How to do this yourself

The people who asked David how the novel was made were not, he thinks, asking for a tour of his repository. They wanted to know whether they could do something similar, and what they would actually have to do on a Tuesday evening to get started. This section is the attempt to answer that. It stands back from the detail in the workflow and the chronicle and tries to say what the process actually was, as steps and habits someone else could adapt. The specific incidents from this project are evidence here, not the point.

Two cautions before the advice. This is one project, ten days long, run by someone who writes software for a living and had used Claude Code heavily before, so the parts that lean on scripting came cheaply to him and may not to you. It was also expensive: it ran on the largest individual Claude subscription, the 20x Max plan, and used a good deal of it, as the setup page explains. And the site describing it was itself generated by the model, from the transcripts, so where a lesson below sounds tidier than the record it came from, the record is the more trustworthy of the two. The transcripts are linked throughout for that reason.

The shape of it in one paragraph

The process was a loop, not a pipeline. David gave the model a goal and some taste, the model produced something, David read it and said what was wrong in specific terms, and then told the model to turn that judgement into something written down, a rule, a checked example, a tool, so the same mistake would not need catching by hand again. David did almost none of the writing of those files himself. His part was to notice, say so, and insist the fix got recorded somewhere the model would read it next time. Then the loop ran again at a slightly higher level. Almost everything on this site, the rule files, the linter, the canon system, the handoff document, the overlap checker, came out of that loop as the residue of a specific failure. Very little of it was designed up front, and the parts that were designed up front (the five-book plan, the first full draft of book one) were largely thrown away.

The pages

Setting up: Claude Code, a folder, and a repository is for readers who have never used a terminal. It says what David used and offers the desktop app as the easier route.

Start with the voice, not the story covers the opening hours of the project, in which the only fiction written was throwaway samples, and argues that this was the best decision in the project.

The loop: how the back and forth actually goes describes a typical exchange between author and model, and what makes feedback usable.

Write everything down where the model will read it is about the files: which ones a project like this needs, and why a rule that is only in a conversation is not a rule.

Turn recurring feedback into checks covers when and how to build tooling, from a list of banned words to a fan-out of review agents, and the one script that found over a hundred fabricated quotations.

Expect to throw work away is about the discarded drafts, what the signs were, and how to make discarding cheap rather than devastating.

Delegate by kind of work, not by volume is about what to do in the main conversation, what to hand to subagents, and how the level of supervision changed over the ten days.

Lessons that should transfer collects the broad claims the rest of the section supports, each with its evidence and a note on how to adapt it.

If you read one page, read the loop. If you read two, add the one on throwing work away, because it is the part people are least prepared for.