Ghostwriter AI
A hybrid AI book-writing automation that helps novelists produce full-length books in roughly half the time, without sacrificing their personal voice or their story's internal consistency.
The problem
AI writing tools promise speed, but on a book-length project they fall apart in predictable ways: the story drifts, characters act out of character, facts contradict earlier chapters, and the structure quietly collapses. Authors end up spending as much time fighting the AI and fixing its mistakes as they would have spent writing.
The goal was to cut a book's production time from around 160 hours to 60-80 — while keeping the author's voice intact and eliminating "AI arguing time."
The core insight
Instead of asking one AI to "write the book," the system shifts human effort away from prose generation and toward outlining and quality control. Prose is cheap to generate; consistency is what's hard. So the architecture attacks consistency directly — validating structure and canon before a single expensive word of prose is written.
The approach — a two-phase AI pipeline
- Validates each scene outline before any prose is generated, so credits are never wasted on a broken outline.
- A non-AI parser enforces structure and fails fast with a clear fix message when formatting is off.
- LLM audits check scene detail and point-of-view drift, then cross-check the scene against a Series Bible for canon conflicts (characters, locations, timeline, rules).
- Outputs a clean PASS/FAIL report — only passing outlines move forward.
- A primary writer AI generates prose scene-by-scene against the verified outline.
- A second verification AI audits every scene for character consistency and drift, enforcing hard stops at point-of-view boundaries.
- A folder-based input system (voice-calibration samples, glossaries, style guides) makes the same engine work across genres without re-coding.
- The result is a logic-verified 'pre-first draft' delivered straight to Google Docs.
Scaling to a series with RAG memory
For multi-book series, the Series Bible is stored as vector embeddings in PostgreSQL + pgvector. A separate ingestion workflow chunks the bible and generates embeddings, so the writer retrieves only the handful of canon facts relevant to a scene — instead of flooding the model's context with the entire bible. That keeps cost down and accuracy up as the world grows across books.
The result
- Book production time cut roughly in half (~160h down to 60-80h).
- Drift, hallucination, and canon contradictions caught automatically before they reach the draft.
- The author's voice preserved through calibrated inputs and a dedicated verification pass.
- A reusable, genre-agnostic pipeline delivered as a done-for-you setup inside the author's own tools.
Want something like this built?
I design AI automations and agents that do real work reliably. If you have a workflow that could be smarter, let's talk.
Get in touch