How I Use AI to Write This Blog
The real unit of AI assisted writing is not the prompt. It is the editorial system around the prompts, and that system starts before any machine runs.
The editorial system, in one picture. It begins before generation: I supply the observation, the provisional thesis, the key passages and the research questions. The assistants investigate, draft, criticize and help produce. Research can overturn the thesis it started from, and every stage can send a piece backward before it is published. The archive then feeds the next article. Diagram generated with AI assistance from the author's workflow and revised by the author.
Posts appear here at close to a daily rate, and a reader may reasonably wonder how one person sustains that pace alongside a full professional life and everything else competing for attention. The short answer is AI. The short answer also conceals almost everything interesting, because a post here does not begin with a request to write an article about something. It begins with an observation: a question raised by a book, a problem at work, a contradiction in a news story, a conversation, or, in one case in September 2026, stepping twice onto a bathroom scale and getting two different numbers.
What happens between the observation and the first draft is the part the short answer hides. I usually enter the process with a provisional thesis, several sentences or examples I already know belong in the piece, and a list of questions that have to be answered before any of it can be trusted. Sometimes the thesis survives the research. Sometimes the research destroys it, which is the more useful outcome. The assistants investigate, challenge, organize and eventually draft around that material, but they start from an intellectual structure I have already seeded.
The popular picture of AI assisted writing has two boxes and an arrow: prompt, then article. I have come to think the picture is wrong in a way that matters. The real unit of AI assisted writing is not the prompt. It is the editorial system surrounding the prompts.
The editorial system
The diagram above is the whole machine, and it runs in three phases: seeding and investigation, an editorial loop of draft and critique and revision, then publication and memory. Drawn as a line it looks orderly, and the line is misleading, because nothing about the process is linear. Research changes the argument and occasionally replaces it, critique sends a draft back to research, scoring sends it back to revision, and a finished post reaches backward into the archive and forward into the next idea. It resembles an editorial room more than it resembles text completion, and the diagram shows who does what in that room.
The machines participate at nearly every stage, and they do not carry equal responsibility at every stage. AI supplies search, comparison, organization, drafting, criticism, continuity and production capacity at a speed and marginal cost that change what one person can reasonably attempt. I decide what deserves attention, which evidence deserves trust, what argument is fair, which criticism is right, and whether the result deserves my name. The distinction is the one the rest of this piece keeps coming back to.
What I supply before generation begins
The idea is still the scarce resource. In August 2026 I stood at the Library of Congress table at the National Book Festival and watched a volunteer work on an unfinished 1917 telegram from a Texas judge to his congressman, part of a crowdsourced transcription program that, by the Library's own count, passed a million completed transcriptions in 2025. That became a post about the limits machine transcription encountered in an archive of difficult handwriting. No prompt produced the idea. I was there, and I recognized something, and the archive gets its continuity from what I notice and choose to pursue.
Recognition is only the first thing I bring. Before the assistants do anything substantial I have usually committed to a provisional explanation of what is going on, and I have written down the formulations I want to survive into the finished article: an anecdote, a comparison, a sentence I have been turning over for a week, the example that makes the abstraction concrete. Some of the language that matters most in a post therefore exists before any generation happens. Alongside it goes the research agenda, which is simply the list of questions that have to be answered for the provisional explanation to hold. The assistants are excellent at investigating a question. They are not the ones deciding which question is worth an afternoon.
What AI changed is the economics of everything after that. Testing an idea costs minutes rather than an afternoon, so many more ideas get tested and most die quickly, which is the correct fate for most ideas. The survivors go to research, where the assistants hunt for sources in a fixed order of preference: the original document, then the institutional source, then the peer reviewed literature, then good reporting, with secondary summaries last.
The bathroom scale shows what research is actually for. My provisional thesis was ordinary and probably what most people would assume, that the scale was simply imprecise and had rounded two readings differently. Research replaced it with a manufacturer's patent on stability detection, a semiconductor company's reference design, and the difference between the precision a display shows and the accuracy a sensor has. The article I published was not the article I set out to write, and a process that cannot reach that outcome is not doing research. It is collecting citations for a conclusion already reached.
The rules and the memory
Two less visible parts of the system do more work than any prompt. The first is a standing set of instructions, accumulated since the blog began, that I carry into every piece of work: a newspaper column rather than a seminar paper, visible sourcing, licensed images, technical claims checked against documents, and a rule that polished language is not evidence of quality. A publishing house keeps that kind of culture in stylebooks, editors and tacit knowledge, and working this way forced me to write some of mine down. A prompt begins one conversation. Standing instructions carry accumulated judgment across all of them, which readers of my Algorithms series will recognize as an algorithmic idea in its own right.
The second is the handover. Longer pieces span several sessions, several days and often more than one AI system, and I do not rely on a later session to reconstruct the reasoning that came before it. At the end of a working session I have the assistant write a document for the next one: what has been established, what remains uncertain, which sources matter, what has already been rejected and why, and what question should come next. When I finished 9 Judges, 2 Votes in September 2026, the handovers from the two posts before it told the drafting session which findings were settled and could be cited in a sentence, which analogy had been tried and abandoned, and where the argument was supposed to land. A different system could walk into the project and pick it up. The handover is editorial memory, and it does something a prompt cannot, because it prevents a rejected argument from quietly returning three sessions later. Continuity includes remembering what not to do.
