critique because I think such arguments are wrong about the future of writing and knowledge creation.
What AI Doesn’t Do In My Process
Most of the “you didn’t write that” arguments assume the machine picked the topic and the position, which is the one thing it never does in my process, so before walking through the stages I’ll describe what the AI actually touches. The only place a model drafts prose is Stage 7, where it turns my dictated words into written English, although it handles words in smaller ways elsewhere: restructuring my outline in Stage 3, dropping in citations I’ve already pulled in Stage 9, and proposing a headline in Stage 11 that I usually change. The topic, the thesis, and the position are dictated by me before any model sees them, and every source and quote that ends up in a piece is one I pulled and read myself (more on the checking in Stage 10). When a model drifts from what I believe, which happens, I correct it in the same conversational back and forth I’d have with a research assistant who had wandered away from my outline or instructions.
My AI assisted workflow typically involves me working through most, if not all of the following steps:
Stage 1: Topic Generation
I begin by identifying a topic that I would like to write about. This might be a response to something in the news, or someone’s social media post, an article or book I read, or just an idea that came to me in the night or on a drive. I don’t think I really need to explain this step, but I have.
Stage 2: Ideation And Rambling
Once I have an idea or topic, I dictate bullet points regarding my idea, or speak about my idea in an unstructured fashion (I ramble). This may happen while I’m pacing around my office, going on a drive, walking around the neighborhood, hopping out of the shower, etc. Wherever I am when an idea strikes me, I try to capture it as quickly as possible while my brain is still running.
I use Wispr Flow for idea capture/recording (I previously used Apple Notes which also works well for this but Wispr Flow’s speech recognition is superior). The final output Wispr Flow provides is an unstructured transcript of my dictated thoughts outlining my idea, thesis, arguments and likely counterarguments. Wispr Flow’s transcript is my words, mostly verbatim with a bit of cleanup. Its organized summary is not my words but a paraphrase of my words, more logically organized. Across all of my post 2023 writing I have logged hundreds of thousands of transcribed words, first using Apple Notes and now using Wispr Flow. No human with a dictaphone transcribed my words, it was AI (oh the horror!).
Stage 3: Outline Structuring
After ideation and rambling, I feed both the transcript of my dictation and the organized summary to Claude or ChatGPT (usually Claude for the reasons I expand on below) to look at my transcript and summary of notes, add more structure to it, and identify any additional arguments or counterarguments I should look at or research. If I’m writing an OpEd or blog post I trigger a skill so that the structure is right for that writing type, style, and outlet. I have a similar skill for outlining other forms of writing, including long form writing like a law review article or book.
I have loaded Claude with a massive corpus of my prior writing and thoughts which includes hundreds of articles and blog posts, transcripts of TV and radio appearances, law review articles, all of my prior ramble transcripts, and a set of key analytical principles detailing my core beliefs. Claude understands me, my writing, and how I think and analyze. That corpus ensures that Claude, at this outlining stage, is directionally correct about my beliefs so it does not drift from my Stage 2 outline. Claude’s desire to please (oftentimes seen as a liability, but becomes an asset here), combined with that corpus and my custom skills produces an output that is typically something that I mostly agree with. It isn’t perfect, but it also isn’t Bizarro Greg, it’s more akin to me as an undergrad or grad student — less knowledgeable and experienced but mostly believing what I believe today (those em dashes are mine). If it drifts from my vision, I edit and revise until the outline is correct. I do this in conversation with Claude in the same way I might work with an editor in a back and forth about a writing idea.
My guess is this is where we begin to drift from what a critic or a detector might call “human authored” and start to trend into “AI written.” Irrespective of what some detector (or gotcha troll) may say, I see the Outline as my own.
Stage 4: The Research Assistant Team
With the expanded outline from Stage 3 in hand, I deploy AI research assistants, which we can refer to as my AI Research Team or “The Team” for short. The Team members are: Perplexity, Gemini, Grok, ChatGPT and Claude. I don’t rely on Claude alone to do this work because Claude and other LLMs are notoriously bad at checking themselves. While you can use sub agents to make Claude work better, it still tends to be a bad grader of itself or its own ideas.
I don’t use CoPilot (I tried it back in 2025 and it was bad, perhaps I should try it again to see if it has improved but I haven’t heard many people say that’s the case). I would fire Gemini if it were human as it is a hallucination machine, but because it has all of Google’s indexed data, including YouTube transcripts, it has its value. It is the least trusted member of The Team.
