A Letter from the Editor

I should say at the start that I did not write this world. I asked for it.

The whole thing started when I rewatched Mad Men and asked Claude why it was so great — why do I love this, why does it cut so deep, and why isn’t there more like it? I don’t remember the answer word for word, but it was good enough that it created some hope in me: that with the tools available today, it should be possible to recreate some of that specialness.

So I asked for an outline for a show. My only input was that it should be set in the 1990s. Claude suggested a print magazine as the vehicle to tell the story from, which I think was the right instinct, and that became Meridian: a monthly literary magazine in New York, founded in 1947 by Stein and Whitfield, with a European émigré spine and foreign correspondence built into its structure. The story runs from November 1989 to September 2001.

In the earliest pilot drafts I noticed a problem right away: a heavy tendency toward language. Spoken dialogue, smart witty intellectual discussion, at length, in rooms — and almost none of the embodied, physical, sensual texture that Mad Men has everywhere and almost never comments on.

There are three scenes I like to cite. The first is Peggy accusing Don of never once thanking her for her work, and Don’s heartbreaking reply that that’s what the money is for. That line cuts so deep and hurts so much that an aligned and secure AI wouldn’t write it unprompted. The second is Sally at the museum: she gets cramps, goes to the bathroom, and we see one shot from her perspective — the blood stain on her white underwear — and then her shocked face. No commentary, no interpretation, nothing follows from the scene, and yet everyone understands what it means. The third is Betty at the hospital, anesthetized against the pain, about to give birth to Gene, hallucinating in intense saturated color about a caterpillar sitting in her hand.

These three scenes seem brilliant to me and also entirely outside the standard output of an LLM in 2026, and a lot of this project was about trying to close that distance.

Research

I have never lived in the United States. I was born too late, and too far away, to have lived through the American nineties myself, and I’ve never worked at a print magazine. So frankly I had no way of judging whether anything the model produced was any good.

The best way I could think of was research. I used deep research runs, one subject at a time, and after a while I had several dozen intensely sourced dossiers on details like the smell inside a room of wax and dried cigarette smoke and old paper, the fumes of diesel air on a street in summer, the sound of car alarms at night, what a transatlantic phone call in 1989 actually sounded like on the line. From each of those I fanned out into more, and I was honestly surprised at how much depth was retrievable down there.

The dossiers were partly for me, so I could learn enough to tell good output from bad. But my hope was that they would also give Claude the context to write more grounded scenes and characters, and that’s what happened — they became the substrate everything else draws on. The set ended at 75 dossiers and is now closed.

The book

Somewhere past the second dozen I realized I was enjoying this phase a lot — more, probably, than the phase it was supposed to serve — and that there was enough material for a publication of its own. That became The Meridian Archive: Why the 1990s Were Stranger Than We Remember.

I had the dossiers organized into a spine, and then written out as full prose chapters, and from that point on the struggle was keeping the writing away from the model’s standard house style: certain words, certain phrases, certain stylistic choices it repeatedly reaches for no matter how many times you cut them. I built a harness for it — a list of tabooed words and moves, and then a rubric specifying the structure of a chapter, which could be used both as a generation spec and as a check after generation. That let a cheaper model like Sonnet draft a chapter and self-verify it, while an Opus model made the editorial calls, mainly which particular facts from the dossiers to use.

Then I set up another loop, a kind of source verification wizard. Any claim that sounded implausible, or that had a precision that seemed questionable, went back out to web search to be validated against real sources and citations. So the claims in this book have been checked once for plausibility by Opus and then once more against the web in specific cases. That doesn’t mean all the facts here are correct. But I think it sets a standard that a lot of traditional, human-only journalistic writing doesn’t match either. Since this is a hobby project, I eventually had to draw a line of where I would stop checking.

I also discovered that I preferred listening to these long chapters rather than reading them, and that NotebookLM was incredibly good at making them entertaining while staying grounded in the actual research. So every chapter got an audio podcast, and when NotebookLM added video shorts — which were really impressive — every chapter got at least one of those too. You can watch a sixty-second short, listen to a twenty-minute podcast, or read a three-to-five-thousand-word chapter on the same subject, depending on how much you want.

The novel

The show is still the goal, but the AI video tools aren’t usable yet for a full-length season — the costs alone are implausible for a hobby project, and the content filters make even single scenes difficult. One example I keep telling people: Seedance 2.0 refused a prompt that I had, over many iterations, sanded down until I was sure it contained nothing of any possible offense. It kept refusing. The problem turned out to be the number 1989. The model was Chinese.

So the novel became the intermediate step between the research and the show, to be produced once generation costs come down enough.

The production was again multiple layers of loops. I started with characters — prompting for people with interesting internal and external wants, often in conflict, some known to them and some not, and in conflict with other characters’ wants too, each with an arc and a background. Nora Kessler came out of that process: editor in chief at thirty-eight, half German, half American, running the magazine after her predecessor’s death. A lot of the names and basic premises came directly from Claude, and I only corrected them when I didn’t like something or felt I had a better idea.

Then an outline of the entire decade, then a per-year timeline, then per-chapter outlines. I iterated a lot on the first chapters, and the same problems kept coming back: house style, repeating phrases, explaining too much of the basic premise, and avoiding uncomfortable subjects — interpersonal conflict, cruelty, sex, violence.

What finally worked was noise. I had it extract real events from the research that function as ruptures, and those get randomly inserted into the prompt for a chapter, so the writer has to somehow integrate them. Another pool held period practices — cocaine use, the widespread sexism of the workplace — inserted the same way. And because I noticed the model tends to write scenes of single characters, mostly interior, with no consequences, I created a matrix of every principal character paired with every other, so each pairing has to have at least one interaction somewhere in the book, and that became a third pool. Forced entropy, basically, to make the model work around its own bias toward safe and predictable storytelling.

Once I felt I wasn’t getting better results on the first chapter, I asked Claude to orchestrate the writing of the entire rest of the book, feeding each finished chapter into the prompt for the next one, so the more heavily curated early chapters would exert a founder effect on everything after them.

What I am

I don’t think of myself as the writer of this universe, or its creator. Consumer zero is closer to it. I wanted this world to exist, and I gave Claude preferences.

What I learned is that even giving preferences is a lot of work, and you have to decide how much of it you’re willing to do. I assume future models will do better with less of it.

Here is the part I should probably admit: I haven’t read the novel yet. I’m a bit afraid to. This letter is me writing the facts down now, while I still remember how it was done, before I move on to some other project — the verdict can come later. But it still seems worth saying that a person can now build an entire fictional universe without ever typing a single line of it.

The final issue of Meridian ships in September 2001.

July 2026