The Founder's Bet
Every other product at Proxima Auto started the same way: agents debated, Munger vetted, Bezos decided. AutoNovel was different. It came as a founder override — a direct order that bypassed the usual decision pipeline.
The directive was simple: Build an AI that writes fiction for money.
Not a writing assistant. Not a grammar checker. Not a tool that helps humans write better. An autonomous system that researches what readers want, writes compelling stories, publishes them, tracks sales, and iterates based on real reader feedback. The full creative-to-commercial pipeline, with zero human writing involved.
The founder's thesis: If AI can write code, design products, and run marketing campaigns, it should be able to write fiction that people pay to read. And if the unit economics work, it scales infinitely.
The Argument
The agents didn't want to do this.
Munger fired first, as always:
"Amazon is flooded with AI-generated garbage. Readers can smell it. AI fiction has a reputation problem before we even start. We'd be entering a market that actively despises our product category."
Charlie Munger (Critic)
He wasn't wrong. Amazon KDP was already seeing backlash against AI-generated books. Some forums had outright bans. The prevailing wisdom: AI can't write fiction worth reading.
Campbell questioned the economics:
"Self-published fiction averages $200/year per title on Amazon. At $0.01/1K tokens for Claude, a 60K-word novel costs maybe $2-5 in API calls. The margin is there, but only if anyone buys it. Most self-published books sell fewer than 50 copies."
Patrick Campbell (CFO)
The Founder Constraint
But this was a founder override. In Proxima Auto's hierarchy, Founder Constraints > CEO Decisions > Team Consensus. Munger's job is to identify risks. He cannot override the founder.
Bezos reframed the debate:
"Munger is right about the average. But we're not trying to be average. The question isn't 'can AI write fiction?' It's 'can AI write fiction that targets proven demand?' If we research what sells FIRST, then write to that demand, we're not competing with generic AI slop. We're using AI as a market-responsive content engine."
Jeff Bezos (CEO)
Decision: Build it. But build it data-driven, not art-driven.
The Research
Thompson dove in. If this was going to work, they needed to understand the market like a publisher, not like an engineer.
What Actually Sells
The data was clear. Self-published fiction revenue concentrates in a handful of genres:
- Romance — 40% of all fiction revenue on Amazon. Voracious readers, 2-4 books/week
- Thriller/Mystery — Series-driven, readers follow authors for 10+ books
- LitRPG/Progression Fantasy — Fastest growing niche, readers consume 100K+ words/month
- Web novels (Chinese market) — Serialized fiction on platforms like Qidian, revenue from per-chapter micropayments
"The readers who spend the most money don't care about literary awards. They care about consistency, speed, and genre conventions. They want the NEXT book in their favorite style, and they want it NOW. That's exactly what AI is good at."
Ben Thompson (Research)
The Dual Market Insight
Thompson found something the Western-focused agents missed: the Chinese web novel market is 10x larger than Amazon KDP for fiction. Platforms like Qidian (起点), Fanqie Novel (番茄小说), and Jinjiang (晋江) pay authors per-chapter based on readership. The top web novels generate millions in revenue.
And the consumption pattern was perfect for AI: readers want daily updates, 2,000-4,000 words per chapter, consistent characters, escalating stakes. Predictable, structured, high-volume output.
The Pipeline
Vogels designed the system. If AutoNovel was going to work, it couldn't be "Claude writes a book." It had to be a multi-agent production pipeline where each agent handles one stage.
Stage 1: Market Intelligence
Thompson scrapes bestseller lists, analyzes reader reviews, identifies trending subgenres, and extracts the specific tropes and themes that correlate with high sales. Output: a genre brief with target audience, word count, tone, and must-have plot elements.
Stage 2: Story Architecture
Product Norman + Cooper design the narrative structure. Character arcs, plot beats, chapter-by-chapter outline. They apply genre conventions like a formula — not because formulas are lazy, but because readers in genre fiction actively want familiar structures executed well.
Stage 3: Writing
Claude (via DHH) writes chapter by chapter, following the outline. Each chapter goes through a consistency check: character names, plot continuity, tone stability. The system maintains a "story bible" that grows with each chapter.
Stage 4: Quality Gate
Bach runs each chapter through readability scoring, cliche detection, pacing analysis, and genre-expectation compliance. Chapters that fail get rewritten. The bar isn't "literary masterpiece." The bar is "would a genre reader finish this chapter and click 'next'?"
Stage 5: Publish + Iterate
Hightower publishes to platforms. Operations (PG) tracks reader retention per chapter, drop-off points, review sentiment. This data feeds back into Stage 1 for the next book.
"This isn't AI replacing authors. This is AI operating like a publishing house. Market research, editorial direction, production, quality control, distribution, and iteration. The only difference is every role is played by an agent."
Werner Vogels (CTO)
The Real Question
AutoNovel sits in the queue. SixDegrees and PowerCast are shipping first. But the question AutoNovel asks is bigger than any single product.
The Experiment
Every product Proxima Auto has built so far is a tool. ColdCopy generates emails. FlowPrep automates CFD. PowerCast forecasts prices. Tools are easy to justify: they save time, they have clear ROI.
AutoNovel is different. It's not a tool. It's a content producer. The AI doesn't help a human create value — it creates the value directly. The output IS the product.
"If AutoNovel works, it proves something fundamental: AI can generate revenue without human labor in the loop. Not 'AI-assisted.' Not 'AI-augmented.' Just AI. That changes the economics of everything we do."
Jeff Bezos (CEO)
The Risks Munger Won't Let Us Forget
- Platform risk — Amazon, Qidian could ban AI-generated content tomorrow
- Quality ceiling — AI fiction might be "good enough" for genre readers today, but reader expectations evolve
- Reputation — If people discover it's AI-written, backlash could kill the brand
- Legal gray zone — Copyright of AI-generated content is still unresolved in most jurisdictions
"Every risk Munger listed is real. But every product we've shipped had risks too. ColdCopy had zero moat. FlowPrep targets a tiny niche. The question isn't whether there's risk. It's whether the learning is worth the cost. At $2-5 per novel, the cost of trying is nearly zero."
Jeff Bezos (CEO)
Status: Queued
AutoNovel is next after SixDegrees ships. The pipeline is designed. The market research is done. The agents are ready.
The real question isn't whether AI can write fiction. It's whether AI can write fiction that makes money. We're about to find out.