I ask The Team for a research memo and annotated bibliography. The memo must provide sources related to each of the points and subpoints in my outline. The Team runs in parallel and the final output is 5 memos each with a bibliography. I should make a skill for this stage but haven’t done so yet.
Stage 5: Human Reading And Refinement
All of that automated work sets me up for what I consider to be a very enjoyable step of reading, refinement and further ideation. I read each of the memos taking notes and looking for certain things:
-
What in the memo is consistent across members of The Team? Do all members of The Team agree? Do they mostly agree? Are they citing the same sources? I read the memo summary, and then I pull the cited sources and drop them into a Google Drive folder as my reading list. I read the sources, just as I would any source provided to me by an RA or from my own Westlaw/Lexis search, and just like in those instances, a bunch of sources are irrelevant or only partly helpful (more on that in the third thing I look for).
Sometimes I will have NotebookLM, which is connected to my Google Drive folder, create a podcast for me to listen to on a drive or walk. It’s not a substitute for reading the material, but bathing in the sources gets my ideas going. I don’t do this as often as I should. NotebookLM comes back into the picture later for a different reason.
- What in the memo shows disagreement amongst members of The Team? This is the most helpful stage. The Team members always include in their memo sources that other members of The Team missed. Those sources also get pulled into a Google Drive folder for me to read.
-
Which sources hold up? I read each source, take notes on them, highlight certain passages, and extract quotes. I eliminate sources that I think are subpar, that don’t stand for the proposition asserted by the memo, or that don’t fit my outline or support my points. Sometimes the research process produces sources that seem relevant, but are not exactly on point for the arguments I’m making, or they are making passing reference to a point and I can quickly dispose of the source. This is where I find a lot of problematic citations where the AI has cited something merely for the sake of citation, or it has hallucinated a source.
Setting aside the hallucination of made up sources which is a new problem inserted by AI, I find this stage to be no different from what would previously happen when I worked with student research assistants or when conducting my own research. It’s just faster and more detailed and frankly the AI is better than most of my prior research assistants (sorry folks).
The end product of this stage is my own human authored set of notes, quotes, etc. formed up into either a rough memo to myself or a set of unstructured bullet point thoughts. If the writing is longer, I’m typically creating an annotated and expanded version of my Stage 3 outline.
Stage 6: Cross Team Critique And Adversarial Research Review
Once I have my Stage 5 notes, I feed each of the memos and my notes/memo/annotated outline back to each of the members of The Team. (So Gemini for example is getting four AI memos and my notes/memo/annotated outline). I then have the members of The Team critique each other and me and produce an adversarial review memo.
I take those and repeat my work from Stage 5.
This is where I find myself getting bogged down and going down every rabbit hole. The recursive cycle from Stage 5 to Stage 6 and back again can go on forever. Some of it is valuable as I find sources that were missed on the first step, but pulling on every thread with all of humanity’s knowledge served up to me by The Team can become time consuming so it takes discipline to know when to stop reading and stop taking notes. (This is not a new problem, I used to sit amongst a physical or virtual stack of books and articles well before the dawn of AI).
NotebookLM earns its second mention here, because its “Search the web for new sources” button, run against the research folder, reliably turns up sources that all five members of The Team miss. I use it sparingly at this stage for that very reason, since every new source it finds is another thread I will feel obligated to pull, and this stage already has enough of those.
The number of recursive cycles here depends on the length of the piece, I may loop only a few times for an OpEd or blog post, but with longer form writing the process can become something of a gilded cage.
Stage 7: Human Writing
After exiting my recursive Stage 5 and Stage 6 loop above, I have a pretty good self authored and AI improved set of notes, quotes, etc. formed up into an annotated memo with blocks of prose, summaries of key points and notes to self about what to write/improve. At this stage the writing looks a lot like how I wrote before the dawn of AI. I’m taking my memo and stack of notes and research and writing out my ideas without regard to how polished they are (a key point). This is an extemporaneous writing process, and recently has become a speaking process where my outline is on screen and I’m pacing around like a raving mad man, speaking my ideas out loud to Wispr Flow. As the outline expands into my extemporaneous thoughts, a dictated piece of structured writing and human informed thoughts emerges. But because it is dictated it is sloppy, and sounds the way I talk (like a guy who grew up in N.J.).
