You paste four chapters into a chatbot and type "continue." What comes back is smooth, grammatical and wrong. The narrator sounds cheerful now. A character you killed off is humming in the kitchen. The scene ends with a tidy little moral your book would never allow.
That isn't bad luck. It's what happens when you hand a model a pile of text and no instructions about what matters in it.
Why "continue my story" goes wrong
A language model reads what's in front of it and predicts what plausibly comes next. Nothing else. It has no memory of your book beyond what you put in the prompt. So three things tend to break.
The voice drifts
Models have a house style: even rhythm, safe adjectives, a tendency to explain feelings and then wrap each scene in a reflective last line. That pull is measurable. A controlled study by Padmakumar and He (2023) found that people co-writing with an instruction-tuned model produced essays more alike each other, mostly because the model's contributions were less varied than the humans'. Your pages pull against that style, but only as hard as the sample you provided. Paste a long stretch and the model drifts back toward its default within a few paragraphs.
The facts slip
If your draft is longer than what fits in the model's working window, the early chapters get cut or blurred: Claude's help centre lists windows from 200K to 1M tokens depending on the model, and summarises earlier messages when a chat nears the limit. Even when everything technically fits, details buried in the middle get less attention than the end; the Lost in the Middle paper measured exactly that drop, and Google's own long-context docs say that with several facts to find at once the model "does not perform with the same accuracy" as the single-needle demos. That's how an eye color changes or a debt is forgotten. We cover the mechanics in why AI forgets your story.
It has no destination
"Continue" gives the model no goal, so it invents one. Usually that means resolving tension early, adding a conversation nobody needed, or introducing a convenient stranger. You asked for the next scene and got the model's idea of a story.
The fix: give it the right four things
Stop pasting the whole manuscript. Give the model a small, deliberate package instead.
- A summary of the story so far. Half a page to a page. What has happened, who wants what, which threads are still open.
- Only the bible entries that matter for this scene. The two characters who are in the room, the place, the rule of the world that's about to be tested.
- The last scene, verbatim. This is where the voice lives. Give it the final page or two, not the whole chapter.
- A beat to hit. One to three sentences on what must happen next, and where the scene should stop.
The order matters: Anthropic's prompting guide says to put long material at the top and the request at the end, which improved results by up to 30 percent in its tests. Summary and bible first, beat last.
Write the summary like a briefing
A summary that says "Elena and Marcus fall out and she leaves town" is useless, because it states the event and drops the situation. A usable one carries the state of play:
- Who knows what. Does Marcus know about the letter? Does the reader?
- What each major character currently wants.
- What's physically true right now: injuries, locations, the time of day, the season.
- The promises you've made to the reader that aren't paid off yet.
Keep it in plain past tense, update it after every chapter, and don't polish it. It's a working document, not prose. If you want a fuller method, how to build a story bible covers the entries and the summary together.
Pick the last scene on purpose
The pages you paste are the voice reference, so choose them like one. A rushed transitional scene teaches a rushed voice. If the last page is weak, paste a stronger page from earlier as a style sample and mark it clearly: "This is the voice to match." Then paste the actual last page so it knows where you are.
Cut anything that's an obvious AI artifact from the sample first. If an earlier draft had a limp closing line, delete it. Whatever's in the sample is what gets imitated.
The beat is the most important line you write
Compare two instructions:
- "Continue the scene."
- "Marcus finally reads the letter, says nothing, and goes out to chop wood. End when the axe hits the log. Don't have him forgive her."
The second one fixes the event, the ending point and the one thing the model would otherwise get wrong. Keep the "don't" lines. They're often the cheapest way to prevent the exact thing you're afraid of.
Ask for a single scene, not a chapter. Short runs are easier to judge, and a wrong turn costs you a minute instead of an afternoon. The same rule is how you write a short story with AI in one evening: six beats, one at a time.
Check the draft the way you'd check a human's
Once the continuation comes back, read it for four things before you read it for prose:
- Facts. Anything contradicting the summary or bible?
- Knowledge. Does any character act on information they shouldn't have?
- Ending. Did it stop where you asked, or did it keep going and wrap things up?
- Voice. Read one paragraph out loud next to your own. Does it sound like the same person?
If something's off, don't argue with the model across five messages. Fix the package (a missing fact, a vaguer beat than you thought) and run it again.
What about continuing mid-scene?
Sometimes you don't want a new scene, you want the model to pick up a half-written one. The same rules hold, with one change: paste the scene so far, stop on a complete sentence, and say in the beat how the scene should end. If you cut off mid-sentence, models often finish that sentence cleverly and then keep going in a direction you didn't choose. A clean stop point plus a clear ending is a much stronger signal than the unfinished text alone.
A quick checklist before you hit generate
- Is my summary current as of the last page?
- Did I include only the entries this scene needs?
- Is the pasted scene one I'd be happy to have imitated?
- Does the beat say what happens, where it ends and what must not happen?
- Am I asking for one scene, not three chapters?
Then start small: take the scene you're stuck on, write the one-paragraph beat, and compare the result with a bare "continue."
Do it by hand, or let a tool do the packaging
Everything above works in any chatbot, and in Docs, though Google's Help me write only sees the document it's open in. You keep a summary document, a bible document and the last scene, and you assemble the prompt each time. It's free apart from your time. The cost is the bookkeeping: after each scene you update the summary, add new characters to the bible, and remember which entries to include next time. It's the step people skip around chapter ten, and that's when the drift starts again.
KudoWrite builds the four-part package itself, for every beat. You write part four, the beat, as a line, a rough idea or a whole passage. Before it writes, the engine plans a brief from the other three: the story so far, the bible entries this passage needs, and the beats around it, plus the chapter's current moment. The chapter-ten step happens as you write: the bible and the running summary keep up with the paragraphs you write. And when a draft drifts anyway, you don't argue across five messages; a one-line steer edits the beat in place and outranks everything else in the brief.
There's one free credit to try it once you confirm your email, then one-off packs from $3.50 for 100 credits, no subscription. Already four chapters in? My Books → Import reads .docx, .txt, .md and more and finds the chapters and scene breaks for free, so you can continue from your last line in KudoWrite.