A waitlist user asked for danmei fanfic. Here is the honest answer.

In late June, someone on our waitlist left a single line in the message box: they wanted danmei fanfiction, Heaven Official's Blessing style. No pleasantries, just the request. We answered by email, then decided the question deserved a public answer, because the discourse on AI fanfic is split between hype accounts and “AI has no soul” dismissals, with very little honest testing in between.

The short version: one-shots work, long fics need scaffolding

We ran the test with a canon most people know: character dossiers from the first arcs, then a 2,000-word festival-meeting one-shot. Both DeepSeek and Claude produced something you could post without the comments section smelling AI on it. Scene blocking held, dialogue rhythm held, and two small details were placed more cleverly than we expected.

Long form is a different animal. Past twenty thousand words with no structural support, characters start saying lines that belong to someone else, and the object planted in chapter five is gone by chapter twelve. The model did not “forget” — the context never carried those facts in. The fix is the same one from our worldbook piece: dossiers for characters, pinned entries for rules of the world, injected every turn. Fanfic is actually the best case for this approach, because the canon already exists. You are organizing, not inventing.

Keeping characters in character, ranked by what worked

Fanfic readers have zero tolerance for OOC, which makes this the discipline that separates fanfic from original work. Ranked by measured effect: example dialogue first — three to five canon exchanges that carry the character's voice lock tone better than any adjective list. Negative constraints second — “he never apologizes first” prevents more drift than ten positive traits. Model choice third: a stronger model helps delicate emotional scenes, but no model rescues an unorganized character sheet.

And the lowest-tech trick is still the best one: for pivotal scenes, generate three candidates and choose. The in-character verdict stays with you. The tool's job is to put the options on the table.

The three-endings problem

Every fanfic writer knows this fork: the fix-it ending, the tragedy, and the self-indulgent what-if, all growing from the same scene. The traditional answer is three documents and increasingly desperate filenames. Branching turns it into a tree: three lines hanging off one divergence point, growing in parallel, switchable at will, never overwriting each other. Our record holder among testers kept seven live lines off a single fork. Her review: “I finally deleted the file called final_v2_REAL.”

The problems no tool has solved, stated plainly

Voice drift: after a few thousand words every model slides toward its own default register, and you have to pull it back with canon passages. There is no set-and-forget fix. Content range: how far romance scenes can go varies sharply by model — one argument for BYOK is that this boundary is set by the model you choose, not by us. Disclosure: archives are tightening AI-labeling rules; label honestly and do not test their patience.

To the person who left that message: we ran your scenario, and it holds up. Build the canon card, organize the worldbook, and fork from the chapter that never sat right with you. The fic you asked for still has to be written by you. The tool's job is to make it writable.