Somewhere around chapter fourteen you will forget whether your heroine's mother is alive. Your readers won't. They'll email you. A story bible is the cheap insurance against that email, and a story bible AI can read is the difference between a model that stays consistent and one that improvises your backstory every time you ask for a scene.
This guide covers what goes in a bible, how to keep it from rotting, and how to hand it to an AI so it gets used. The manual method comes first, because it works anywhere and you should understand it before you automate it.
What a story bible is (and isn't)
It's a reference document about your book, written for whoever is about to write the next page. That reader used to be you. Now it's also a model that has never seen your manuscript.
It isn't an outline, and it isn't a wiki of everything you ever imagined. A bible full of trivia is as useless as an empty one, because nobody reads it, human or machine. Keep it to facts that the next scene might need.
What goes in it
Five sections cover almost every novel.
Characters
For each person who matters, a few lines, not a dossier:
- Name, age, job, and how they look on the page (the details you've actually shown).
- What they want, and what they're afraid of.
- How they talk. A verbal habit, a sentence length, a word they'd never use.
- What they know, and what they don't. This one prevents the most errors. Who's in on the secret?
- Relationships that change the scene: ally, rival, lover, owes money.
Places
Name, what it looks like in two or three concrete details, and anything the plot depends on. A door that sticks. A road that floods. Distances between locations if anyone travels.
Threads
Open promises to the reader. The unexplained letter, the debt, the stranger at the funeral. List each thread with its status: planted, developing, paid off. This is the section that stops you dropping a plotline at chapter twelve and forgetting it existed.
Rules
Whatever is true in your world that a reader would call a cheat if you broke it. Magic costs. Technology limits. Legal systems. Timeline constraints, like "the war ends the winter before the story starts."
Voice and style
Point of view, tense, how much interiority, what your prose does and doesn't do. If you ban yourself from adverbs or from opening on weather, write it down. A model can follow a style note far better than it can guess one.
Keeping it current
The usual failure isn't a bad bible. It's a stale one. Three habits help.
Update after every chapter, not every month. Five minutes while the facts are fresh. Add what the reader now knows, change statuses, note new characters.
Write down what the page says, not what you planned. Drafts drift from outlines. If your villain became sympathetic in chapter nine, the bible should say so.
Record dates and ages. Timelines are where continuity quietly breaks.
Also keep a running summary, a paragraph per chapter covering what happened and what changed. The bible holds the lasting facts, and the summary holds the order of events. You need both.
How to give it to an AI
Here's the manual method. It's tedious, but it works with any model.
- Don't paste the whole thing every time. Models have limited context (200K to 1M tokens on Claude, depending on the model), and a bloated prompt buries the part that matters in the middle, which is where models read worst. Pick the entries relevant to the scene. NovelAI's Lorebook and AI Dungeon's Story Cards automate this with trigger words; by hand, you are the trigger.
- Lead with a short summary of the book so far, then the relevant bible entries, then the last scene or two, then what the new scene must do. That order isn't taste: Anthropic's prompting guide puts long material at the top and the request at the end, which improved results by up to 30 percent in its tests.
- State the beat plainly. "Mira finds out about the debt, and does not confront him yet" beats "write the next scene."
- Tell it what not to invent. "Use only characters listed here" cuts a surprising amount of random new cousins.
- Check the output against the bible, then fold anything new back into the bible.
That's also the answer to why models drift when you skip this. Without context they fill gaps with the most likely guess, which is rarely your guess. Our guide to why AI forgets your story covers the mechanics in more depth.
Letting the tool do the bookkeeping
The manual routine has one weakness: you. It's easy to skip the update on a tired night, and one skipped chapter is where inconsistencies start.
This is what KudoWrite automates. You sketch a beat, the AI writes it, you Rewrite, Expand or Split until it's right, and then you Commit it into the chapter. Every commit updates a story bible of characters, places, items and events, plus a running summary of the book. The bible stays readable and editable, so what a character knew in chapter 6 is a click away instead of a scroll.
It doesn't remove the need to think like an editor. You still read the bible, and you still correct it when it records something wrong. Treat it as a draft of the document you'd otherwise keep by hand. One limit to know: it's a browser app with an English interface. Existing notes don't have to be retyped, though: Sudowrite's Story Bible CSVs, a Novelcrafter codex, a NovelAI lorebook or AI Dungeon story cards import as bible entries, and the manuscript itself comes in from .docx, .epub, Scrivener and most other formats.
Three examples of an entry that works
Vague entries produce vague scenes, so here is the level of detail to aim for.
Character. "Mira Okafor, 34, ferry mechanic. Short sentences, swears in her grandmother's language when scared. Wants to buy back the family boat. Afraid of being the one who stayed. Does NOT know her brother sold the boat's papers. Owes Dell money."
Place. "The Low Landing: a ferry dock below the cliff road. Planks stay wet year-round, the only light is a sodium lamp that buzzes, the tide floods the lower step at dusk. Eleven minutes on foot from Mira's flat."
Thread. "Unsigned letter found in the engine room (ch. 4). Status: planted. Mira has not shown it to anyone. Pays off in ch. 22 when the handwriting is matched."
Notice what they share: concrete nouns, a status, and a note on who knows what. None of them is a paragraph of lore. A model can use a three-line entry. It will skim past a three-page one.
Common mistakes
- Writing it once and never touching it again. A bible from chapter one is fiction about a book you no longer have.
- Putting spoilers where the AI will repeat them. If a twist lands in chapter 30, mark which facts are secret and from whom, or the model may cheerfully reveal it in chapter 5.
- Recording adjectives instead of facts. "Brave" tells a scene nothing. "Walks into the burning shop for the cat" does.
- Letting the model rewrite your canon unchecked. If an AI proposes a new detail and you like it, it becomes canon only when you decide it does.
A small starter template
If you're building from scratch today, copy this shape into any document:
- Premise in two sentences.
- Summary so far, one paragraph per chapter.
- Characters, five lines each at most.
- Places, three details each.
- Threads, with a status.
- Rules, as bullets.
- Voice, a short paragraph.
Fill in what you know, leave the rest blank, and stop. A short, accurate bible beats a long, wrong one every time.
Where it pays off
The payoff shows up late. Chapters one through five barely need it. By chapter thirty, with a cast of twenty and three timelines, it's the only thing standing between you and a manuscript that contradicts itself. If you're drafting a whole book with a model, read how to write a novel with AI next for the full workflow around it.
Start with the thread list. It's the section writers regret skipping most.