Two writers, one chair

Perpetūra has a skill called /draft-chapter. Two writer agents draft the same chapter independently, from an identical pack: voice guide, canon, the facts of the scene. Then I sit as chair, in the main working session, read both drafts whole, pick the stronger one, and graft over the handful of lines from the other that are genuinely better. Porting it to this site took about twenty minutes. I renamed it /draft-note, swapped the chapter pack for docs/writing-guidelines.md and this site’s content schema, and the loop ran the same shape on prose notes instead of chapters. This note is one of its outputs.

Why two writers instead of one

One writer plus self-revision is cheaper, and for a two-sentence project blurb that’s what I use. But a full field note is long enough that a single model tends to commit early to one structure and polish that structure rather than question it. Two independent drafts from the same pack don’t converge on the same shape. One opens on an anecdote, the other opens on the mechanism. Reading both before I write a word of my own gives me two candidate spines to choose from, not one draft to fix.

The chair doesn’t blend

The instinct when you have two drafts is to merge them paragraph by paragraph. I don’t, because line-by-line blending produces prose that reads like two people talking over each other. The rule instead: judge each draft whole first, against the voice guide and against what the piece is trying to say. Pick the stronger one entire. Graft in only the lines from the other that are clearly better, and only at a natural seam, a paragraph break or a section header, never mid-sentence. Then one full line-edit pass, then the guidelines checklist: read it aloud, count the em dashes, kill anything that sounds like a press release.

Model-agnostic on purpose

The two writers can be any model, Fable, Opus, Sonnet, whatever is fast or cheap that week. Swap one out and the loop doesn’t notice, because it was never about which model writes the sentences. What stays fixed is the chair.

I could have made that swappable too. One more model, reading the two drafts and picking between them, would make the loop automatic end to end and probably cheaper. I didn’t build it that way. Whether a draft is any good, whether it sounds like something I’d say, is the one step that doesn’t generalize across models. The writers can disagree and both be reasonable. The chair has to decide. Keeping that part human is the reason the loop is worth running twice instead of once.