ZenStory AI

Practical guide · 实用指南

Writing fiction with DeepSeek: choose chat, the DSH plugin or a workbench

用 DeepSeek 写小说：聊天、DSH 插件与工作台怎么选

Source checked 2026-09-12 · 源码核对日期

Choose by where the writing needs to live. Use an ordinary DeepSeek model chat for a disposable brainstorm, outline critique or scene rewrite that you will copy into your own manuscript. Use DeepSeek Harness (DSH) plus the community Oh Story DSH plugin when you want a local writing folder, persistent sessions, creation skills and a file/editor/Chat workbench. Use the separate hosted ZenStory service at app.zenstory.ai when you want an account-based browser project instead of running DSH locally. These are different layers: DeepSeek is the model/provider choice, DSH is the agent host, @oh-story/dsh is a community plugin for that host, and hosted ZenStory is a separate application. Installing the plugin does not add it to an ordinary DeepSeek chat, and this guide does not claim that hosted ZenStory natively uses DeepSeek. For the broader difference between chat, skills and agents, see AI agent skills for writers; for a field-by-field comparison of the three ZenStory AI writing environments, see the workflow comparison.

按作品需要保存在哪里来选。只想临时脑暴、点评大纲或改一段文字，并准备自己复制回稿件，用普通 DeepSeek 模型聊天即可；希望围绕本地写作目录持续创作，由会话、Skills 和文件/编辑器/Chat 工作台共同管理项目，可选 DeepSeek Harness（DSH）加社区插件 Oh Story DSH；不想在本地运行 DSH，而想用账户化的浏览器项目，则选独立的 ZenStory 托管服务 app.zenstory.ai。四者不是同一层：DeepSeek 是模型/Provider 选择，DSH 是 Agent 宿主，@oh-story/dsh 是装进该宿主的社区插件，ZenStory 托管版是另一套应用。安装插件不会把它装进普通 DeepSeek 聊天，本指南也不声称 ZenStory 托管版原生使用 DeepSeek。 若要先理解聊天、Skills 与 Agents 的通用差异，可阅读写作者的 AI Agent Skills 指南；若要逐项比较 ZenStory AI 的三种写作环境，可查看工作流对比。

Oh Story DSHSource on GitHub

Before you start · 开始之前

Steps · 操作步骤

  1. Pick the smallest environment that matches the job For one question—three premise variants, feedback on a synopsis, or a rewrite of one supplied scene—ordinary chat keeps setup low. For an 80,000-word draft or a serial whose settings, chapter plans, prose and continuity must survive later sessions, consider DSH plus the plugin. If the manuscript should live in an account-based browser project and you do not want to run a local DSH process, use hosted ZenStory. If you already use another supported coding-agent host, follow the separate agent-skills guide rather than assuming the DSH plugin installs there.
  2. Install the community plugin in DSH Web Run the versioned command in the example. It temporarily supplies pnpm because DSH's plugin add needs it, installs @oh-story/dsh into the web profile, and then starts DSH Web. Keep the terminal running. If a browser does not open, use the complete local URL printed by DSH; do not share its first-auth token.
  3. Configure the model at the host layer In DSH, add the provider and model under Settings → Models, or start DSH with DEEPSEEK_API_KEY present when that is the provider you intend to use. This configuration belongs to DSH, not to the Oh Story plugin. Viewing existing work can be deferred without a model key, but generating new prose requires a configured model.
  4. Open the writing folder before prompting Use the plus button beside Workspaces, add the folder that should own the novel, choose that workspace and open a session. An empty folder initially stays in native DSH Chat. The novel workbench appears after recognized story files are created; a visible tab is a view of the same DSH workspace, not a second agent or a ZenStory account.
  5. Start with a bounded book brief Use /story-long-write 帮我开书 (or the English intent in the example) and state genre, reader promise, protagonist, central pressure, length target, viewpoint, tone, hard exclusions and the desired stopping point. The bundled workflow's default opening path creates the project, core settings, volume outline and the first ten chapter blueprints, then stops before prose. Review the direction instead of asking for an entire novel in one turn.
  6. Write one chapter and inspect the files After approving the setup, explicitly request /story-long-write 写第1章. Check the actual file under 正文/, its matching file under 大纲/, and the relevant settings rather than judging only the chat summary. Give corrections as concrete constraints—what fact, voice, reveal boundary or ending beat must change—then request a revision of the named chapter.
  7. Continue from the persistent workspace On a later visit, reopen the same DSH workspace and an existing or new session, then ask /story to inspect the current project before continuing. Name the next scope, such as one chapter, and tell it not to replace manual edits. The long-writing workflow distinguishes a named chapter, daily continuation and revision; a bare command diagnoses project state instead of silently producing chapters. Back up the manuscript folder separately from host/profile configuration.
  1. 按任务选择最小够用的环境 只问一个问题——例如给出三个梗概、点评简介或改写你提供的一场戏——普通聊天准备最少。已有八万字稿件，或需要让设定、逐章细纲、正文与连续性状态跨会话保留的连载，可选 DSH 加插件。希望稿件位于账户化网页项目、又不想在本地运行 DSH，则用 ZenStory 托管版。若本来就在其他受支持的 coding-agent 宿主里工作，请按单独的 Agent Skills 指南选择，不要假设 DSH 插件也能直接装进去。
  2. 在 DSH Web 中安装社区插件 运行示例中的固定版本命令。命令临时提供 pnpm，因为 DSH 的 plugin add 会用到它；随后把 @oh-story/dsh 装入 web profile，并启动 DSH Web。终端需要保持运行。浏览器没有自动打开时，使用 DSH 在终端打印的完整本地地址，但不要分享其中用于首次认证的 token。
  3. 在宿主层配置模型 在 DSH 的“设置 → 模型”中添加 Provider 和模型；如果准备使用对应的 DeepSeek Provider，也可以在启动 DSH 前提供 DEEPSEEK_API_KEY。这是 DSH 的配置，不是 Oh Story 插件的配置。查看已有作品可以稍后再配模型，但生成新文本需要可用的模型配置。
  4. 先打开作品目录，再发写作请求 点击 Workspaces 旁的加号，添加真正存放小说的目录，选中 workspace 并打开会话。空目录起初仍显示 DSH 原生 Chat；Agent 写出可识别的小说文件后，小说工作台才会出现。这个标签只是同一 DSH workspace 的视图，不是第二个 Agent，也不是 ZenStory 账户。
  5. 用有边界的开书请求起步 输入 /story-long-write 帮我开书，并写清题材、读者承诺、主角、核心压力、预计篇幅、视角、语气、明确禁区和本轮停靠点。内置流程默认先建立项目、核心设定、卷纲与首批十章细纲，然后停在正文前。先审方向，不要一轮要求生成整本小说。
  6. 明确写一章，并检查真实文件 确认设定和细纲后，再明确输入 /story-long-write 写第1章。检查 正文/ 中的实际章节、大纲/ 中对应细纲及相关设定，不要只看 Chat 摘要。修改意见要落到可执行约束：哪个事实、人物声音、信息揭示边界或章尾节拍需要变化，然后点名要求修改该章。
  7. 从持久化 workspace 继续 下次打开同一 DSH workspace，进入已有或新会话，先让 /story 检查当前项目再继续。明确本轮只写下一章等范围，并要求保留人工修改。长篇流程会区分指定章节、日更续写与回炉修改；裸命令只诊断项目状态，不会悄悄批量写正文。备份时应把稿件目录与宿主/profile 配置分别保管。

