____              _
    | __ )  ___   ___ | | __
    |  _ \ / _ \ / _ \| |/ /
    | |_) | (_) | (_) |   <
    |____/ \___/ \___/|_|\_\
     ____                      _
    / ___| ___ _ __   ___  ___(_)___
   | |  _ / _ \ '_ \ / _ \/ __| / __|
   | |_| |  __/ | | |  __/\__ \ \__ \
    \____|\___|_| |_|\___||___/_|___/  V4

21 AI skills. One idea. A publish-ready book.

Quick Start · Pipeline · Genesis Score · Skills · Why It's Different · Architecture · FAQ

Book Genesis is the only open-source system that turns a single idea into a full manuscript using 21 specialized AI skills that research, write, evaluate, disrupt, and revise -- the same pipeline used to produce a 68,000-word memoir scoring 9.0 on an empirically-calibrated quality framework benchmarked against 15 bestsellers representing 350M+ copies sold.

No external APIs. No databases. No build step. Just Claude Code and markdown files.

Quick Start

Prerequisites: Claude Code CLI installed + Claude Pro ($20/mo) or Max ($100/mo for full novels)

Install (one command):

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/PhilipStark/book-genesis/main/install.sh | bash

# Windows (PowerShell)
irm https://raw.githubusercontent.com/PhilipStark/book-genesis/main/install.ps1 | iex

Write a book:

claude
> /book-auto pt-br "thriller sobre ex-policial vingando a família"

That's it. The orchestrator runs the full 17-phase pipeline autonomously. It pauses at 3 checkpoints for your approval. Everything else is automatic.

What's New in V4

V4 is a ground-up rebuild. We ran 5 genre tests, hired a publishing consultant to analyze the output, identified 9 systemic problems, and fixed every one. Here's what changed:

The 10 New Anti-AI Patterns

V3 caught the obvious stuff (forced symmetry, empty metaphors, rule of three). V4 catches what makes AI prose feel AI even when it's technically clean:

Pattern #11 (Explanatory Extension) appeared in every chapter of a 14-chapter test manuscript. It was the single most AI-identifiable pattern. V4 kills it at the sentence level.

What's New in V4.1

V4.1 adds persistent entity tracking and standardizes the pipeline for global use:

The 17-Phase Pipeline

  IDEA
   |
   v
Phase 1:   RESEARCH ----------> /book-researcher
Phase 1.5: READER PERSONAS ---> /reader-persona
Phase 2:   FOUNDATION ---------> /narrative-foundation
Phase 2.5: VOICE DNA ----------> /voice-fingerprint
   >>> CHECKPOINT 1: You approve the foundation <<<
Phase 2.7: ENTITY TRACKING ----> /entity-tracker (BUILD)          [NEW in V4.1]
Phase 2.8: CONTINUITY ---------> /continuity-guardian (outline)
Phase 3:   WRITING ------------> /prose-craft
Phase 3.1: DIALOGUE POLISH ----> /dialogue-polish
Phase 3.2: HOOK CRAFT ---------> /hook-craft
Phase 3.5: DISRUPTION ---------> /chaos-engine
Phase 3.7: ENTITY UPDATE ------> /entity-tracker (UPDATE)         [NEW in V4.1]
Phase 3.8: MECH PREPROCESS ----> /mechanical-preprocess
Phase 4:   EVALUATION ---------> /beta-reader
Phase 4.5: QUALITY GATE -------> /quality-gate
   >>> CHECKPOINT 2: You approve the manuscript <<<
Phase 5:   REVISION -----------> /book-editor
Phase 5.5: ENTITY UPDATE ------> /entity-tracker (UPDATE)         [NEW in V4.1]
Phase 5.6: CONTINUITY ---------> /continuity-guardian (manuscript)
Phase 6:   DELIVERY -----------> /editorial-package + /production-prep
   >>> CHECKPOINT 3: Book delivered <<<

Every phase has a gate. Nothing advances without passing. The revision loop (Phases 4-5) continues until the Genesis Score floor hits 8.0+ or 3 cycles are reached.

Why This Is Different

The Problem

Every AI writing tool produces the same thing -- competent, forgettable prose. You can generate 50,000 words in an afternoon and every page reads like it was written by the same polite, analytical, emotionally-managed narrator. Literary fiction reads like thrillers. Memoirs read like self-help. Five genres, one author.

This is the Claude House Style: similes that get explained, dialogue in orderly turns, characters who notice their emotions and manage them successfully, settings described with encyclopedic confidence, and themes that echo in every detail like a hall of mirrors.

It scores a 7 on any rubric. It gets rejected by every editor who reads past page 3.

What We Did

We ran 5 genre tests (literary fiction, thriller, memoir, prescriptive non-fiction, commercial fiction). A publishing consultant analyzed the output and identified 9 systemic problems. The full analysis is in docs/system-analysis.md.

Then we rebuilt the pipeline from the ground up.

The Key Insight

The gap between "competent" and "unforgettable" is not about better prompts. It is about controlled chaos.

Real humans have irrelevant thoughts during important moments. They fail to manage their emotions. They remember things at the wrong time. They interrupt each other. They notice the wrong details. Their prose rhythm breaks under pressure.

AI does none of this. AI is relentlessly competent.

That is why V4 has a Chaos Engine -- an agent whose entire job is to make things worse in a specific, human way. It injects intrusive thoughts that have nothing to do with the plot. It makes characters fail to control their composure. It breaks the prose rhythm when a character is stressed. It adds texture details that carry zero thematic weight.

