Correcting an AI Narrator Means Naming the Error

Correcting an AI narrator starts before you touch any button: name the error. A wrong entry in the story’s record, a scene that misread a record which was already right, and a turn you simply do not want are three different failures, and each answers to a different repair. Reach for the wrong one and the detail comes back two scenes later, wearing the same face.

• A record error survives a retell, because the retold scene is handed the same wrong fact.
• A prose error dies on a retell, because the record it should have used is already correct.
• A correction phrased as a statement of what is true works better than one phrased as a complaint about the mistake.
• SillyTavern’s documentation describes lorebook entries as prompt injections that appear only when their keywords are matched, so an untriggered correction never reaches the model at all.
• In Fabledrift, an interactive fiction app for Android, a reader’s note wins over the story record when the two disagree.

Three kinds of wrong an AI narrator can be

Every mistake you notice on the page is one of three things, and they are not repaired the same way. The first is a record error: something false was written into the story’s memory, and the narrator is faithfully reproducing it. The second is a prose error: the memory is correct and this particular scene ignored or garbled it. The third is not an error at all — the story went somewhere you did not want, which is a question of authorship rather than accuracy.

Symptom on the page	What is actually wrong	The repair that works	The repair that wastes a scene
The same wrong name or fact in scene after scene	The record holds it	Correct the entry, then step back	Retelling the newest scene
One scene contradicts what came before, then it passes	The scene misread a correct record	Retell that scene	Editing the memory record
A character acts out of character once	Usually the scene	Retell, or answer in the next move	Rewriting the character entry
The plot turned somewhere you dislike	Nothing	Step back and write a different move	Calling it a bug and retelling

The fourth column is the expensive one. Most of the frustration readers report with AI stories is not that the repair failed; it is that they applied a scene-level fix to a record-level problem, four times, and concluded the app was broken.

How do you tell a record error from a prose error?

You read the record. Any app serious about continuity keeps a short, human-readable account of what the story believes — names, places, promises, open threads — and lets you open it. Fabledrift calls it the story journal; other apps call it World Info, lorebooks, or story cards. If the wrong detail appears in that record, you have a record error. If the record says the right thing and the page said the wrong thing, you have a prose error. The test takes about twenty seconds and it decides everything after it.

One caution on reading the record: it shows what the story has written down, not everything the model saw. Recent scenes travel with the record, so a detail can be absent from the journal and still present in the prompt because it happened four paragraphs ago. That is why record errors tend to be the persistent ones and prose errors the one-off ones.

When the error is neither

A repeated beat is its own category. If the narrator keeps returning you to the same doorway or restating the same revelation, no single fact is wrong and no correction will land, because the failure is structural rather than factual. That is a loop, and loops break differently — usually by giving the story a new element rather than by fixing an old one.

Correcting an AI narrator works best as a plain statement

Write the correction as the fact you want to be true, not as an account of what went wrong. “Tam is Mira’s younger brother; he is fifteen” is a sentence a narrator can use. “You keep getting Mira’s brother wrong, please stop saying he’s older” contains a grievance, a negation and, buried inside it, the wrong version repeated once more. Anthropic’s prompting guidance makes the same point about models generally: tell the model what to do instead of what not to do, and give the desired behavior rather than the prohibited one.

Three habits make a correction more likely to hold:

• One fact per note. A paragraph containing four corrections is harder to match and harder to keep true than four short entries.
• Name the subject explicitly. “He is fifteen” resolves against whoever the model thinks “he” is; “Tam is fifteen” does not need resolving.
• State it in the present. “Mira carries her mother’s knife” ages better than “Mira picked up the knife in the mill,” which is an event and will read as one more thing that happened.

The wording matters more than it should, and that is a real cost of the format rather than a trick. You are writing a line that will be read by a narrator, repeatedly, out of context, for the rest of the story. Short, present-tense, one subject.

Why does a correction stop working a few scenes later?

Usually because it never reached the model. A memory record is not a document the narrator reads front to back; entries are selected and injected. SillyTavern’s World Info documentation is explicit that its entries are prompt injections that behave “like a dynamic dictionary that only inserts relevant information from World Info entries when keywords associated with the entries are present in the message text,” and that entries with larger order numbers are inserted closer to the end of the context, where they carry more weight.

Two consequences follow for readers. A correction keyed to a word that stops appearing goes quiet exactly when the story moves on — your note about Tam sleeps through every scene that calls him “my brother.” And when two entries disagree, something has to decide which one the narrator believes, which is a rule the app sets rather than a matter of the model’s judgment.

That precedence rule is worth finding out before you need it. Our own is blunt on purpose: a reader’s note wins over the story record, because the reader is the authority on their own character. It is the same reason we treat the record as something you can read and edit rather than an internal structure kept out of sight.

Can you correct the narrator inside the story?

Often, yes, and it is the lightest repair available. If the error is recent and small, write the true version as your next move — say the brother’s name in dialogue, or reach for the knife that should be at your belt. An attempted action still happens on the page, so the correction enters the story as an event rather than as an instruction, and nothing about the reading breaks. Some apps, ours included, also let you ask the narrator a question without advancing the story, which surfaces what it currently believes without costing you a scene.

The limits are real. An in-story correction cannot undo a fact that has already been written into the record, and it cannot fix something that contradicts several scenes at once without the page turning into an argument. When the error is older than a scene or two, step outside, fix the record, and use the back button — which is four different buttons wearing one name, so it is worth knowing which one you are pressing. For us, stepping back to your own earlier scene is free, always; retelling is free on the Apprentice and once a day on the Chronicler and the Oracle.

The next wrong detail, in order

Read the record. If the wrong fact is in it, correct the entry as a short present-tense statement and step back past the scene that first used it. If the record is right, retell the scene and move on. If nothing is false and you simply wanted the story to go elsewhere, step back and write a different move, because that is authorship and not repair. Correcting an AI narrator is mostly diagnosis; the repair itself is the short part. The one move to avoid is the reflexive retell before the diagnosis, which spends a scene to learn nothing.