NousResearch/autonovel

autonovel

An autonomous pipeline for writing, revising, typesetting, illustrating, and narrating a complete novel. From a seed concept to a print-ready PDF, ePub, audiobook, and landing page — all generated by AI agents.

Inspired by karpathy/autoresearch: the same modify-evaluate-keep/discard loop, applied to fiction.

First novel produced: The Second Son of the House of Bells — 19 chapters, 79,456 words. See the autonovel/bells branch.

Quick Start

# Clone and setup
git clone <repo-url> && cd autonovel
cp .env.example .env    # Add your API keys

# Install dependencies
uv sync

# Generate a seed concept (or write your own in seed.txt)
uv run python seed.py

# Run the full pipeline
uv run python run_pipeline.py --from-scratch

The Pipeline

Phase 1: Foundation

Build the world, characters, outline, voice, and canon from a seed concept. Loop until foundation_score > 7.5.

Phase 2: First Draft

Write chapters sequentially. Evaluate each one. Keep if score > 6.0, retry if not. Forward progress over perfection.

Phase 3a: Automated Revision

Adversarial editing → apply cuts → reader panel → generate briefs → rewrite chapters. Plateau detection stops the loop when scores stabilize.

Phase 3b: Opus Review Loop

Send the full manuscript to Claude Opus for dual-persona review (literary critic + professor of fiction). Parse actionable items. Fix the top issues. Repeat until the reviewer runs out of major items.

Phase 4: Export

Rebuild docs, typeset in LaTeX, generate art, produce audiobook scripts, build ePub, create landing page.

See PIPELINE.md for the full technical specification.

Tools (27 Python scripts)

Foundation

Drafting

Evaluation

Revision

Art & Cover

Audiobook

Orchestration

File Structure

FRAMEWORK (reusable, on master):
  program.md             — Agent instructions per phase
  CRAFT.md               — Craft education (plot, character, world, prose)
  ANTI-SLOP.md           — Word-level AI tell detection
  ANTI-PATTERNS.md       — Structural AI pattern detection
  PIPELINE.md            — Full automation specification
  WORKFLOW.md            — Step-by-step human guide

TEMPLATES (filled per-novel on a branch):
  voice.md               — Part 1: guardrails. Part 2: discovered per novel
  world.md               — World bible template
  characters.md          — Character registry template
  outline.md             — Chapter outline template
  canon.md               — Hard facts database
  MYSTERY.md             — Central mystery (author-only)
  state.json             — Pipeline state tracker

TYPESETTING:
  typeset/novel.tex      — LaTeX template (EB Garamond, trade paperback)
  typeset/build_tex.py   — Chapters → LaTeX with vector ornaments
  typeset/epub_*          — ePub metadata, CSS, and front matter

ART:
  audiobook_voices.json  — Character → ElevenLabs voice mapping
  landing/index.html     — Responsive landing page template

CONFIG:
  .env.example           — API keys (Anthropic, fal.ai, ElevenLabs)
  pyproject.toml         — Python dependencies

How It Works

The novel is five co-evolving layers:

  Layer 5:  voice.md          — HOW we write
  Layer 4:  world.md          — WHAT exists
  Layer 3:  characters.md     — WHO acts
  Layer 2:  outline.md        — WHAT HAPPENS
  Layer 1:  chapters/ch_NN.md — THE ACTUAL PROSE
  Cross-cutting: canon.md     — WHAT IS TRUE

Changes propagate both down (lore change → outline change → chapter revision) and up (writing reveals a gap → update lore → check downstream). The pipeline tracks propagation debts in state.json.

Two Immune Systems

1. Mechanical (evaluate.py, no LLM): regex scans for banned words, fiction clichés, show-don't-tell violations, sentence uniformity.

2. LLM Judge (evaluate.py, separate model): scores prose quality, voice adherence, character distinctiveness, beat coverage.

The Opus Review Loop

After automated revision cycles, the full manuscript goes to Claude Opus with this prompt:

"Read the below novel. Review it first as a literary critic and then as a professor of fiction. Give specific, actionable suggestions for any defects you find. Be fair but honest. You don't have to find defects."

The dual-persona review catches what automated tools can't: prose-level repetition, character thinness, ethical gaps, structural monotony. The loop continues until the reviewer's items are mostly qualified hedges rather than real problems.

API Keys

The pipeline uses three external services:

Copy .env.example to .env and fill in your keys. Only the Anthropic key is required for the core pipeline. Art and audiobook are optional.

Production History

The first novel, The Second Son of the House of Bells, was produced through this pipeline:

• Foundation: World bible, 8 characters, 24-chapter outline, voice discovery
• Drafting: 24 chapters, 75,698 words, sequential with evaluation
• Revision: 6 automated cycles + 6 Opus review rounds
• Structural: 24 → 19 chapters through 4 merges
• Art: Linocut cover (Nano Banana 2), 19 woodcut chapter ornaments (vectorized)
• Audiobook: 19 chapters parsed into 4,179 speaker-attributed segments
• Final: 79,456 words, 6 review rounds, all major items resolved

Inspiration

• karpathy/autoresearch — the autonomous research loop
• Brandon Sanderson's writing lectures (Laws of Magic, character sliders)
• K.M. Weiland's Creating Character Arcs
• Blake Snyder's Save the Cat
• Ursula K. Le Guin's "From Elfland to Poughkeepsie"
• slop-forensics and EQ-Bench Slop Score