It is the only writing pipeline with an agent designed to break its own output.

The Benchmark

We calibrated the Genesis Score V3.7 against 15 bestsellers representing 350M+ copies. The system correctly identifies what makes bestsellers work and what makes them different from merely good books.

Average accuracy: 85%. Worst case: 70% (The Alchemist -- parable fiction has a documented accuracy ceiling due to non-craft factors like gift-ability and translation ease).

Autonomous Mode

V4.1 introduces /book-auto — a single command that runs the entire pipeline:

/book-auto en "a grief counselor discovers her five patients are connected to the same murder"

The orchestrator agent runs up to 200 turns autonomously. It only pauses at 3 checkpoints:

1. Foundation approval — You see the characters, outline, and voice specs before writing starts
2. Manuscript approval — You see the full score breakdown before the editorial package
3. Delivery — You receive the finished book

Everything between checkpoints is automatic: writing, dialogue polish, hooks, disruption, mechanical cleanup, evaluation, quality gate loops, revision, continuity checks.

For manual control, use /book-genesis instead — same pipeline, you drive each phase.

Genesis Score V3.7

A 7-dimension FLOOR scoring system. Your book's score equals its weakest dimension. You cannot hide weak characters behind brilliant prose.

Anti-inflation protocol: Evidence required for every score. Max +0.5 per iteration. Reference comparison against published titles for scores >= 9.0. Cross-dimensional verification (if prose goes up, check pacing didn't drop).

New in V3.7:

• CVI bifurcated into CVI-Launch (first-year sales prediction) and CVI-Legacy (20-year sales prediction)
• 5 engagement types with adjusted CVI weights (Aspiration books get Identity Effect at 25%)
• 8 genre profiles with different thresholds across all dimensions
• Scope declaration -- the evaluator states what the score covers and what it cannot measure
• Oscillation analysis -- detects scores bouncing between iterations (sign of prompt-gaming)

Full scoring methodology: docs/genesis-score.md

The 21 Skills

Each skill works standalone or orchestrated. Use /book-auto for full autonomous mode, /book-genesis for manual control, or any individual skill for specific tasks.

Character Chaos System

V3 characters had wound, lie, and arc. Solid, but every AI writing tool does this now.

V4 characters have 5 chaos markers -- the things that make fictional people feel like actual people:

These are not plot points. They are human noise. They exist because people are not narrative vehicles.

Proof of Concept

Protocolo Nao Encontrado -- a 68,000-word memoir in Portuguese that went through the full V4 pipeline.

• Genesis Score floor: 9.0
• 3 scoring iterations: 8.70 -> 8.95 -> 9.04
• 13 chapters across 3 parts + introduction + epilogue
• Full editorial package (query letter, synopses, cover brief)

Artifacts in examples/protocolo-nao-encontrado/:

Additional translations available: synopses-en.md (English), synopses-es.md (Spanish)

The full manuscript is not included (author IP). These artifacts demonstrate what the system produces.

Age of Aquarius -- a 97,000-word hermetic fantasy in English that scored 9.16 on Genesis Score V3.7.

• Genesis Score floor: 9.16
• 4 scoring iterations: 7.80 -> 8.40 -> 8.90 -> 9.16
• 28 chapters across 4 parts
• Full evaluation artifacts

Artifacts in examples/age-of-aquarius/:

Benchmarks

The Genesis Score V3.7 was calibrated against published bestsellers:

*The Alchemist represents a documented accuracy ceiling for parable fiction -- non-craft factors (gift-ability, translation ease, spiritual timing) account for ~30% of its commercial success and fall outside any craft-based framework.

Average accuracy: 85%. The framework correctly identifies what makes bestsellers work -- and more importantly, what separates them from the thousands of competent books that sell 500 copies.

Requirements

Installation

One command

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/PhilipStark/book-genesis/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/PhilipStark/book-genesis/main/install.ps1 | iex

Manual

git clone https://github.com/PhilipStark/book-genesis.git
cd book-genesis
# macOS / Linux
./install.sh
# Windows
./install.ps1

The installer copies skill files to ~/.claude/skills/ and verifies Claude Code can find them.

How It Works Under the Hood

1. Skills are markdown files (SKILL.md) installed in ~/.claude/skills/
2. Claude Code loads them as custom slash commands (e.g., /book-genesis)
3. The orchestrator reads PROJECT_STATE.yaml to resume across sessions
4. Every artifact is a file -- outlines, scores, characters, the manuscript itself
5. The revision loop is driven by the Genesis Score: write -> disrupt -> score -> identify weakest dimension -> revise -> re-score
6. The FLOOR system means the weakest dimension drives all revision effort

No external APIs. No databases. No build step. No dependencies beyond Claude Code.

Contributing

We want genre testers, skill translators, and anyone who has tried to make AI write something worth reading.

• Test it -- Write a book and report what works and what breaks
• Translate skills -- All core skills are now in English; translations to other languages welcome
• New genre profiles -- Test with specific genres and submit calibrated profiles
• Improve the Genesis Score -- Better rubrics, better anti-inflation, better calibration data
• Add examples -- Share your proof-of-concept artifacts

See CONTRIBUTING.md for guidelines. Open an issue before starting major work.

How It Compares

Star History

License

MIT -- use it, fork it, sell books with it.

Built with Claude Code by Philip Stark
_{If this is useful, a star helps more than you think.}