One writer used the scene almost verbatim. The other called Fanon useless. Both were writing the same kind of scene, a confrontation between characters with a long shared history, and both used the same tool and model. One supplied a bare prompt and received a generic argument. The other supplied three paragraphs of context on each character's history, what each needed from the confrontation and would not say aloud, plus the scene's emotional function in the chapter arc. The results were entirely different.
The cause was neither talent nor luck. It was understanding what the tool does.
The autocomplete mistake
Many writers approach AI writing tools as extremely fast autocomplete: provide the beginning of a sentence, and the system completes the thought. In creative work, that premise leads to poor results because writing is not sentence completion. It is a chain of decisions about what comes next, informed by what came before and what still needs to happen.
If you use AI for sentence completion, you provide fragments and expect it to reconstruct a story from context it has never received. It can produce a plausible continuation of the fragment, but that continuation may have no connection to your actual story. The result feels generic because it is generic. Without the details that permit specificity, the system has no basis for producing anything else.
A better model is this: language systems are strong at recognizing patterns and continuing them within rich context. Their limits track the amount of useful context available. Supply the character's voice, the scene's emotional stakes, the conflict's history, and constraints established in earlier chapters. Make the world sufficiently detailed that the system has something real to draw on. The resulting prose will gain that same depth.
What you need to give the system
For fan fiction, a lore-aware system generally needs several kinds of context. Character context covers established facts about everyone in the scene: their desires, fears, speech patterns, and the things they would refuse to say aloud. Scene context covers what happened immediately before and what must be ready afterward. World context covers the setting, the universe's rules, and limits established by earlier chapters or the source material.
Strong scene prompts explain the scene's emotional and narrative purpose before specifying its action. "Kael and Orren have not spoken since the betrayal in chapter four. This scene needs to establish that Kael knows more than Orren thinks he does, without revealing what Kael will actually do with that knowledge. Kael's register in the source material is dry and controlled. He would not raise his voice in this moment even though he is furious." That gives a writing system usable direction. "Write a scene where Kael confronts Orren" does not.
Voice generation versus plot generation
These are different uses, and AI handles them with different levels of success.
Voice generation, or writing in a particular character's established register, is a task language models handle well when examples and constraints are clear. Provide several paragraphs of source prose featuring the character, along with notes on speech patterns and this scene's emotional register. The result can be genuinely useful as a starting draft or as a test of whether your own prose has captured the intended voice.
Plot generation, or choosing what should happen next, is a much weaker use of AI. A language model cannot judge what would satisfy the needs of your particular story. It creates plausible continuations from patterns in its training data and the context you provide. It does not know that a beat should echo a theme from two chapters earlier, or that the payoff you are building requires one more scene of sustained tension. Such decisions depend on seeing the entire story arc, and that understanding belongs to the writer.
The most effective pattern combines both roles: use the system for voice and scene drafts, while keeping plot decisions for yourself. First decide what must happen narratively, then request several versions. Choose the strongest version or merge parts of several. Do not ask the system to decide what happens next.
The editing relationship
AI output is raw material, not finished prose. Any passage produced during co-writing needs a writer's editorial judgment before entering the story. That is not an obstacle to overcome; it is the productive shape of the relationship.
For many writers, revising generated prose takes less time than starting from a blank page, and revision provides quality control at the same time. You encounter the passage as both reader and author, making strengths and weaknesses visible. The generated scene may have the right emotional arc but the wrong wording, or the right voice but poor pacing. Correcting those problems is quicker than beginning again and leaves creative judgment in your hands.
Where AI is genuinely weak in this context
Scene-level consistency is a real capability. Long-range narrative coherence is not. A language model may write a scene that fits the immediate context while contradicting something established twenty chapters earlier and omitted from its context window. Lore-safety features like the ones in Fanon reduce this risk by checking generated text against your lore documents, but they do not remove it. Generated content still needs to be read against the whole story, not only the nearby context.
Language models also struggle with genuine emotional surprise. They produce the feeling that seems plausible from the setup. The unexpected moment that stays with a reader, one that feels exactly right in hindsight although it was unforeseeable, is difficult to generate. Those moments arise when writers disrupt patterns in ways readers could not predict. That creative choice remains yours.
The tool helps most with mechanical work: putting a scene's first draft on the page, keeping a voice consistent across a long project, and producing variations for comparison. It helps least with the work requiring deepest creative judgment: deciding what the story means, choosing a direction when the characters' next step is unknown, and locating the emotional truth of a difficult scene. Those decisions remain entirely the writer's responsibility.