Finishing a dropped novel for yourself: the whole workflow, honestly

The serial I am thinking of stopped in the spring of 2024, at chapter 214, on a cliffhanger the author clearly intended to resolve within a week. Two years of silence now. No hiatus announcement, no goodbye post — the update schedule just flatlined, the way web serials actually die. I reread the whole thing last winter, hit chapter 214 again, and did the thing I used to only joke about: I gave it an ending myself. Not to post anywhere. Just so the story would stop living in my head as an open loop.

This is the workflow, start to finish, including the parts that did not work on the first try. None of it requires writing skill or technical setup beyond installing a reader app — what it requires is a couple of evenings and a clear memory of why you loved the book.

Step one: get the book out as a file

Whatever you feed an AI has to exist as text you control, and this step is where most people stall. My serial was easy — I had a full txt export from the reading app I originally followed it in. Hiatus fanfic is even easier: every public AO3 work has a Download button at the top of the page with five formats behind it, and the epub option grabs a multi-chapter fic as one file, no account needed. If what you end up holding is an epub and you want txt, conversion is a solved problem: Calibre, the standard open-source ebook manager, converts epub to txt (and between dozens of other formats) in one step. The point either way: end up with a file on your phone that is yours, not a cached copy inside someone else’s app that can vanish in an update.

Import itself is the anticlimax of the process. Foreverse reads txt and epub directly, and it was built for hoarder-scale libraries — the bulk-import test shoved 315 real webnovel txt files, about 1GB, through it in roughly 8 seconds. One dropped serial does not register.

Step two: find the last chapter that was still good

Here is the part no tool does for you, and the part that most decides whether the continuation feels right. Serials rarely die suddenly; they trail off. Rereading with two years of distance, I could see the author losing the thread around chapter 209 — a side-plot that goes nowhere, pacing that stalls, a character making decisions that only make sense as setup for something that never arrived. Chapter 214 was the last chapter posted. Chapter 209 was the last chapter that felt like the book.

Finding the spot took one focused evening, not a full reread. I skimmed the final thirty chapters from the chapter list, reading first and last scenes, and marked where three things last lined up: the protagonist still wanted something specific, the subplot count was still stable, and the prose still had the author’s rhythm instead of filler. All three pointed at the same chapter. Old reader comments, if the serial had them, are a shortcut here — the chapter where the comment section turns from theories to “is the author okay?” usually marks the cliff better than memory does.

So I forked at 209, not 214. That is a reader’s judgment call, and it is worth making deliberately: continue from the wrong anchor and the AI will faithfully extend the decline instead of the book you loved. Five orphaned chapters is a small price. They are still there on the canon line anyway.

Step three: grow the ending on a branch

The mechanics: select the passage where the fork should happen, tell the AI where the story should head — a direction, not a script; “the siege breaks tonight, and the cost lands on the captain” is plenty — and the continuation grows as a new branch from that exact point. Read it. Where it turns wrong, fork again from the last good paragraph and steer. My ending took an evening and a half, about forty continuations, three abandoned side branches, and one full restart when I realized I had let the tone go soft for ten consecutive segments.

What I actually typed into the direction box got shorter as the night went on. Early on I wrote paragraphs of instructions; the model obeyed them and produced something dutiful and dead. What worked better was one plot beat plus one constraint — “she finds the letter; nobody cries in this scene” — and letting the imported text do the stylistic heavy lifting. Constraints phrased as prohibitions beat descriptions of mood every time, which turns out to generalize; the prompting field notes unpack why negative rules are the only ones a model follows reliably. The other habit that stuck: when a continuation was 80% right, I stopped regenerating and just edited the two wrong sentences by hand. Regenerating for perfection is how an evening becomes a week.

Two settings did real work. Auditioning models first: I ran the same chapter-209 passage through three models before committing, and one was eliminated on sight for narrating emotions the author never names. And switching models mid-book for different scene types — the battle-heavy stretch and the quiet epilogue chapters wanted different pens, which stopped surprising me after seeing the blind-review data on how differently models imitate the same book.

The honest section: what the machine cannot hold

Style drift over long runs is real, measurable, and not fixed by any prompt I have seen. When we benchmarked nine models continuing novels for 20 consecutive rounds — output fed back into context each time, exactly what a long personal continuation does — every failure fit one of three patterns: restart loops, mid-run freezes where the model repeats its own recent output verbatim, and one model that held the author’s voice flawlessly for 17 rounds and then rewound the entire plot. The structural cause is unglamorous: as generated text fills the context, the model increasingly imitates itself instead of the author.

In practice this means a long continuation is not one heroic run; it is a series of short runs with a human editor — you — trimming, steering, and occasionally switching models to break a loop before it locks in. I hit one mild freeze around continuation thirty: two segments in a row reusing the same door-and-lantern image almost word for word. One continuation on a different model snapped it out, and I switched back after. That is the loop-breaking move the benchmark data recommends, and it costs a single segment. Plan for it. The evening-and-a-half figure above includes that editing, and the editing is most of why the result reads like the book.

If the author comes back

This was my actual hesitation before starting — some superstition that writing an ending forecloses the real one. The file structure answers it better than the superstition deserves: the canon line is untouched, so official chapters slot back exactly where they belong, and my branch stays what it always was, one reader’s private alternate ending. If chapter 215 ever posts, I will read it the same hour. I would genuinely love for my version to become the wrong one.

Chapter 214 stopped being an open loop in February. The author’s ending, if it ever comes, gets to be a gift instead of a debt — and that trade cost me an evening and a half and well under a dollar in tokens.