What it does

autonovel is a fully automated fiction factory. Feed it a seed concept and it loops through world-building, outlining, drafting, adversarial revision, and typesetting until it emits a complete novel with cover art, chapter ornaments, and multi-voice audiobook narration. The first output, The Second Son of the House of Bells, clocked in at 79,456 words across 19 chapters after 6 automated revision cycles and 6 rounds of critique by Claude Opus.

The interesting bit

The architecture treats a novel as five co-evolving layers (voice, world, characters, outline, prose) with a state.json ledger tracking “propagation debts” when changes upstream require fixes downstream. More unusually, it runs two immune systems against AI slop: a mechanical scanner (regex for banned words, clichés, show-don’t-tell violations) and a separate LLM judge for prose quality and voice adherence. The final Opus review loop uses a dual-persona prompt—literary critic plus professor of fiction—with a stopping condition that halts when the critic’s complaints dwindle to “qualified hedges.”

Key highlights

• 27 single-purpose Python scripts, not a monolithic model call: seed generation, world bible, character registry, foreshadowing ledger, canon cross-reference, Elo chapter tournaments, adversarial word-cutting, revision brief generation
• Explicit craft education baked in: CRAFT.md, ANTI-SLOP.md, and ANTI-PATTERNS.md encode Brandon Sanderson lectures, K.M. Weiland, Blake Snyder, and Ursula K. Le Guin as guardrails
• Production-grade outputs: LaTeX typesetting (EB Garamond, trade paperback), vectorized woodcut ornaments, print-ready Lulu/KDP covers, ePub with CSS, responsive landing page
• Optional audiobook pipeline parses prose into 4,179 speaker-attributed segments and maps characters to ElevenLabs voices
• Only Anthropic API key required for core pipeline; fal.ai and ElevenLabs are optional extras

Caveats

• The README is upfront that this is “forward progress over perfection” with a chapter keep threshold of merely score > 6.0
• No open-source model support mentioned; locked to Anthropic (Sonnet for drafting, Opus for review), with commercial art and audio APIs
• First and only demonstrated novel is on a separate branch; no comparative metrics or human reader feedback cited

Verdict

Worth studying if you’re building complex multi-agent pipelines or care about structured methods for reducing LLM slop in long-form generation. Not a drop-in replacement for human writing, and probably not cheap at scale—Opus review loops on 80k words add up.