Guide · 13 August 2026
How to keep a hundred-chapter series consistent
The mistakes readers write about are never the interesting ones. Nobody posts a review to say your magic system has a philosophical hole in it. They post to say the sword was on his left hip in chapter nine and his right in chapter sixty-one, and that the innkeeper you buried in book one poured them a drink in book three.
Somewhere past chapter forty you stop writing a book and start maintaining a database with a plot attached. It has no schema, no keys, and one very tired administrator. What follows is a method for keeping that job small. It works in a plain text file, and nothing in it requires buying anything.
Only five things actually break
Long-series continuity feels enormous because it gets discussed as one problem. It is five.
- Who knows what, and when. The invisible one. A character acting on information nobody gave them.
- Who is where. Position, travel time, and the two people who cannot be in the same room but are.
- What things cost. Money, distance, casting time, cooldowns, days on the road. Any number a reader can check.
- Who is dead, maimed, married or exiled. Status meant to be permanent, and therefore never re-read.
- How the rules work. Magic, politics, the System. Rules you invented under pressure in chapter thirty and contradicted under the same pressure in chapter seventy.
Physical description is not on the list. It is the one everybody worries about first and it matters least. Grey eyes in chapter four and green in chapter sixty-one is a blemish. A character who knows their brother's killer eleven chapters early is a scene that stops working.
Knowledge is the one that sinks books
Knowledge leaves no trace in the prose. You can search a manuscript for a sword; you cannot search it for the absence of a conversation. It also multiplies. Eight characters can each hold a different, partial, wrong version of the same event, and every version has to hold for as long as the secret does. So record it explicitly. Not "the truth about the massacre" but a line per character per fact, with the chapter they got it in and who told them. Overkill for three chapters. The only reason chapter eighty-one works.
Why the spreadsheet stops working around chapter forty
A spreadsheet records the present. You need the past. Three things follow from that.
It only holds the current state. You update the cell and the old value is gone. So it answers "what is her level now", which you rarely need, and not "what was her level in chapter forty", which is what you ask every time you edit chapter forty. Purpose-built tools mostly do not model it either. Plottr handles a series properly, with unlimited books in one project and a series-level view over per-book timelines, and its character fields are still free text with no state that changes over time.
Updating costs more as the world grows. At chapter ten there are a dozen things to keep current. At chapter eighty there are hundreds, and a twenty-minute chore stands between you and the end of the day. So one Tuesday you skip it, then a second.
It fails silently. A stale sheet looks exactly like a current one. No error, no warning, nothing to say the record and the manuscript diverged eleven chapters ago. You find out from a reader.
Record the change, do not reconstruct the state
This is the whole discipline. When something changes in a scene, write the change down as you write the scene. Do not maintain a picture of the world and try to keep it accurate. Keep a list of what happened and let the picture follow.
The difference is not stylistic. Reconstruction is a memory operation and memory degrades. Recording is clerical, and clerical work stays cheap forever. Forty seconds at the end of a scene beats a weekend of archaeology producing a bible you half-doubt.
An entry needs four things and no more:
- what changed, named the same way every time
- from what, to what
- which scene or chapter it happened in
- whether the reader saw it, or only you know
In practice: ch.31 — Kael learns the debt is owed to Vessine, told by Marra, on the page. Or ch.62 — Marra rides for the coast, eight days, off the page. Seconds each, and neither asks you to decide anything.
Treat the log as append-only. When you retcon something, do not edit the old entry: write a new one saying what changed and from which chapter it applies. The old entry is evidence about the chapters that shipped with it, and revisions need it.
Three records mean no record
The common setup is an outline in one tool, a bible in a second, the prose in a third. It survives book one, then fails the first time you deviate while drafting, which is every time. Now the outline is fiction about your fiction.
Pick one record, make it authoritative, and let the others go stale. Which one matters less than picking, but it should be the one you touch daily: a record you have to remember to open is one you will stop opening.
Scrivener shows the cost even at its best. The binder is still the deepest manuscript-structure tool there is, and a character in it is a text document in a folder. Each book is its own project, with no first-class way to share a character between book one and book seven short of copying the folder or syncing a bible by hand.
Give the record a shape before you need one
Free-form notes are wonderful for three months and unusable at book three, because nothing enforces that chapter sixty-one uses the same field names as chapter four. Writers keeping a bible in Obsidian hit this hardest: the flexibility is real, and so is a stat block being whatever front matter you invented that morning. Fix a small number of fields and never add a sixth on a whim.
