I published a three-pass AI editing system on this newsletter a while back. Pass 1 handled scene-level story and prose, Pass 2 handled scene-level copyediting, and a third pass, which I called the Dead Metaphor Pass, read the whole manuscript afterward hunting for repeated imagery and physical tics.
I ran all three on real manuscript, not test snippets, and the first two held up well. Scenes got sharper; the copyedit pass caught things I’d stopped seeing after the fifth read-through, without flattening my voice into something blander along the way.
The Dead Metaphor Pass was the one that changed the most once I started running it on real chapters.
The rule I’d written for it was mechanical: if a gesture, image, or physical tic showed up three or more times across the manuscript, the model flagged it and rewrote every instance past the second. Clean logic on paper, too blunt in practice, because the rule treated frequency itself as the defect. It couldn’t tell me which occurrence was actually weak. Sometimes the first use of an image is the flat one and the fourth is the best line in the chapter. Sometimes a character rubbing her jaw four times in three hundred pages is characterization, not a tic that slipped past me. A rule firing on the count alone can’t see any of that.
So I rewrote the rule: the pass that used to trigger an automatic edit at instance three now treats three or more as a reason to look, nothing more, because frequency alone should never force a change. That’s the one number worth sitting with: the old version rewrote everything past occurrence two by default; the new version rewrites nothing by default, and asks the model to justify each flag on what the repetition is doing, not on how often it happens.
Rewriting that rule surfaced a bigger question. The three-pass system could tell me whether a scene worked and whether the prose was clean, and once revised, whether an image or gesture had gone stale across four hundred pages. What it never told me, not once, was whether the novel worked as one complete story: whether the setups paid off, whether the middle actually escalated instead of repeating itself with new scenery, whether the ending resolved the thing the book had actually been about, or something merely adjacent to it.
No amount of scene-level polish answers that, because it isn’t a scene-level question; the system is now four passes instead of three, with a new manuscript-level structural audit sitting between the copyedit pass and the renamed pattern pass. What each one does, why the reordering matters, and the full prompts for all four are below.
