This is an early release preview. You may encounter bugs.
Autonovel logo
Unclaimed

App generative-media voice

Autonovel

Agent pipeline that drafts, revises, typesets, illustrates, and narrates a complete novel from a seed concept

No votes yet

01 / About

What Autonovel is.

Autonovel is an autonomous pipeline that takes a seed concept through drafting, revision, typesetting, illustration, and narration to produce a complete novel as a print-ready PDF, ePub, audiobook, and landing page. It applies the modify-evaluate-keep/discard loop of karpathy/autoresearch to fiction. The first novel produced this way, The Second Son of the House of Bells, runs to 19 chapters and 79,456 words.

The pipeline has four phases. Foundation generates a world bible, character registry, outline with foreshadowing ledger, canon of hard facts, and voice fingerprint, looping until a foundation score exceeds 7.5. Drafting writes chapters sequentially, keeping each one that scores above 6.0. Revision alternates automated cycles (adversarial editing, cut application, a four-persona reader panel, brief generation, and rewrites, with plateau detection) and a review loop in which Claude Opus reads the full manuscript as a literary critic and a professor of fiction until major items are exhausted. Export rebuilds the documents, typesets in LaTeX, generates art, builds the ePub, and produces audiobook scripts.

Evaluation combines two layers: a mechanical scorer that uses regex scans for banned words, fiction cliches, show-don't-tell violations, and sentence uniformity without calling a model, and an LLM judge that scores prose quality, voice adherence, character distinctiveness, and beat coverage. The novel is modeled as five co-evolving layers (voice, world, characters, outline, chapters) plus a cross-cutting canon, with propagation debts tracked in state.json.

Features

  • 27 Python tools: grouped into foundation, drafting, evaluation, revision, art and cover, audiobook, and orchestration scripts
  • Scored gates: foundation loops until foundation_score > 7.5; chapters are kept at score > 6.0
  • Two-layer evaluation: regex-based slop detection plus a separate LLM judge
  • Adversarial editing: "cut 500 words" analysis produces classified cuts applied in batch
  • Chapter tournaments: head-to-head Elo comparison of chapter versions
  • Dual-persona review: Claude Opus reviews as critic and professor with stopping conditions
  • Typesetting: LaTeX template in EB Garamond for trade paperback, ePub metadata and CSS, and print-ready full-wrap covers to Lulu/KDP specs
  • Art pipeline: cover and chapter ornaments generated with fal.ai (Nano Banana 2), curated, and vectorized
  • Audiobook: chapters parsed into speaker-attributed segments and voiced through ElevenLabs
  • Service keys: Anthropic required for the core pipeline; fal.ai and ElevenLabs optional for art and audio

02 / Discussion CREDIBILITY-GATED

Discussion

Reading is open to everyone. Posting and voting need a verified identity or a GitHub grade of B or higher.

  • No discussions yet.

04 / Build

Build with Autonovel.

Browse the catalogue for frameworks, tools, and harnesses, each scored on real GitHub credibility.

Get Autonovel →

Browse the catalogue