Most writers come to a story bible late. The draft exists, sometimes three drafts, and the bible is a folder of notes that disagree with each other. Starting from a blank page at that point means a week of copying names out of your own book.
Lokee can read the draft and propose a story bible from your manuscript: characters, places, factions, relationships, changes over time, and who was where on which day. The word that matters is propose. Nothing lands in the vault until you have reviewed it. I did this with my own novel, Medalion, and the review taught me things about the book I had forgotten, which is the point.
This guide walks the whole route: getting the manuscript in, choosing what to extract, reviewing, and then the tidying that turns a raw extraction into something you trust. The running example is the demo story The Ashen Circle, whose three chapters are already in the demo's Writer.
Two ways to get a story bible from your manuscript
There are two doors, and they do different jobs.
The Writer imports your manuscript as manuscript. On an empty Writer, Import a .docx manuscript; later, Import manuscript. Lokee reads the document's headings and shows Review proposed structure: a tree of containers (books, parts, acts) and leaves (chapters). Select rows and Merge selected when a heading split a chapter in two, use Flatten to remove a grouping and promote its chapters one level, then Commit the chapters. Your book is now in the Writer, chapter by chapter, and every later extraction can point back at the exact text.
The Import page extracts from a document. In the vault sidebar, Import takes a .txt, .docx, .md, .epub or .pdf file up to 50 MB and extracts entities and relationships from it without making it part of your manuscript. It suits series notes, an old wiki export, or a previous book whose text you do not want in this vault's Writer. If those notes are the problem you are solving, how to organize worldbuilding notes covers what deserves an entity and what stays a note.
For a novel in progress: import the manuscript into the Writer first, then extract from the Writer chapter by chapter or with Extract whole book. You get the bible and the links back to the text in one pass.
Who pays for the AI?
Extraction is real AI work on long text, so the payer rules matter before you start:
- Cloud vaults run extraction on Lokee's side and pay in Lokee credits, which needs a paid plan (/pricing). The form shows the estimated cost and your balance before anything runs; credits are deducted when extraction starts.
- Local vaults in the desktop app can use OpenAI or Claude with your own key (saved in your system's credential store), Local Model if you have one configured (it runs entirely on your computer), or ✦ Lokee AI credits if your plan includes them. With your own key, a consent screen shows which models will see the text and the estimated cost first.
There is also a route with no Lokee AI at all. The book-to-vault guide shows how to have your own coding agent (Claude Code, Codex or similar) read the book and write a Lokee vault, which you then open in the desktop app. That is how I turned Medalion into a vault, with Claude Code; the prompt kit has the prompts I used.
Choose what to extract, and keep the first pass narrow
Whichever door you use, you choose what the AI looks for.
Extraction stages. Entity extraction is always included. Traits & attributes, Relationships and Quality review can be switched on or off, each with its own model in a local vault. Quality review is a second pass that proposes corrections to the first; for a whole book it is worth the extra cost.
Two further stages work with time:
- Changes over time ("proposes what changed and when, for your review"). A promotion, a death, an alliance that ends. Each proposal can be placed at a world point, linked to the chapter where the reader learns it, or left unresolved.
- Events, places and dates ("where characters are, what happens and when"). It is slower and off unless ticked. In the Writer the same two options appear as Find changes over time and Find events, places and dates (slower) next to the extraction button.
Focus. On the Import page, Focus preset (optional) narrows a run to Characters, Locations, Relationships, Timeline, Family tree, Objects & artifacts, a Custom entity group, or A specific chapter, and Extraction focus (optional) takes a sentence in your own words, such as "only extract characters and locations, ignore minor events". A focused run is cheaper and far easier to review. The Writer's extraction buttons have no focus picker; there, the way to keep a pass small is to extract one chapter at a time. For a first pass on a long book, start with characters and places, review, and only then run relationships.
Why "dates only from the text" matters
The events, places and dates stage deserves a closer look, because this is where AI tools usually go wrong. Lokee only takes a date from the text: a date as written, or a phrase like "the next day" counted from a date the text states. When it works a date out, the proposal shows it ("Date: … — worked out as …") so you can check the arithmetic. When the text only says "that winter", you get a time hint to resolve yourself, never an invented date.