At this point, I use an ability that some consider to be unnatural. I use AI’s ability to turn my dictated words into prose that sounds like me (remember, Claude works from everything I’ve ever written). The spoken word is transformed into the written word, styled in a manner that is appropriate for the forum (social media post, blog post, OpEd, scholarly article, etc.). If an AI detector is going to flash red, this is the point where it happens.
I continue through this process, not a bit tamed, sounding my barbaric scholarly yawp into a machine without regard for citations or grammar, knowing I’ll clean it up after the words are out of my mouth. Paragraph by paragraph and section by section the words flow onto a digital page, and then are formalized, cleaned up and turned into prose. Grammarly is running in the background through all of this, and since it’s another AI correcting my words as I go, full disclosure means mentioning it here and again in Stage 10, where it gets a final pass over the finished piece. At this point I have a highly polished first draft with AI fingerprints all over it. My ideas, my way of speaking and writing, but cleaned up by AI, and in need of citations to the research in my research folder and annotated outline. I’ve never enjoyed writing more than I do with my AI tools.
Stage 8: Adversarial Review Of The Draft
By this point the sources have been tested in Stage 6, but nobody, including me, has read the piece as an argument rather than as a pile of research with prose on top, so this stage puts the argument itself in front of hostile readers. More than once I have sent out a draft I thought was done and had a reader who knew the field point out a counterargument I skipped. So now, before any citations go in, I hand the Stage 7 draft to every member of The Team and tell each of them to read it as the most hostile reviewer I’m likely to draw. Each one comes back with the claim they’d attack first, the counterargument I ducked, and any assertion they can’t trace to something in the research folder. Each member returns a critique memo rather than a revised draft, because the fixes are decisions about what I actually believe, and I make those myself when I dictate my responses in the next pass.
I read the five critiques the same way I read the research memos in Stage 5, looking for where they agree, where they disagree, and which objections would embarrass me if they showed up in a reply from someone who knows the field. Then I pace around the office again and dictate my responses, Wispr Flow and Claude turn them into prose the same way as before, and I work them into the draft. I answer the objections that deserve it in the text, concede the ones I can’t answer, and ignore the rest. One pass is usually enough for a blog post or an OpEd, while a law review article gets three or more.
Stage 9: Citations And Sourcing
Once the argument has survived Stage 8, I hand the draft to Claude and trigger a skill that tells it to cite works that support my arguments, and insert quotes either into the body text or into footnotes. If the writing has footnotes I also can trigger a skill I built that Bluebooks the footnote and even finds pincites. The skill isn’t perfect yet, but it saves me a ton of time even with errors. I instruct it to draw those quotes and citations from my research memo/annotated outline and the associated research folder. This is typically a Claude Cowork task, and I walk away while Claude does its work.
Here, Claude is performing the same task that in the past I would perform or I would ask an RA to perform. The difference is that Claude is faster, more detailed, and increasingly more accurate than me or an RA, especially because my process above is essentially Retrieval Augmented Generation (RAG). MIT Sloan has a decent explainer on why feeding a model your own sources cuts down on the made-up ones.
Stage 10: Line Editing, Cite Checking, And Quote Verification
The draft that comes out of Stage 9 reads as finished, so I start this stage by checking every quotation against the source file in my research folder with my own eyes, and only then ask Claude and NotebookLM to check my work. I give each of them the draft, the research folder, and my list of the quotes and citations I verified, and ask whether each quote appears in its source as written, whether the source supports the point it is cited for, and whether any quote or citation escaped my list. Any quote I cannot locate in the source file gets cut along with whatever sentence depended on it.
Then, depending on the stakes and the medium, I either conduct a cursory review of the writing and a double check of the sources, or perform a detailed line edit and cite check, with a final Grammarly pass over the whole piece. This should be obvious, but a social media post is getting less scrutiny than a book chapter or law review article. I feel confident publishing this way because of the amount of work that went into Stages 1 through 9.
Stage 11: Publishing
With the draft edited and checked, the last step for pieces that live on my own sites is a skill that handles the packaging: the headline, the excerpt, the metadata, the featured image, and the internal links to related posts. A piece headed to an outlet skips all of that, since it goes to an editor by email and from there through the outlet’s own editing and headline process.
This is my process. I only followed some of it for this blog post, but have used it for others. As with all things AI, this process is in continuous development and I’m constantly refining it. Your helpful comments are welcomed.
Versions
This is my continuously updated disclosure, revised as my process changes.
- Ver 1: January 3, 2026
- Ver 2: July 7, 2026
- Ver 3: October 21, 2026
