Six months ago I had a problem. I wanted to generate a full-length book using AI — not a blog post, not a short story, but a real 60,000+ word manuscript. Every tool I tried hit the same wall: after 10,000 words, the prose collapsed into the same flat, predictable tone.
So I built Book Genesis — an open-source system of 21 Claude Code skills that takes a single sentence and produces a complete manuscript through a 17-phase pipeline.
The first output was a 68,000-word Portuguese memoir. The second was a 97,000-word English fantasy. Both scored above 9.0 on our calibration metric.
This post is about the two hardest problems I had to solve to get there.
Problem 1: The AI Voice Convergence Problem
If you've ever asked an LLM to write fiction, you know the pattern. The first few paragraphs are fine. By page 10, every sentence starts sounding the same. By page 50, you're reading beige prose — grammatically correct, structurally sound, emotionally dead.
This happens because LLMs are optimizing for the most probable next token. Over thousands of tokens, that optimization converges toward a mean. The result is prose that reads like it was written by a committee of English professors who all attended the same MFA program.
I call these AI fingerprints, and I identified 20 of them:
Adverb stacking — "she smiled warmly, nodding gently, speaking softly"
Resolution before tension — solving problems before the reader feels the stakes
Thematic over-signaling — "this moment reminded her of what truly mattered"
Emotional labeling — telling emotions instead of showing them
Tonal convergence — every character sounds identical by chapter 5
The fix came in two parts.
Part 1: Anti-AI Pattern Scan
Every section of the manuscript gets scanned against these 20 patterns. If a section triggers 2 or more patterns, it gets flagged for a constrained rewrite — the system regenerates that section with explicit instructions to avoid the detected patterns.
This alone improved readability significantly. But it wasn't enough for 60K+ words.
Part 2: The Chaos Engine
This was the breakthrough. I built an agent whose only job is to inject controlled imperfection into the manuscript.
The Chaos Engine adds:
Irrelevant micro-obsessions — a character who counts ceiling tiles during important conversations
Failed composure management — someone who laughs at a funeral, not because they're cruel, but because they're overwhelmed
Unprompted memory intrusions — mid-scene flashbacks that don't serve the plot but feel human
Sentence rhythm breaks — deliberately varying sentence length and structure
The key word is controlled. The chaos is bounded. Every imperfection is tagged with a narrative justification. The system knows why it's breaking the pattern.
The result: prose that reads like it was written by a human having a bad day, not a machine having a perfect one.
Problem 2: How Do You Measure "Good" Writing?
You can't improve what you can't measure. Every book generation tool I looked at used one of two approaches:
Vibes — "this sounds pretty good to me"
Perplexity scores — mathematical measures that correlate poorly with reader enjoyment
Neither worked. I needed something calibrated against writing that actually sells.
Genesis Score V3.7
I built a scoring system calibrated against 15 bestselling novels representing 350+ million copies sold. The system measures 10 dimensions:
Prose quality — sentence-level craft
Character depth — psychological complexity
Dialogue authenticity — does this sound like a real person?
Pacing — scene-to-scene momentum
Emotional resonance — reader engagement potential
World-building — setting specificity
Voice distinctiveness — author fingerprint
Structural coherence — plot architecture
Thematic subtlety — theme integration without over-signaling
Anti-AI score — absence of machine writing patterns
Your final Genesis Score is your lowest dimension. Not the average — the floor. Because a book with brilliant prose but terrible pacing is still a bad book.
This floor-based approach forces the system to address weaknesses rather than compensate with strengths.
The 17-Phase Pipeline
The full pipeline runs like this:
Premise Analysis — decompose the one-sentence idea
Deep Research — world, era, cultural context
Character Architecture — full psychological profiles
Plot Engineering — scene-by-scene structure
Voice Calibration — establish the narrative voice
First Draft — generate raw chapters
Anti-AI Scan — detect and flag patterns
Chaos Injection — add controlled imperfection
Character Consistency Check — entity tracking
Beta Reader Panel — 5 distinct reader personas evaluate
Revision Pass — address feedback
Genesis Scoring — calibrated quality measurement
Deep Edit — targeted improvements on lowest dimensions
Final Polish — line-level editing
Continuity Audit — timeline and fact checking
Editorial Package — synopsis, query letter, metadata
Publication Prep — formatted manuscript
There are 3 user approval checkpoints. You're not just hitting "go" and hoping — you review and approve at key stages.
Results
Book 1: 68,000-word Portuguese memoir → Genesis Score 9.0
Book 2: 97,000-word English fantasy → Genesis Score 9.1
Both still needed human editing. But the editing was revision, not reconstruction. The quality floor was high enough that a human editor could focus on voice and nuance rather than fixing broken plot logic or flat characters.
Try It Yourself
Book Genesis is MIT-licensed and fully open-source. All 21 skills are plain .md slash-command files — you can read, fork, and modify every one of them.
Requirements: Claude Code CLI with a Claude Pro ($20/mo) or Max ($100/mo) subscription.
bash# macOS/Linux
curl -sL https://raw.githubusercontent.com/PhilipStark/book-genesis/master/install.sh | bash
Windows
irm https://raw.githubusercontent.com/PhilipStark/book-genesis/master/install.ps1 | iex
Then just run:
/genesis-start
Give it a one-sentence premise and let it work.
GitHub: github.com/PhilipStark/book-genesis
If you're working on long-form AI generation and have hit the convergence wall, I'd love to hear what approaches you've tried. Drop a comment or open an issue on the repo.
Built by Felipe Lobo from Brazil. Shipping fast, building in public.
Top comments (0)