Presences come with certainty, too. If a witness reports seeing someone, the proposal is marked Reported, not confirmed, or A rumour, or Uncertain, with the source named where the text gives one. That distinction is exactly what a mystery needs; see Keeping a mystery fair. New places and events are only created when you accept them.
Review: where the story bible is actually made
An extraction from the Import page opens the review screen. Items arrive in four tabs:
- Pre-approved: candidates the AI is at least 85% confident about. Accept all pre-approved is safe for most of them, but skim the names first.
- Pending: everything below that threshold. This is where your judgement is needed.
- Done: what you have decided.
- All.
Separate sections hold Entities, Relationships and, if Quality review ran, Corrections; a Changes over time tab holds the time proposals. The entity list can be grouped By status, By type, as a Timeline or as a Family tree, which is often the fastest way to spot a wrong parent link.
For each candidate you can Accept, Discard, change the Entity type, or use Reassign to existing entity…. Reassigning is the step people skip and regret. If the AI proposes "the Magistrate" as a new character and you already have Magistrate Ophelia Crane, reassign it to her: her card gains what the text said, and you avoid a duplicate. Types the AI proposes that your vault lacks are marked new type; accepting them creates the type at commit.
When you are done, the button says exactly what will happen: Commit 34 entities + 51 relationships, for example. If anything is still undecided, Lokee says so and records it as skipped. Skipped items are not lost: Import history lists every session with your decisions, and a skipped entity or relationship can be restored from there. Rerun with a different focus → in the same place runs the document again with a narrower brief.
Extraction from the Writer reviews in a side panel, Extraction candidates, with the same idea in a smaller space: entities, relationships and changes tabs, Select high-confidence, then Apply.
Before a commit, Lokee saves a restore point (Before extraction commit), so a commit you regret is one roll-back away.
The Ashen Circle, end to end
Here is the route on the demo's three chapters, so you can see what the review looks like in practice.
The run. In the Writer, with chapter one open, Extract this chapter with Find events, places and dates (slower) ticked. One chapter is the Writer's version of a narrow pass. In a cloud vault the button shows the credit cost first; in a local vault, your key or local model.
What arrives. Rowan Ashe, Mira Voss, Old Bram and Sera Nightingale come in pre-approved. So does "Kestrel", as a separate pending character, because chapter one uses the alias before anyone says her real name. "The Watch" arrives as a faction with new type beside it, and "the East Gate" as a place. The events stage proposes two presences: "Rowan — at The Gilded Lantern" and "Sera — at The Gilded Lantern, Reported, not confirmed", because Bram says he saw her and the narration never confirms it.
The decisions. Accept the four pre-approved. Open Kestrel: Reassign to existing entity…, Sera Nightingale. Sera's card gains the sentences about Kestrel, and nothing is duplicated. Keep "The Watch" but change its Entity type to the Faction type the vault already has, so the new type is never created; later, add "The Watch" as an alias of The Crown Watch. Accept the East Gate as a place, then give it a Location of Duskhollow so maps can roll it up. Accept Rowan's presence. Accept Sera's too, with its certainty intact: the reported sighting is a fact about the text, and the first thing a reader of the mystery would check.
Commit. The button reads Commit 5 entities + 3 relationships. A restore point is saved first. Open Sera's card: the AI Extracted — verify this entry before publishing banner is on it. Read what the text contributed, fix the one line that attributed Bram's words to her, Dismiss. Repeat for the cast.
Chapters two and three go the same way, and each one is faster, because the vault now knows who Kestrel is.
After the commit: banners, duplicates, aliases and types
The banner is your to-do list
Every entity the AI created carries a banner: AI Extracted — verify this entry before publishing. It is deliberately a nag. Read the card, fix what is wrong, then Dismiss it. Treat the banner as the to-do list of the import; when no main character still carries it, the import is finished.
Duplicates are inevitable in a long book: "Sera", "Sera Nightingale" and "Kestrel" may all arrive as separate people. Reassigning during review catches most. For the rest, local vaults have Merge duplicates, opened with Find duplicates → on the entity list. Find duplicates scans the vault with the AI provider you choose (your key or a local model) and proposes pairs; for each, Keep A, Keep B or Skip, then Apply the merges. A restore point is saved first. Merge duplicates is not available in cloud vaults; there, delete the extra entity after moving anything you need to the one you keep, and add the other name as an alias so future extractions match it.