Draft, attack, revise
A draft is where the popular picture of AI writing begins and ends. In my process it is where the editorial work starts. One system drafts under the standing rules. A different session, sometimes a different system, is then asked to attack it: find the unsupported assumption, the technical claim that outgrew its source, the missing counterargument, the paragraph that repeats an earlier one, the sentence that sounds like a machine, the title that explains too much. Fluent prose disguises weak reasoning, and much of this system exists to stop fluency from becoming the standard. I am the referee, which means I decide which objections are valid and which are a critic being clever.
Scoring is the part of the process that looks strangest from outside. I have every draft rated from 0 to 100% on a fixed list of dimensions: opening, originality, technical credibility, accessibility, momentum, voice, sourcing, title, ending. A draft may score 92% overall and 75% for technical credibility, and the useful number is 75%. The percentage does not measure literary quality, and I would not defend it as a measurement of anything, because its purpose is to force a criticism into explicit form. A low score demands an answer to one question, why is this not better, and the answer is always specific: a source that does not say what the draft says, an overstated cause, two essays competing in the middle, an ending that summarizes. It's Mine & It's Wrong went through that loop until the loop stopped finding anything, and the thesis that survived was not the thesis the first draft had.
Verification comes last and is not delegated. Material factual and technical claims go back to primary or authoritative sources before publication, and the assistants make those documents remarkably fast to find. I still read the claim with the source open rather than accept a summary of it, because deciding whether a document supports a sentence is a judgment, and the responsibility for what the post asserts is mine.
Publication and the archive
An image is chosen the way a paragraph is, to clarify or argue rather than decorate, and I confirm licensing and credit the artist or institution before it goes up. The photograph on the bathroom scale post shows a display reporting a decimal place the sensor cannot resolve, which is that article's argument in a single frame. Publishing is then a stage, not an ending. After a post goes live I open it and read it as a reader rather than an editor, and something is nearly always different on the page: a link, a caption, a line break, a sentence that read fine in a draft and does not in public. Then comes the question that has changed my sense of what this blog is. Does the new post change the archive? An older entry may need a link to it, or may belong in an older post's further reading, and sometimes a category appears that was not there before. When the bathroom scale post went up, the research for it sent me back through the archive, and I found that posts written months apart about aggregated ratings, personalized pricing, prediction markets, AI measurement and quantum verification were about the same thing: the encoded decision underneath an ordinary result. The Algorithms series existed before it had a name.
The archive is therefore both output and input. Earlier posts are searched before a new one is drafted, and a series can emerge retrospectively from entries never planned as one, because patterns invisible when individual pieces were written begin to appear across the whole. AI makes that continuity easier to recover, and deciding what belongs together remains an editorial judgment. The machinery produces posts. The archive reveals the larger project.
All of which leaves one question the machinery cannot answer about itself. If AI participates in research, drafting, criticism, revision and production, who is the author? That deserves its own article, and it gets one in Who Wrote This?
First published September 2026 and revised the same month, when this account was split from the authorship essay that follows it. The account is revised when the method changes.
Further Reading
Posts mentioned in this piece
- AI Can't Read Teddy Roosevelt's Letters - the idea that came from standing at a table.
- I Weigh How Much? - a trivial observation, a patent, and the first entry in the Algorithms series.
- It's Mine & It's Wrong - the post whose thesis changed under critique.
- 9 Judges, 2 Votes - the post the handovers made possible.
- Where Does Supergirl Rank? - how indices turn opinions into a trusted measure; an Algorithms post from before the series had a name.
Other posts the archive search recognized as Algorithms entries
- The Price You See May Not Be the Price I See (April 2026) and You Seem Like Someone Who Should Pay More (April 2026) - personalized pricing.
- A Misplaced Bet: Prediction Markets at the Edge of Governance (March 2026) and What Would Cassandra Think of Kalshi? (February 2026) - prediction markets.
- How Should We Measure AI? (August 2026) - measurement.
- IBM Confronts Quantum Computing's Verification Problem (August 2026) - verification.
Sources cited in this piece
- About By the People - Library of Congress crowdsourced transcription program, launched 2018, the table I stood at in August 2026.
- Method and system to quickly determine a weight, US 10,024,709 - Withings, granted 2018; the stability test behind the bathroom scale post.
AI Assistance Statement ▾
This blog publishes at a near daily rate, and that pace is possible because AI tools do a substantial share of the work between the idea and the published text. Preparation of this entry included assistance from Anthropic's Claude and from OpenAI's ChatGPT (GPT-5 series reasoning models). I use them to research a topic and gather primary sources, to organize ideas and propose structure, to draft and revise prose, to check factual claims against the cited sources before publication, and to score drafts against a set of house style rules. Longer pieces are often developed across several sessions. A written handover carries the argument, sources, and open questions from one session to the next, and the same tools help prepare those handovers. The tools also help identify candidate images and confirm that selected images appear to be released for reuse, for example through public domain or Creative Commons licensing.
The process is also an AI experiment in its own right. There is a live argument about what AI-assisted writing does to originality, and whether the result is thought or slop; the August 2026 dispute over a Wall Street Journal op-ed that its author acknowledged drafting with AI, and the Journal's subsequent defense of the practice (WSJ is behind a paywall, but for a public summary: click here, and here), is one newsworthy example. I would rather run the experiment openly than pretend it is not happening. This blog is one sustained attempt to find out whether a person with an argument, working with these tools every day, produces writing that is still recognizably that person's, and I disclose the method so readers can judge the result.
The judgment is mine. I choose the topic, decide the argument, supply the personal and professional experience the pieces draw on, read and edit every draft, verify the sources and image licensing, and take full responsibility for the final published content. Where a post contains my own recollections, the AI did not invent them.
Statement revised September 2026.