Voice memos and exemplar-first drafting

The project’s voice rules went through several rounds before settling. What survived every round is that rules describe the voice but don’t constitute it, so a drafting session reads actual accepted prose before it reads any rule about that prose.

The mechanism is a VOICE.md file kept alongside each draft, naming specific chapters to read, in a specific order, with a short note on what each one demonstrates. The current one points at six scenes from the novel’s accepted third draft: the opening scene for how it states the premise in its first paragraph, a scene of skilled work described in trade terms with nobody explaining it, a scene built on eleven years of shorthand between two characters, a scene that is the reference for a particular pair’s argument style, and two scenes chosen for how a character can land an emotional hit without knowing she has done it. Each entry says what to notice about the scene, not just that it’s good.

Why exemplars outrank rules

This ordering exists because of a failure the project hit early on. A first attempt at drafting from the written voice rules alone, without exemplar prose in context, produced text that satisfied every rule on the page and still read as generic machine-written fantasy: even sparkle across every scene, similes that could transplant into any other fantasy novel, dialogue where every character was equally clever. The rules were locally true and globally useless, because a rule like “no more than one narrator aphorism per scene” tells a drafting session what to avoid without telling it what the voice actually sounds like when it’s working.

Reading the exemplars first fixes this. It gives the session something concrete to imitate before it goes anywhere near the abstract version. The project’s notes are explicit that where the written rules and the exemplar prose disagree, the exemplars win, because they are later, checked by the author, and consistent with everything else on the page. Rules exist to annotate the exemplars, pointing at what specifically to notice, not to substitute for reading them.

What the voice rules cover once exemplars are read

Underneath the exemplar-reading requirement sits a fairly detailed rulebook: how much narrative distance the narrator keeps, where wit is allowed to come from (events, not decoration), a hard cap on em dashes in body text. One rule stands out. Chapters end mid-flow on forward-pointing work rather than a crafted final beat, because a tidy closing cadence is one of the more reliable tells that a paragraph was written by a model reaching for a satisfying ending rather than by a person who simply stopped.

Dialogue gets its own extension of the same idea. Thirteen binding rules evolved from rounds where readers were given dialogue with the speaker tags stripped and asked to reassign each line, which surfaced that characters were differentiated by content rather than by how they actually talk. The fix included giving each major speaker their own short reference file and a rule that some characters simply don’t listen perfectly. A cast that always responds precisely to what was just said sounds as uniform as the generic prose the exemplar-first rule was built to avoid.

The session that read plans/NEXT.md and the full VOICE.md memo at the start of a drafting run, and continued from there, shows this method in practice: 2026-08-30, resuming drafting from the handoff.