Use AI to check novel character continuity
For a long novel, use AI as a contradiction finder over a small, dated character reference—not as the authority on canon. Keep one source of truth, record facts with their scope and evidence, and ask the model to compare a specific draft passage against only the relevant entries. Review every flag before changing the manuscript. This workflow helps you catch continuity errors while preserving intentional change, unreliable narration, and unresolved questions.
Why a character reference needs more than a list of traits
A flat list such as “Mara is cautious, left-handed, and collects postcards” mixes different kinds of information. “Left-handed” may be a stable fact; “cautious” is an interpretation that can vary by situation; and her postcard collection could change during the story. If the list does not say when and how each statement applies, an AI checker may call meaningful character development a mistake—or miss a real continuity problem.
Build the reference around claims that can be checked. For each one, capture the character, the claim, its time or scene scope, and where it came from. Add a status such as confirmed, planned, uncertain, or superseded. Distinguish what is true in the story from what a character believes, says, remembers, or tells another person. Those categories are not interchangeable.
Keep entries atomic: one claim per row. “Mara is 32, left-handed, cautious, and owns a blue bicycle” is difficult to check or update safely. Separate claims mean a revision to her age does not accidentally overwrite her handedness or an unresolved ownership detail.
A practical workflow for checking a new scene
1. Freeze the draft passage and retrieve relevant canon.
Before asking for a check, decide which manuscript version you are checking. Pull the character entries relevant to the passage: physical details used in the scene, current possessions, relationships, knowledge, location, and timeline. Include earlier facts only if the scene depends on them. Do not send a whole novel as an undifferentiated prompt and assume the model has noticed every constraint.
This caution has a research basis, but its limits matter. Liu and colleagues’ 2023 paper, [“Lost in the Middle: How Language Models Use Long Contexts”](https://arxiv.org/abs/2307.03172), reports experiments in which tested models often used relevant information less reliably when it appeared in the middle of long inputs than when it appeared near the beginning or end. That finding is not a direct test of novel continuity, nor does it establish how any particular current model will behave. It does support a modest practical choice: retrieve and restate the relevant facts close to the question instead of relying on a large context alone.
2. Ask for comparisons, not general proofreading.
Give the model the scene excerpt and the relevant canon entries. Ask it to identify only direct conflicts, quote the exact lines on both sides, and explain whether the discrepancy could instead be a deliberate change, an in-world belief, or a time difference. Require “no conflict found” when it cannot point to specific evidence. A useful prompt is:
“Compare the excerpt with the canon entries below. Return a table with: canon claim and source, excerpt wording, conflict type, confidence, and a question for the author. Flag only incompatible claims within the same time and viewpoint. Separate direct contradiction from ambiguity and possible character development. Do not rewrite the scene or invent missing facts. If evidence is insufficient, say ‘uncertain’.”
This is a proposed editorial procedure, not a guarantee that a model will follow the instructions accurately. OpenAI’s [prompt engineering guide](https://developers.openai.com/api/docs/guides/prompt-engineering) recommends clearly specifying the task and organizing the input; that documentation describes prompt guidance, not an assurance that outputs will be complete or correct. Treat the comparison as a list of leads to verify against the manuscript.
3. Triage flags before revising.
For every flag, inspect the cited scenes and classify it. A practical four-way decision is:
Do not accept an AI-generated fix without checking the surrounding scene. A suggestion that changes a fact can introduce a second contradiction—for example, changing a prop detail in one scene may break the timeline elsewhere. Once you choose a revision, update the manuscript and the reference together, then recheck affected later scenes.
Worked example: a notebook changes color
Suppose the confirmed reference says: “Mara carries an indigo pocket notebook in chapters 2 through 12.” In a chapter 13 draft, she takes a green notebook from her bag and reads an entry made the previous week. A checker should quote both passages and ask whether this is the same notebook, a replacement, or a deliberate mismatch in what another character notices. It should not declare a contradiction solely because two colors appear.
The author checks the manuscript. If chapter 12 shows Mara buying a green cover for the indigo notebook, the apparent difference has an explanation, but a reader might need one small cue in chapter 13. If there is no cover change and the older entry cannot be in a second notebook, the draft may contain a genuine continuity mistake. The author can change “green” to “indigo,” add a supported transfer scene, or decide that the character picked up a different notebook; each choice has consequences for nearby passages.
The useful test is whether the claims refer to the same object, at the same point in time, from the same viewpoint. If any part is unresolved, flag it as a question. A model can point to the two sentences, but the author decides which notebook exists in the story and updates the reference only after making that decision.
Revision control: make the canon change traceable
After resolving a flag, record the editorial decision next to the affected claim: what changed, which chapter supports it, and whether older references are superseded. Keep an earlier version or dated note when it helps explain why other scenes may still contain the previous detail. Then search the manuscript for that character and the relevant detail—such as notebook, cover, indigo, or green—and review each occurrence in context. A literal search can find wording the AI missed, though it cannot decide whether each use is narratively consistent.
Run the check again on the revised excerpt and any scenes affected by the change. Avoid asking the model to silently edit the canon sheet: first review the proposed change, then apply it yourself or through a controlled document revision. If you use a project workspace or document feature, verify its current official instructions before relying on any specific storage or retrieval behavior; the continuity method works with a plain document and does not depend on a particular platform.
Limits: what AI can and cannot establish
A model can compare text you provide, suggest places to inspect, and help format a consistency report. It cannot decide your intended canon from conflicting drafts unless you resolve the conflict. It may overlook a detail, misread a timeline, mistake subtext for fact, or produce a plausible but unsupported explanation. Long-context research is one reason to avoid treating “I included the whole book” as proof that every relevant sentence was considered; it is not evidence that all systems fail on long manuscripts.
The author remains responsible for canon, voice, and meaning. Use AI output as an audit trail: every proposed contradiction should point to passages; every accepted fix should have an explicit rationale; and unresolved items should remain visibly unresolved. For a long novel, that modest discipline is more dependable than asking a model to remember character facts implicitly across chapters.