AI小说质量差的原因分析
Sherpa用多代理架构解决AI小说连贯性问题，比单窗口聊天更懂叙事逻辑。
AI生成的小说质量不佳，主要因为聊天机器人一次只能处理一个提示词，上下文窗口有限，无法记住40集前承诺的内容。Sherpa系统采用不同代理分别处理创意、结构、研究、文本和审查，同时使用"叙事世界模型"跟踪故事历史。通用大模型难以同时维持这些要素。
AI fiction reads like slop for several reasons:
> a chatbot answers 1 prompt at a time > context windows are limited. has no idea what the story already promised 40 episodes ago. > can’t keep readers hooked beyond a point
The difference is architecture. Sherpa runs separate agents for ideas, structure, research, prose and review, while a separate "Narrative World Model" keeps track of the story's history. A single chat window can't do that job.
Serialized fiction needs software that knows who's alive, who knows what and which places exist. General LLMs suck at exactly keeping all that together. This article breaks it down really well.