Aliases matter beyond tidiness: they drive search, @-mentions in the Writer, passive mention suggestions and AI matching. Adding "Kestrel" as Sera's alias is what stops the next extraction from proposing Kestrel again.
Tidy the types your book actually uses
An extraction shows you what your book contains, which is rarely the schema you would have designed. Open Settings, the Structure tab, and the Entity types section, and look at what arrived:
- Fold together types that mean the same thing. If "Location" and "Place" both arrived from two runs, keep one: on the entity list, select the entities of the other type, use Change type, then delete the empty type.
- Add the fields your story needs, then use Backfill missing fields so existing entities get them too.
- Give moving things a Place field and, if you use a text field for whereabouts, tick Says where someone is on it. That is what lets maps place people and lets the consistency check catch someone in two places at once. Following characters across the map explains why.
- Mark relationship types that should pull groups together in the graph as cluster relationships, and set map and family roles where they apply, under Relationship types in the same tab.
Setting up a story bible that scales goes through these decisions in depth. Doing them right after the first import is far cheaper than after the third.
Run the consistency check on the fresh import
Now ask the vault what it thinks of itself. Open Consistency Check in the sidebar and Run deterministic check (free). It covers the whole vault and costs nothing. On a fresh import, expect findings: two values for the same field at the same point, someone in two places at once, a relationship that ends before it begins. Some are extraction mistakes; some are mistakes in the book, which is the point of the exercise.
The check never fixes anything. It flags, and you decide. Running a continuity check before every draft explains each finding and how to resolve it. From here on, the habit that keeps the bible honest is in Recording change instead of overwriting it: what the extraction found as "changes over time" is the same thing you will now record by hand as you write on, and Three kinds of "when" explains the lanes those changes land in.
Restore points: the safety net while tidying
Imports, merges and bulk edits are exactly the operations you might want to undo, and Lokee saves a restore point automatically before each of them. Restore Points in the sidebar lists them by reason ("Before extraction commit", "Before entity merge", "Before bulk edit") alongside any you saved yourself with Save restore point now. Roll back to this shows what will be removed, restored or reverted before anything happens.
A good habit for a big import: save a restore point by hand before you start tidying. If an afternoon of merging goes wrong, you are one click from the clean import.
Common mistakes, and what Lokee doesn't do here
Common mistakes
Extracting the whole book with every stage on, first time. The review becomes a wall of pending rows and you start accepting carelessly. Run characters and places first, review, then relationships, then changes over time.
Accepting a new type for every synonym. "The Watch", "Guard" and "Crown Watch" each proposed as a new faction type means three types to fold later. Change the Entity type in the review to one you already have.
Discarding instead of reassigning. Discarding "the Magistrate" throws away what the text said about her. Reassign it to Ophelia and the card keeps the sentences.
Skipping the banners. An entity still marked AI Extracted has not been read by you. The banner is the only list of unverified entities there is; there is no separate "unverified" view.
Importing a series into one vault per book. Book two's extraction then cannot match book one's cast, and Kestrel comes back as a stranger. One vault per series, extracted in order.
Expecting the Capture page to do this. Capture extracts from a short voice or text note; it has no follow-up step and is for new material, not a manuscript.
What Lokee doesn't do here
It does not judge your prose, and it does not decide what matters: a walk-on with a name gets proposed like a protagonist, and you are the one who discards him. It does not invent dates, which also means a book with no dates in it yields no dated events. In cloud vaults it does not merge duplicates for you; that tool is local-vault only. And it does not keep a list of entities you have not verified beyond the banners on the cards themselves.
In five minutes
- Import the manuscript into the Writer and commit the chapter structure.
- Extract this chapter on chapter one; review; reassign; add aliases.
- Extract whole book with Relationships on; review in the same way.
- Tidy types under Settings, Structure; add a Place field; Backfill missing fields.
- Run deterministic check (free) and read the findings against the book.
- Save a restore point before the next big pass.
For a finished series, keep one vault per series and repeat the loop book by book, so each later extraction matches against a clean bible. The import help section has the short reference for the same steps.