Expect to throw work away

The book on Royal Road is the third draft. The first was a complete draft of book one, all three acts, written on the second day of the project and reclassified as groundwork on the fourth. The second included a full second act of some sixty-five scenes, drafted in a single push on 30 August and pulled the same evening. Neither will be read by anyone. Both were necessary. Discarding them cheaply mattered more to the published book than any individual rule about prose.

With a model producing text this fast, it’s tempting to treat every draft as an asset. Most of them are not. They are how you find out what the rules should have been.

The two big discards

The first draft was written fast, before the voice rules had been tested against real output at length and before the dialogue work described on the loop page. By 27 August enough had been learned about what was wrong with it that fixing it in place was going to cost more than starting again, so it was declared groundwork. The chronicle entry records that this made a batch of earlier decisions about the book’s voice moot, and nothing published depended on them.

The second discard was sharper. On 30 August a whole act was drafted with the canon and editorial checks running as it went, and it passed them. Then the overlap checker, written minutes earlier that same evening, showed that the new prose was too close to earlier rejected material. It read as recycled. The act was moved into a folder marked do not open, a rule about this specific kind of contamination went straight into the voice guide, and one scene was redrafted from a clean brief that night, with a second the next morning, to confirm the new process worked before anyone committed to redoing the rest.

What the signs were

In both cases the individual scenes read acceptably, but the draft as a whole was wrong. The problem was in the whole book, not any one scene: a voice that had drifted generic across it, or an act whose phrasing echoed something that should have been forgotten. Checks that run scene by scene do not catch this. A reader reading a stretch cold does, and so does a tool that looks across the whole corpus. That’s why the project’s most consequential tools, the census and the overlap checker, are corpus-level rather than scene-level.

The other sign was a fix that kept growing. When repairing a draft turns into re-engineering scene after scene, and each repair reveals the next, that is the sign the underlying rules have changed since the draft was written.

Making discarding cheap

A few things made the project’s discards survivable. Drafts lived in version control, so throwing one away meant moving a directory, not losing anything. Whatever a bad draft taught the project was already written into the rule files by the time it was discarded. And the exemplars, the small set of pieces David actually liked, were kept separately from the drafts, so a discard never took the voice with it. The one time this went wrong, early exemplar chapters were deleted as scaffolding on the first day. It took a few hours that same afternoon to work out that the voice had drifted because of it, and to restore them.

Steps you can take

1. Treat the first full draft of anything as a probe. Its purpose is to show you which of your rules are wrong. Have the model write the rules it reveals into your files before you decide whether to keep it.
2. Keep exemplars separate from drafts, and under a rule that they are never deleted without a replacement being chosen first.
3. Every so often, read a long stretch cold, as a reader, and ask the corpus-level questions. Does this sound like one voice throughout? Does any of it feel familiar in a bad way? Have the model build tools that ask the same questions mechanically. The overlap checker on this project was requested in one message and built in minutes.
4. When you decide to discard, do it completely. Move the material out of anywhere a drafting session might read it, and if you need to check new work against it, do so through a tool that reads it from history rather than from the working tree. The overlap checker page explains why this matters.
5. Before redoing everything, redo a small piece under the new rules and check it. The project redrafted one scene the night of the big discard for exactly this reason.