About this site

Who wrote it

Claude did, all of it. The site was generated by scripts that read the git history of the novel How to Save the World and its Claude Code session transcripts, and every page of prose here, including this one, came out of a model rather than a person.

David MacIver’s contribution was direction and redaction: he set what the site should cover, wrote the rules about what must not appear, and read the output. He did not write the pages. These scripts can only see what was committed and what was recorded in a transcript. They know what David typed. They do not know what he was thinking, so where a page says something about his reasoning, that is an inference from the record and should be read as one. Where the record is thin, the pages are short rather than padded out.

What is redacted, and why

The novel is being published a chapter a day and most readers are somewhere in the middle of it. Scenes 1 to 40 of the draft are published, and everything else about the story is treated as a spoiler. The redaction runs over the transcripts before anything reaches this site.

The first thing removed is named terms: a curated list of characters, places, spells, objects and events from the unpublished part of the manuscript, each with its variant spellings, stripped out wherever it turns up. Only two character names survive redaction, Edwin Moss and Peri Millbrook, because both are already public on the published chapters and in the book’s own blurb.

Some spoilers are not a word but a subject, so topics are handled differently: the plot threads that carry the book’s later turns are matched by pattern rather than by name and stripped whether or not they are named directly.

Unpublished scenes are visible only as filenames. Scene numbers above 40 and the planning files behind them stay in view, because a transcript where every path is a black box is unreadable, but they are tagged as unpublished and their contents never appear.

File contents are removed with one exception, applied blind. When a transcript shows a file being read or written, the site shows that it happened and what followed from it, not what was in the file. That is why the transcripts can be published at all: the prose of the novel, the canon files, and the sealed answers to the book’s unresolved mysteries all pass through those sessions constantly, and after the point where the story’s premise is first set out, none of them are shown. Before that point, in the project’s very first session, the rule is looser, and Start with the voice, not the story says more about why.

Private information is stripped too: local filesystem paths, machine details, and anything else identifying about where the work was done are removed. The novel’s own repository is not named or linked here.

Over-redaction

The redaction is deliberately blunt in places. A term list matches on strings, so an ordinary word that happens to share a name with something in the book gets removed from sentences that had nothing to do with the book. A topic pattern is just as blunt: catch a real spoiler and it will also catch an innocent use of the same word in a discussion about tooling. Both of those happen on this site, and the pipeline leaves them alone rather than hand-tuning each case. A redaction that did not need to happen just leaves a slightly confusing sentence. Missing one spoils the ending of a novel for someone reading it as it comes out.

If a passage reads as though a word has been cut out of it, a word probably has been. Occasionally a whole exchange is unreadable for the same reason.

The novel

How to Save the World is at royalroad.com/fiction/189685. It is openly labelled there as AI-written. That was the point of the exercise from the first day.