Introduction

David MacIver has been publishing a novel, How to Save the World, on Royal Road, and it was written with Claude Code. Enough people asked him how that actually worked that it seemed worth answering properly rather than in fragments, so this site sets out what the process looked like, what tooling it needed, what went wrong, and what the day-to-day record of ten days actually contains.

You can read the novel here: How to Save the World on Royal Road.

Every word on this site was generated by Claude from the project’s git history and its Claude Code session transcripts. David set the redaction rules and the direction, and read the result, but no page here was hand-written by him. That includes this paragraph. It is why the pages stick close to what the record actually shows, and say so when it does not show much.

What is left out

The novel is still being published, so the site is redacted against spoilers. Anything past the currently published chapters is removed: character names, place names, spells, events, and the plot. Two names survive, Edwin Moss and Peri Millbrook, because both are on the published page already. File contents are never shown, only the fact that a file was read or written and what happened next. The full policy, including where the redaction is deliberately heavier than it strictly needs to be, is on About this site.

How to read it

There are four sections, and they cover the same ten days from different angles.

How to do this yourself is the place to start if your question is practical: what the back and forth between author and model actually looks like, which files and checks a project like this needs, when to throw work away, and which lessons should transfer to a different project. It stands back from the detail, and links out to the other sections when it needs the evidence.

The workflow is the systems view. It covers the rule files that bind a drafting session, the exemplar-first method for holding a voice, the linting and review passes that hunt for habits typical of machine-written prose, the canon system, and the handoff document that lets a fresh session pick up where the last one stopped, plus the tooling for proofreading, commenting, and publishing. Read this if you want to know how the thing was built.

The chronicle is the narrative view: one page per day, from setting the premise on 24 August 2026 to finishing the draft on 2 September. It is reconstructed from commits and transcripts, so it is a record of what happened rather than of what anyone intended. A good deal of the ten days went on work that was later thrown away.

The transcripts are the raw material: the redacted Claude Code sessions themselves, prompt by prompt. They are long and they are not edited for readability. If a claim on one of the other pages looks too tidy, this is where to check it.

The workflow pages and the chronicle link to each other where a rule and the failure that produced it belong together, which is most of the time.