Example · 示例

Suggested input — original example, not an observed model answer:

Terminal — install the pinned plugin and start DSH Web:
npx -y --package pnpm@11.7.0 --package @deepseek-ai/dsh@0.1.5-rc.1 dsh plugin --profile web add @oh-story/dsh@0.1.9 &&
npx -y @deepseek-ai/dsh@0.1.5-rc.1 web

In the DSH session for your chosen empty workspace:
/story-long-write 帮我开书
Create a 240,000-word near-future mystery serial for adult readers.
Premise: a municipal archivist discovers that one ordinary resident disappears from every public record at 02:17 each night, but her handwritten corrections return by morning.
Reader promise: procedural investigation plus the cost of deciding which version of a person deserves to survive.
Protagonist: meticulous, conflict-averse, and personally implicated by a missing sibling.
Use close third person, restrained prose, and no omniscient explanation of the system in the opening volume.
For this turn, create the project settings, volume outline and first ten chapter blueprints only. Stop before prose and list the files created.

After reviewing the plan:
/story-long-write 写第1章
Use the approved chapter blueprint. Write only chapter 1, preserve the reveal boundary, and end when the protagonist sees her own correction handwriting in a file she has never opened. Save the prose in the project; do not continue to chapter 2.

In a later session in the same workspace:
/story Please inspect the current novel files and continuity state first. Continue with only the next planned chapter. Preserve my manual edits to chapter 1, report any missing or conflicting blueprint before writing, and summarize the files changed when done.
建议输入——以下为原创示例，不是实测模型回答：

终端——安装固定版本插件并启动 DSH Web：
npx -y --package pnpm@11.7.0 --package @deepseek-ai/dsh@0.1.5-rc.1 dsh plugin --profile web add @oh-story/dsh@0.1.9 &&
npx -y @deepseek-ai/dsh@0.1.5-rc.1 web

在你选定的空 workspace 对应 DSH 会话中输入：
/story-long-write 帮我开书
写一部约 24 万字、面向成年读者的近未来悬疑连载。
核心梗：市档案员发现，每晚 02:17 都有一名普通居民从全部公共记录中消失，但她手写的修订次日又会恢复。
读者承诺：程序化调查，以及“哪个版本的人值得留下”带来的选择代价。
主角：严谨、回避冲突，并因失踪的手足与事件有私人关联。
使用第三人称限知和克制文风；第一卷不得全知解释系统真相。
本轮只建立项目设定、卷纲和首批十章细纲，停在正文前，并列出创建的文件。

检查并确认计划后输入：
/story-long-write 写第1章
严格使用已确认的本章细纲，只写第 1 章，保留信息揭示边界；在主角发现一份从未打开的档案里出现了自己的修订笔迹时收尾。正文写入项目文件，不要继续第 2 章。

以后在同一 workspace 的会话中输入：
/story 请先检查当前小说文件和连续性状态，再只续写计划中的下一章。保留我对第 1 章的人工修改；如果细纲缺失或冲突，先报告，不要猜写。完成后说明改动了哪些文件。

Expected files · 预期文件

Verify the result · 验证结果与边界

Sources and version notes · 来源与版本说明