Book Author Self-Publishing — One Person, One Book, Shipped

What's in this pack

You are one person writing one book. There is no agent, no publishing house, no production editor. You are also the cover designer, the EPUB technician, the audiobook narrator, the launch marketer, and the person who reads the KDP terms of service at midnight. Every existing "AI for authors" listicle is either (a) a vendor's affiliate page or (b) a generalist tool wall with 50 picks and no order.

This pack is the opinionated ten-tool rig across the six layers a self-publishing author actually crosses:

• Outline — book structure, chapter beats, scene cards
• Draft — AI co-writer + long-form editor for daily wordcount
• Edit — second-pass tooling for prose-level revision
• Cover — AI image generation tuned for book covers (not Instagram art)
• Format — Markdown → EPUB / MOBI / print-ready PDF via Pandoc + Calibre
• Publish — Amazon KDP for paperback + Kindle, plus a Ghost/Substack-style home for serialised long-form

Everything here is open-source or self-hostable. The two paid alternatives commonly named in this space — Scrivener and Sudowrite — are good products; this pack just covers the OSS-first path so you don't end up with five subscriptions for a book that pays $40/month.

Install in this order (outline → draft → edit → cover → format → publish)

1. Novel — Notion-Style AI Editor Component (outline + draft) — Start here. A clean WYSIWYG block editor with slash-commands and a Notion feel, plus an embedded AI completion that's good enough to ghostwrite the boring connective sentences. Use it to lay down a chapter outline and the first 500 words of each section before you go anywhere near a heavier editor.
2. Claude Code Agent: Technical Writer (AI co-draft) — The agent originally tuned for API docs, but the prose-tightening it does is the single most useful thing AI can do for a long manuscript: cut padding, enforce voice consistency, surface contradictions between chapter 2 and chapter 11. Drop it in .claude/agents/ and run it on each chapter once before human edit.
3. Zettlr — Open-Source Markdown Editor for Academic Writing (daily writing surface) — Distraction-free long-form Markdown editor with Pandoc baked in, citation manager, and a manuscript-aware sidebar (chapters, word count per file, project totals). This is where you put in the 1,500-words-a-day grind.
4. MarkText — Simple and Elegant Open-Source Markdown Editor (edit pass) — Lightweight WYSIWYG-Markdown hybrid for the second pass. Open one chapter at a time, edit prose without the file-tree noise of a manuscript app. Pair with Zettlr (Zettlr for writing, MarkText for revision) rather than picking one.
5. Stable Diffusion Web UI by AUTOMATIC1111 (cover, first draft) — Local Stable Diffusion install for cover concepting. Generate 50 covers in an hour at zero per-image cost. Use SDXL or Flux models, not the 1.5 defaults — covers need photographic detail.
6. InvokeAI — Professional Creative Engine for Stable Diffusion (cover, finishing) — Professional canvas with layer-based inpainting and ControlNet. When you've found a concept in AUTOMATIC1111 but the title text needs to sit in a specific dead-zone of the image, this is where you finish it. Treat #5 as ideation and #6 as composition.
7. Pandoc — Universal Document Format Converter (format) — The Markdown-to-everything converter. One command turns your chapter .md files into a valid EPUB, a print-ready PDF (via LaTeX), and a Word .docx for the editor you might still hire. Learn --reference-doc and --epub-cover-image early.
8. Calibre-Web-Automated — Self-Hosted eBook Library with Auto Import and Conversion (format, EPUB QA) — Self-hosted Calibre wrapper. The killer use here is EPUB QA: import the Pandoc output, view it in three reader engines, fix the table-of-contents and CSS bugs before you ever upload to KDP. Also: it auto-converts EPUB → MOBI for legacy Kindle.
9. F5-TTS — Flow Matching Text-to-Speech (audiobook narration) — Open-source, zero-cost TTS that's good enough to read non-fiction chapters and bad enough that you'll still want a human pass on dialogue-heavy fiction. The legal point matters: Audible/ACX accepts AI-narrated audiobooks but requires explicit disclosure as of 2025. Don't omit it.
10. Ghost — Professional Publishing Platform for Modern Journalism (publish + serialised long-form) — Self-hosted Ghost is the open alternative to Substack for serialising chapters as a newsletter while the manuscript is still in progress. Free tier of readership beats a paywalled launch. Substack is fine too; pick Ghost if you want to own the email list and the URL.

