Repository navigation
The skill now builds a rewrite from the writer's notes. The post follows the notes and drops the tutorial scaffolding the draft added around them, like a "basics" section or a closing summary. Sections end on their last fact and the rewrite leaves no placeholder from the draft behind. The skill also names the sincerity markers Claude adds: earlier rewrites put honestly in a third of posts, and genuinely now counts in the AI lexicon.
After the loop converges, a fresh agent reads the text cold, once, and the rewriter fixes the reasons that point at a sentence. Stories get their own rules in references/fiction.md, about the story's shape: callbacks and bookends, a clockwork arc and a moral stated outright.
Measured with blind Claude Opus judges who read each text alone (human originals score 4 to 8): blog posts went from a mean P(AI) of 47.8 to 34.9 on 20 titles, better on every one. The first of these changes took held-out titles from 36.7 to 31.8. Spanish posts went from 38.8 to 29.2, held-out stories from 74.0 to 69.7, and rewrites by GPT-5.6 through Codex from 67.0 to 55.3. The harness behind these numbers is in eval/e2e/, and references/numbers.md records them along with the candidates that were measured and left out.