- a canonical name, plus every alias and title the prose uses
- the handful of physical details you describe repeatedly
- what they know, as dated lines rather than a paragraph
- what they carry, and what it is worth
- status, and where they were last seen
Aliases are the field people skip and then regret. If your record says Marra and your prose says the Ashfall widow, nothing you do later connects the two. Novelcrafter's Codex handles this well, with aliases and automatic detection of mentions in prose. Copy the idea even if your bible is a folder of text files.
A model is not the record
A model that has read your book has a very good memory for impressions and a very bad one for facts. Ask ChatGPT about your protagonist and it remembers she is a reluctant healer. It does not remember that her mana regeneration hit 14 per second in chapter sixty-two, and where it does not know, it produces something plausible rather than admitting the gap. Fluent invention is the one error you cannot catch by reading, because it reads correctly.
Attaching the bible is better, and still bounded. A Claude Project holds files that persist across every conversation in it, and context windows are now large enough that a substantial bible fits. What it cannot do is change: it is a snapshot you re-export by hand, and the model cannot write a state change back.
Tools that do the attaching for you have a window too, and it is worth knowing its size. Sudowrite feeds prior narrative into generation automatically, capped at roughly 20,000 words across up to 25 linked chapters. That covers the recent past well. It does not reach the skill unlocked in chapter eight, or the debt owed since chapter thirty-one.
So the model reads the record and drafts against it. The record is written by you, or by a system you can audit. Never by recollection.
Where to keep it, if you are starting today
You do not need to buy anything to do this properly.
- A folder of Markdown files costs nothing and cannot be taken away from you. You are the schema designer and the support desk.
- Kanka is free with unlimited entries across eighteen entity types, so a long series never hits a content wall. The vocabulary is tabletop-first, so you translate as you go.
- Novelcrafter's Scribe tier, if you would rather not build it yourself, is $4 a month for unlimited books, series and universes, with no AI attached.
Ask any tool two questions before committing years to it. Can it say what was true as of chapter forty, not just what is true now. And can you get everything out. LegendKeeper is the clearest case of the second: full HTML and JSON export at any time, with a stated commitment never to paywall it.
Some of this a machine can check, most of it cannot
Once state changes are data rather than prose, a whole class of error becomes checkable by comparison alone. A character in a scene after their death. An item in two places at once. A level that went down. A stat over its maximum. A date running backwards. A character using a name nobody told them. None of that needs judgement or a model. They are lookups, they should be automated, and they are the errors that survive a careful read.
Everything else is still yours. Whether the reveal lands, whether she would really say that, whether the rules you have been so consistent about are any good. Continuity is the floor, not the ceiling. It earns the effort only because falling through it costs you the reader who was going to buy book four.
The habit, in one paragraph
Keep one record and let the others rot. Write changes down when they happen, in the same shape every time, with the chapter number attached. Track knowledge as carefully as possessions. Never edit history, append to it. Automate the checks that are lookups. Do that from chapter one and chapter one hundred is an ordinary day's work.
Questions people ask about this
How long should this take per scene?
Under a minute once the habit sets. You are copying facts you have just written while they are still in your head, not deciding anything. If it takes five minutes you are summarising the scene as well, which is a job you do not need.
Do I need to record everything?
No, and trying to is the commonest reason people give up. Record only what a later chapter could contradict: knowledge, position, possession, status and rules. A character's opinion about the weather is not state. A character learning who killed their brother is.
What if I have already written forty chapters with no record?
One read-through with a log open, fixing nothing on the way. Note the state change and the chapter, keep moving, and collect contradictions in a separate list for afterwards. Budget roughly a day per hundred thousand words. It is dull and it is finite.
Can an AI just do this for me?
It can extract state changes from a scene you paste in, which saves real time, and you should check what it extracted. What it cannot be is the record itself. A model remembers impressions and fills gaps with invention rather than admitting a gap.
Does this work if I write without an outline?
Better, in fact. Recording state as it happens assumes nothing about what comes next, so it suits a discovery writer. You are keeping a ledger of what is true, not a plan of what will be.
If you would rather not keep the log by hand
New Quest does the recording part for you. Paste a drafted scene and it extracts what changed. It keeps the history, so chapter forty stays answerable, runs the lookup checks above without calling a model, and serves the record over MCP to whichever LLM you draft with. One book is free, which is enough to find out whether the habit suits you. If a folder of Markdown files is already working, keep it.
No card. Export everything whenever you want.