Patrick Rothfuss's Kingkiller Chronicle has been waiting for Book 3 for 14 years. I spent 8 months building a system to finish it. The result is 113 chapters, 301,000 words, free at thethirdsilence.com
Here's how:
Story Bible (56,000 words)
Before generating any fiction, I treated both novels as a codebase. Every character, location, lore element, and unresolved thread got extracted into structured reference docs. Every Chekhov's gun catalogued. Both timelines (the inner narrative and the frame story) mapped with convergence constraints.
Drafting
Used Claude primarily. Per-chapter workflow: feed the relevant story bible entries, previous chapters for continuity, the chapter outline, and 3-5 representative passages from the original books for voice matching. Generate. Read. Write revision notes. Regenerate. Repeat 3-8 times per chapter.
Raw output was never usable as-is. The specific failure modes I kept hitting:
Flattened voice. LLMs default to generic literary prose. Rothfuss has an extremely specific cadence — short declarative sentences, metaphors that feel inevitable, silence as punctuation. Required extensive examples in-prompt and aggressive post-editing.
Conflict avoidance. Models resolve tension prematurely. Characters reconcile too fast. I had to explicitly instruct "leave this unresolved and painful."
Referential drift. By chapter 80, the model has forgotten chapter 12. The story bible helped but I still caught contradictions every pass.
The em-dash problem. First draft: ~40% of paragraphs had em-dashes. Final: under 10%.
Emotional genericism. "She felt a wave of sadness" vs. making the reader feel it through concrete detail without naming it. Hardest thing to fix.
Continuity Checking
Built a system that cross-referenced every chapter against the story bible and all preceding chapters. Flagged character inconsistencies, timeline errors, lore contradictions, repeated phrases, and LLM prose tells. This caught hundreds of errors. It later became its own product (galleys.ai).
Voice Editing
Three full passes just for voice. Stripped hedging language, overexplained motivations, stacked metaphors, wrong-rhythm sentences. Cut 80,000 words total.
The Split
Roughly 30% of final prose originated from LLM output that survived editing. 70% is rewriting, restructuring, and original bridging material. Plot, structure, and creative decisions are entirely mine.
The book is free: thethirdsilence.com (browser, EPUB, PDF)
The editorial pipeline became Galleys: galleys.ai
Happy to answer questions about any part of the process.
Geobond•1h ago
Flattened voice. LLMs default to generic literary prose. Rothfuss has an extremely specific cadence — short declarative sentences, metaphors that feel inevitable, silence as punctuation. Required extensive examples in-prompt and aggressive post-editing. Conflict avoidance. Models resolve tension prematurely. Characters reconcile too fast. I had to explicitly instruct "leave this unresolved and painful." Referential drift. By chapter 80, the model has forgotten chapter 12. The story bible helped but I still caught contradictions every pass. The em-dash problem. First draft: ~40% of paragraphs had em-dashes. Final: under 10%. Emotional genericism. "She felt a wave of sadness" vs. making the reader feel it through concrete detail without naming it. Hardest thing to fix.
Continuity Checking Built a system that cross-referenced every chapter against the story bible and all preceding chapters. Flagged character inconsistencies, timeline errors, lore contradictions, repeated phrases, and LLM prose tells. This caught hundreds of errors. It later became its own product (galleys.ai). Voice Editing Three full passes just for voice. Stripped hedging language, overexplained motivations, stacked metaphors, wrong-rhythm sentences. Cut 80,000 words total. The Split Roughly 30% of final prose originated from LLM output that survived editing. 70% is rewriting, restructuring, and original bridging material. Plot, structure, and creative decisions are entirely mine. The book is free: thethirdsilence.com (browser, EPUB, PDF) The editorial pipeline became Galleys: galleys.ai Happy to answer questions about any part of the process.