A 79,456-Word Novel, Written by 27 Python Scripts

What it does

autonovel is a Python framework that drives AI agents through a four-phase pipeline to produce a complete novel. Starting from a seed concept, it builds a world bible, character registry, and outline, then drafts chapters sequentially with automated evaluation. The system revises the manuscript through adversarial editing, reader-panel scoring, and a dual-persona review by Claude Opus, then exports a print-ready PDF, ePub, cover art, and a multi-voice audiobook.

The interesting bit

The project treats a novel as five co-evolving layers—voice, world, characters, outline, and prose—where changes propagate up and down with tracked propagation debts in state.json. It also runs two immune systems against AI slop: a mechanical scanner that regexes for clichés and banned words, and an LLM judge that scores voice adherence and beat coverage.

Key highlights

• Shipped one real novel: The Second Son of the House of Bells (79,456 words, 19 chapters) on a dedicated branch.
• Uses a mechanical “slop” scanner (evaluate.py, regex-based) plus a separate LLM judge to evaluate drafts, with explicit score thresholds for keeping or retrying chapters.
• Implements an Opus Review Loop where Claude Opus acts as both literary critic and fiction professor, iterating until the review exhausts major actionable defects.
• Tracks cross-cutting canon in a hard-facts database and propagates lore changes downstream to outline and prose automatically.
• Generates print-ready LaTeX typesetting (EB Garamond), vector cover art, and multi-voice ElevenLabs audiobook scripts.

Caveats

• Requires paid API access: Anthropic keys are mandatory, while fal.ai and ElevenLabs are needed for art and audiobook output.
• The first novel required six automated revision cycles and six Opus review rounds, so compute costs and wall-clock time are significant.
• Only one complete novel has been produced so far, leaving the pipeline’s generalizability across genres unproven.

Verdict

Worth studying if you are building agentic long-form generators or automated publishing workflows; avoid if you expect a cheap, one-click bestseller machine.