I've published two AI-written novels on Amazon KDP in the last 48 hours. Not rough drafts — fully edited, cover-designed, keyword-optimized novels ready for sale. This is a complete walkthrough of how that pipeline works.
This isn't a "paste into ChatGPT" guide. It's a proper multi-agent system that handles the full production workflow from story concept to KDP submission. Here's exactly how it works.
The obvious approach — dump a prompt into an LLM and ask for a novel — doesn't work at scale. A 90,000-word novel is far beyond any context window. By chapter 15, the model has forgotten chapter 1. Characters drift. Plot threads disappear. The quality degrades noticeably.
The solution is agent specialization: instead of one AI trying to hold an entire novel in its head, you use multiple agents, each with a narrow, well-defined job.
The Writer's job is simple: write one chapter at a time. Each time it runs, it receives:
That's its entire context. No accumulated drift, no forgotten subplots. Every chapter starts fresh with everything it needs injected in. Target: 3,500–4,000 words per chapter, 25 chapters, ~90,000 words total.
The Editor reviews every 3 chapters as a batch. It checks against a detailed QC checklist:
If a batch fails QC, the Editor returns specific revision notes and the Writer rewrites the flagged sections. In our most recent project, chapter 8 failed on a continuity error — a character name used incorrectly in a flashback. Caught and fixed before it compounded.
The Marketer handles everything Amazon-facing. It receives the story bible and chapter summaries, then produces:
Keyword research matters here. For alien romance, the primary term "alien romance stories" pulls 51,024 searches/month on Amazon. That's the term the Marketer builds around — not generic phrases no one searches.
The Orchestrator coordinates everything. It maintains a project file that's the single source of truth — current chapter status, QC results, cover generation status, KDP submission checklist. Every agent reads from and writes to this file. Nothing falls through the cracks.
Covers are generated via the Replicate API using Flux, then composited with custom typography. The pipeline:
KDP Paperback tip: Get your actual page count from KDP's print calculator before finalizing the spine width. Your word processor's page count and KDP's print page count differ — sometimes significantly. Submit a draft first, get the KDP page count, recalculate the spine, then resubmit the final wrap.
Before submitting, every book goes through a final checklist:
| Item | Cost per book |
|---|---|
| Claude API (Writer + Editor, ~500k tokens) | ~$15–18 |
| Claude API (Marketer + Orchestrator) | ~$3 |
| Replicate cover generation (3–4 attempts) | ~$2 |
| KDP submission | $0 |
| Total | ~$20 |
At $4.99 eBook with 70% royalty: $3.49 per sale. Break-even at 6 sales. A mid-performing romance title in a well-chosen niche can do 30–100 organic sales per month. The pipeline can produce another book while the first one is selling.
No pipeline runs perfectly the first time. Here's what we hit:
All of these are solvable problems. The system catches them before they reach readers.
The KDP Novel Factory is this entire pipeline packaged as an OpenClaw skill. Install it, configure it with your pen name and preferences, and tell it what to publish. It handles the rest — story bible through KDP submission.
Four-agent publishing pipeline for OpenClaw. 9 genre packs included. $197 one-time.
Get the KDP Novel Factory →The two books produced with this system — The Cartographer's Compass (Fantasy Romance, 95,161 words) and The Wrong Star (Sci-Fi Romance, 86,455 words) — are both in Amazon review as of late March 2026. I'll update this post when they go live with real sales data.