Prompting AI to continue a story: what holds up after twenty rounds

“Continue this” is the most common fiction prompt in existence and the weakest. The model is missing three decisions — how long, at what pace, and whose style is the reference — so it fills them with the statistical average of its training data: medium-length sentences, even paragraphs, a register best described as generically competent. One round of that is tolerable. The problem is what happens over many rounds, because continuation is never one round.

Why bare prompts decay: the 20-round evidence

Continuation has a feedback loop built in: each round’s output goes back into the context for the next round. Run that twenty times, as we did across six models on the same 8.9M-character novel, and the context gradually becomes the model’s own prose instead of the author’s — so the model increasingly imitates itself. The numbers are blunt: the original’s sentence-length variation coefficient was 0.84; all six models drifted into a 0.46–0.59 band. Left unmanaged, the loop also produces harder failures — restart loops, verbatim freezes, plot rewinds — and the freeze variant has been documented since the GPT-2 era: Holtzman et al.’s degeneration study showed repetition is self-reinforcing, with each repeat raising the probability of the next.

The practical conclusion from those runs is that a working continuation prompt is not one clever sentence. It is a small system with two layers, and most of the failures we see come from stuffing layer-two problems into layer one.

The two layers, and what goes in each

The instruction layer says how to write: output length, pacing rules, point-of-view and tense locks, and a banned-phrase list. The material layer says what is true: the recent story text, any setting or character sheets, and — critically — labels for which passages are original prose and which are earlier AI continuations. Instructions govern diction. They do not govern world state or canon; twenty rounds into our benchmark, models given style rules but weak material anchoring had collectively wandered from a blood-cloud ritual array into sunlit forests. No adjective in the prompt fixes that. Structure the layers separately — the same separation OpenAI’s prompt engineering guide recommends as instructions versus context, with clear delimiters between them.

One material-layer directive deserves its own paragraph, because it produced the largest single swing we have ever measured. Labeling earlier AI output “for plot continuity reference” caused one model to treat those passages as non-canon and restart from the original text’s ending — 20 rounds out of 20. Rewriting the label as two explicit statements (these passages are canonical events that already happened; continue from the very last paragraph; anchor style to the original prose only) brought regressions to zero. If you build your own context by hand, include that canon declaration. It is one sentence of insurance against the most common long-run failure there is.

Four copy-ready templates

These four are written in English for this post, following the same principles as the official packs we ship — three-part structure: output rules, craft rules for the genre, banned list. Paste one after your story material, adjust the numbers to taste, and extend the banned list with whatever tics your model keeps producing. That last habit matters more than any starting template: a banned list is a living document.

1. Fast-paced action

When to use: progression fantasy, thrillers, anything where the failure mode is the AI padding three paragraphs of scenery while the protagonist “gathers his thoughts”.

2. Interior / emotional

When to use: romance arcs and slow-burn relationship lines, where one unguided continuation will happily jump from first eye contact to a confession.

3. Suspense

When to use: mysteries, horror, anything where the AI’s instinct to be helpful makes it dump the truth two paragraphs in — or, worse, announce that danger approaches.

4. Literary restraint

When to use: quiet literary fiction, where default model prose is too eager — too many metaphors, emotions named outright, every scene tied off with a bow.

What templates cannot fix

Three things, stated plainly so you spend your debugging time in the right layer. Model choice dominates: the gap between models imitating the same book is larger than the gap between prompts on one model, and the ranking flips by genre — audition before committing. Long-run drift is a material problem: it needs source labeling and the canon declaration above, not more forceful adjectives. And world-state consistency belongs to setting sheets and lore entries injected as material, not to a “remember all the details” line that a model will nod at and ignore.

Disclosure, since this post would be incomplete without it: Foreverse ships 23 official template packs in-app — 11 continuation styles, 3 branch-direction generators, plus video and illustration sets — free to import from the community tab. They are Chinese-first, written for the webnovel genres our earliest users read, and the full teardown of their contents is currently published in Chinese. The four English templates above are not translations of those packs; they apply the same three-part structure to the genres English readers ask about. Import a pack or paste a template — in the reader either becomes a saved option on the continuation form, and an imported template is yours to edit.

The honest endpoint of all this tuning is one number. Mean sentence length and dialogue ratio can be prompted and model-picked into range; sentence burstiness cannot — the author sat at 0.84, the six-model benchmark ranged 0.46 to 0.59, and even our best labeled follow-up run only climbed back to 0.64. If a continuation of your favorite book still reads slightly too even, too composed, that remaining gap is what you are noticing, and for now the fix is editorial, not prompt-shaped.