Another accidental book

Last year you might remember I had NotebookLM write a 24K-word book as a test. I fed it my chat transcripts and ran various reports to shape the book and write each chapter.

One year later I gave a similar task to Claude Fable 5 on Cowork. Claude had my notes, research, and other context in hand, and it worked autonomously to wrote a whole book. It burned through a mountain of tokens and produced a 65k-word draft for Nobody Reads Your Book: How Writing Actually Works in the World. Total processing time was less than two hours. This was an autonomous eight-stage process that used multiple agents working in parallel. The plan had been its own. I just supplied the ideas and said “have at it”.

I’m still reading the draft, but dang, it’s so much better than last year’s draft. In fact, I’m having it rewrite that book next. These aren’t books I planning to publish. The books are just for me at this point. I’m reading them to see what it looks like when my ideas are expressed in a 65k-word book.

Sample (split into shorter paragraphs for the web)

Here is what actually happens between a finished manuscript and a stranger’s hands, and as you read it, watch for the moment the text itself stops being the thing that moves.

A 100,000-word manuscript reaches an agent as a 300-word query letter—not a compression of the book but a different object, engineered to resonate at the frequency agents actually process: is this sellable to an editor? The agent re-encodes it for submission; the acquiring editor, to buy it, re-encodes it again, as a profit-and-loss spreadsheet—the book as projected unit sales, the only frequency a publishing board can hear. Before publication, the sales force compresses it once more into the industry’s most revealing genre: the handle, one sentence calibrated for a sales conference, of the form “Educated, but set in rural Appalachia”—the book rendered entirely in terms of previously successful books, because comparable to what has sold is the frequency bookstore buyers process.

Those buyers, facing 40,000 to 50,000 new titles a year, spend seconds per title; the reps presenting to them have read almost none of what they present—John B. Thompson, the sociologist whose 280 interviews anatomized this industry, records a buyer waving off a season’s offerings with “Skip the next title… actually, skip the next twenty.” The chain even has its perverse elegances. A debut novelist with no sales history can be easier to sell upstream than a modestly successful veteran, because a veteran’s numbers constrain the projection while a debut’s blank record, as the industry saying goes, sets the imagination free—the system preferring a fantasy it can price to a fact it cannot inflate. And the people inside know what the compression costs: when James Daunt took over Barnes & Noble, he dismantled much of the chain’s central buying apparatus and pushed decisions down into individual stores, on the explicit theory that the relay had been filtering out the only information that mattered.

The books that survive this gauntlet arrive in stores as physical signals: co-op payments from publishers buy placement on the front tables, and 10 copies face-out where a browser will graze for three seconds carry a message two spine-out copies do not—quantity itself becomes a quality signal. The critic Gérard Genette named this whole apparatus—covers, blurbs, placement, the book’s entire clothing—the paratext: “a fringe of the printed text which in reality controls one’s whole reading.” The fringe is the system’s handwriting on your book, and every reader meets the fringe first.

Even the advance, the sum paid to the author, functions mostly as internal messaging: some 70 percent of advances never earn out, which would make them insane as investments, but they are not investments—they are the organization signaling to itself, and to the chain, how hard everyone downstream should push. Thompson’s summary deserves its italics: “Big books do not exist in and by themselves: they have to be created.”