How they fit together

Outline (Novel #1)
   │
   ├── chapter structure + first 500 words per section
   │
Draft loop:
   Zettlr (#3)  ⟵──── daily writing
      │
      └── Technical Writer agent (#2) ── prose-tightening pass per chapter
           │
           └── MarkText (#4) ── human revision pass
                 │
                 ▼
           Final Markdown manuscript
                 │
                 ├── Pandoc (#7) ───► EPUB / PDF / DOCX
                 │       │
                 │       └── Calibre-Web-Automated (#8) ─ EPUB QA, MOBI conversion
                 │
                 ├── SD WebUI (#5) ─► concept ─► InvokeAI (#6) ─► final cover PNG
                 │
                 └── F5-TTS (#9) ──► audiobook WAV/MP3 per chapter
                 │
   Publish:
      ├── Amazon KDP upload (EPUB + cover + paperback PDF)
      ├── Audible/ACX upload (audiobook + AI-narration disclosure)
      └── Ghost (#10) ── serialise chapters as newsletter for ARC readers

The load-bearing trio is Zettlr + Pandoc + Calibre-Web-Automated: writing surface, format converter, QA viewer. Without all three, you spend launch week fighting EPUB validator errors instead of marketing.

Tradeoffs you'll hit

• AI co-writer vs human-only draft — Letting AI write paragraph one of a chapter saves time. Letting AI write the whole chapter produces second-chapter-quality-dropoff: the model loses tone consistency around the 4,000-word mark. Best practice: AI for connective tissue and outlines, human for every scene that carries emotion or argument.
• KDP vs traditional publishing — KDP gets you to readers in 72 hours with full rights retained. Traditional gets you bookstore placement and a print run you don't have to fund, but expects a 12-18 month timeline and exclusivity. For non-fiction with a defined audience, KDP almost always wins on time-to-revenue. For literary fiction looking for prestige reviews, the calculus differs.
• Paperback vs eBook-only — KDP print-on-demand makes paperback nearly free to add, and a physical book on Amazon dramatically improves the eBook's credibility. Default to both unless your topic is so niche that no one will browse the physical edition.
• Substack vs self-hosted Ghost — Substack: zero ops, instant audience cross-promotion, takes 10% of paid revenue. Ghost (#10): you own the list, no platform cut, you run the server. Pick Substack to start, migrate to Ghost when the list crosses ~1,000 paid.
• AI cover (#5/#6) vs Fiverr designer — A $35 Fiverr cover is competent and forgettable. A 6-hour AI session can be either striking or amateurish. The rule: if you can describe your book's tone in three concrete visual references, AI wins. If you can't, hire someone.

Common pitfalls

• AI chapter-two quality dropoff — The first chapter the model writes is fresh and on-voice. Chapter two onwards drifts. The fix: re-prime the agent with a one-paragraph style sheet (voice: skeptical / second person / no semicolons) at the start of every chapter session, and run #2 over the whole manuscript at the end to catch drift.
• EPUB formatting chaos — Pandoc-generated EPUBs are valid but ugly out of the box. Custom CSS via --css and a --reference-doc style template are mandatory, not optional. Always preview in Calibre-Web-Automated AND on a real Kindle before upload — the web preview lies.
• Cover violates KDP guidelines — KDP rejects covers with cropped text, AI-watermarks left in by mistake, or trim-line bleed in the wrong place. Read the KDP cover spec once, save it as a checklist, and check every export. Rejections cost 48 hours.
• No ARC (advance reader copy) program — Launching with zero reviews is the single biggest self-publishing mistake. Use Ghost (#10) to serialise the last three chapters to a closed list 30 days pre-launch and ask for honest Amazon reviews on launch day. Twenty real reviews beat any ad spend.
• Audiobook royalty math wrong — Audible/ACX pays 40% royalty on exclusive contracts, 25% on non-exclusive. AI narration is allowed but must be disclosed as of 2025, and some authors report lower discoverability for disclosed AI titles. Don't assume audiobook revenue without running the math against your actual list price and runtime.