autonovel is an autonomous pipeline that takes a novel from a seed concept through writing, revising, typesetting, illustrating, and narration, ending with a print-ready PDF, ePub, audiobook, and landing page. It borrows the modify-evaluate-keep/discard loop from karpathy's autoresearch project and applies it to fiction, orchestrating 27 Python scripts across four phases: foundation (world bible, characters, outline, voice, canon, looped until a quality threshold), sequential chapter drafting with per-chapter scoring and retries, automated revision (adversarial editing, a four-persona reader panel, revision briefs, with plateau detection), and export.
Quality control combines two "immune systems": a mechanical, no-LLM scorer that regex-scans for banned words, fiction cliches, show-don't-tell violations, and sentence uniformity, plus an LLM judge scoring prose quality, voice adherence, and beat coverage. After automated cycles, the full manuscript goes through a Claude Opus dual-persona review (literary critic + professor of fiction) that repeats until major actionable defects are exhausted. The novel itself is modeled as five co-evolving layers (voice, world, characters, outline, prose) plus a canon of hard facts, with propagation debts tracked in a state file so changes flow both up and down the stack.
The core pipeline needs only an Anthropic key; cover art and ornaments (via fal.ai) and multi-voice audiobook generation (via ElevenLabs) are optional stages. The repo documents one completed production run: a 19-chapter, 79,456-word novel that went through six automated revision cycles and six Opus review rounds.
