文档目录 · Contents
ZenStory AI ZenStory Workbench 工作台文档
AI 创作助手
AI 可以讨论故事、起草正文、检查设定，也可以读取和操作项目文件。每一轮说清要做什么、依据什么，以及写到哪里停止。
提出一项明确任务
一个好用的请求包含目标文件、已确认事实、允许改动的范围和停止点。例如：
先读取《第1场》和《三场景提纲》。
把第2场写入当前正文文件，采用林禾的有限视角。
包还在窗口，林禾不知道报错原因。
写到沈敏说“包不是我的”时停止，不继续写结局。
讨论方向时写明“只给方案，不修改文件”。直接入稿时说明替换还是追加，避免把同一场重复写入。
提供上下文
| 资料 | 提供方式 |
|---|---|
| 当前文件 | 打开目标文件；AI 根据焦点读取已保存内容 |
| 人物、提纲和旧章节 | 点名要求先读取，或选中关键段落添加到对话 |
| 项目素材 | 在素材条目上添加到对话 |
| 素材库条目 | 从项目素材标签中选择并附加 |
| 长期背景和进度 | 在 AI 记忆中保存简短摘要 |
素材附件最多 5 项，文本引用另有 5 段上限。移除不相关项，给本轮留出需要的资料。上下文有长度预算，附加一本书不等于每次读完全书。
发送前保存刚改过的设定。文件名不唯一时，让 AI 先列出候选，再指定目标。
使用技能和生成模式
在输入框用 / 选择技能，或从项目的技能标签添加。本轮最多主动选择 3 个技能；选择后仍要写清任务。也可以点名某个技能，让 AI 按需使用。
输入框提供快速和深度模式。根据任务选择；模式不会替代具体要求，也不保证固定耗时或字数。
看回复，也看文件
操作卡片用于核对创建、编辑和读取结果。回复里出现候选正文，未必已写入文件；直接打开目标文件查看。
可以在生成过程中补充要求，界面会提示是否已接收。需要结束这轮时点击停止。停止会结束继续生成，已发生的文件修改仍会保留。
中断后继续
先看当前文件和本轮操作记录，确认实际写到哪里，再发有范围的续写请求：
先读当前《第2场》，列出已完成与尚缺的部分。
保留已写段落，只补缺口，不重新追加整场。
只有未写入任何内容且页面提供重试时，才按提示重试原任务。不要在连接状态不明时反复发送同一请求。
新建会话
对话顶部的新建会话切换聊天，不删除项目文件或 AI 记忆。换会话前把已采用决定放进文件，留下实际进度；新会话再指定目标和依据。
遇到今日消息用完的提示，等北京时间次日 00:00 恢复，或查看 Pro。用量见订阅权益。
AI writing assistant
The AI can discuss stories, draft prose, check consistency, and read or operate on project files. State the task, its evidence, and where to stop each turn.
Describe one clear task
Include a target file, established facts, permitted changes, and a stopping point:
Read Scene 1 and Three-scene outline first.
Write Scene 2 into the current draft file, in Lin He's limited viewpoint.
The bag remains at the desk; Lin He does not know the cause of the mistake.
Stop when Shen Min says “The bag isn't mine.” Do not write the ending.
For planning, say “Give options only; do not change files.” For drafting, specify whether to replace or append so a scene is not written twice.
Supply context
| Evidence | How to provide it |
|---|---|
| Current file | Open the target; the focused file identifies saved content to read |
| Characters, outlines, earlier chapters | Name files to read, or select key text and add it to chat |
| Project materials | Attach the material item to chat |
| Library items | Select and attach from the project's Materials tab |
| Long-term background and progress | Save a short summary in AI Memory |
Up to 5 material attachments and, separately, 5 text quotes are supported. Remove irrelevant items. Context has a length budget; attaching a book does not mean reading all of it every turn.
Save recent changes before sending. If filenames are ambiguous, ask for candidates before selecting the target.
Skills and generation modes
Use / in the input to select skills, or add them from the project's Skills tab. Select up to 3 for a turn, then describe the task. You can also name a skill for the AI to use as needed.
The input offers fast and deep modes. Choose for the task; a mode does not replace clear instructions or guarantee a fixed duration or word count.
Check replies and files
Operation cards help verify file creation, edits, and reads. Candidate prose in a reply may not be saved; open the target file to check.
You can add instructions during generation; the interface reports whether they were received. Choose Stop to end further generation. File changes already made remain.
Continue after interruption
Inspect current files and this turn's operation records before continuing:
Read the current Scene 2 and list what is complete and what is missing.
Keep existing paragraphs and fill only the gap; do not append the whole scene again.
Retry the original task only when nothing was written and the interface offers that option. Do not repeatedly send the same request while connection state is unclear.
Start a new chat
New Chat switches the conversation without deleting project files or AI Memory. Before switching, save accepted decisions in files and record actual progress. Identify the target and evidence again in the new chat.
If daily messages are exhausted, wait until 00:00 Beijing time the next day or check Pro. See Plans & Benefits.
Related: local revision, AI Memory, and writing workflow.