Practical guide · 实用指南

Agent skills for writers: prompts, skill packs, plugins and MCP explained

不会编程也能用写作 Skill 吗？提示词、技能包、插件与 MCP 怎么分

Source checked 2026-09-12 · 源码核对日期

An agent skill packages task instructions and optional resources so a compatible AI agent can reuse a workflow. For a writer, that might mean turning an idea into an outline, importing an existing manuscript or revising a chapter against project notes—not writing software. Oh Story is a fiction-writing skill pack, not a new model or a standalone chat app. Start with the writing job you need done, then choose the host and workflow.

Agent Skill 把任务说明和可选的参考资料、脚本等打包，让兼容的 AI Agent 复用一套工作流程。对作者来说，任务可以是把灵感整理成细纲、导入已有小说，或结合项目设定修改章节，不一定是编程。Oh Story 是小说创作技能包，不是新模型，也不是独立聊天软件。先想清楚要完成哪项写作任务，再选运行环境与流程。

Oh Story Source on GitHub

Before you start · 开始之前

One concrete writing task and material you own or are authorized to use. A premise, a chapter needing revision or an unfinished manuscript is enough to choose a starting point.
For Oh Story, a configured supported agent host and a writing-project folder. You do not need to author code for the writing tasks below, but you do need to open the right folder, follow installation instructions and review requested file/tool access. If you prefer not to manage that setup, see the separate hosted workbench.

一项具体写作任务，以及你拥有或获准使用的材料。一个故事点子、一章待修改的正文，或一部未完成的小说，都可以作为起点。
使用 Oh Story 需要先配置受支持的 Agent 宿主，并准备写作项目目录。下面的写作任务不要求你编写代码，但需要会打开正确的目录、按说明安装，并判断是否允许文件与工具操作。不想管理这套环境，可以了解独立的托管写作工作台。

Steps · 操作步骤

Decide whether a prompt is enough For “suggest five titles” or a one-off paragraph edit, a clear prompt may be all you need. Prompts can also be reusable templates. Consider a skill when you repeatedly paste the same process, refer to the same resources or need a consistent handoff between research, outlining and revision. Packaging makes the process reusable; it does not make every answer better.
Distinguish the instruction from the package A skill has a SKILL.md entry with metadata and instructions, and may include references, templates or scripts. A compatible host discovers it and loads relevant instructions when needed. “Skill pack” is an informal name for a collection of related skills, not a separate object defined by the Agent Skills standard. Oh Story groups writing, import, analysis and revision tasks this way.
Separate the model, host, plugin and MCP connection The model generates responses; the host runs the session and manages tools and permissions. In Claude Code, a plugin is a distribution container that can include skills, agents, hooks and MCP servers. MCP is a protocol for connecting AI applications to external data and tools, not a writing model or a synonym for a skill. These layers can work together; installing a skill does not train a new model.
Choose the entry by what you already have Only an idea: use story-long-write to plan a long-form project, or story-short-write for a short story. Existing manuscript: use story-import before continuing, rather than asking for chapter 21 in an empty project. A draft needing editorial feedback: use story-review; a passage needing less formulaic wording: use story-deslop. Unsure which to pick? Ask the story router to choose. Oh Story specializes in Chinese web-fiction workflows; its rubrics are not universal standards for all fiction.
Install one matching route, then work in your book folder The Oh Story project page shows the npx skills installation path. In Claude Code, that path uses /story-setup, followed by a fresh session and /story. The Claude Code marketplace plugin has namespaced commands such as /oh-story:story-setup instead. Follow the path you actually installed; other hosts can use different command syntax and may not support the same hooks or specialist agents.
Give a brief with a clear stopping point Specify genre, reader experience, viewpoint, constraints and the deliverable. In the example below, the first request is for planning, not prose. Read the outline and resolve story choices, then request the first chapter explicitly. The example is a suggested brief, not output from a benchmark or a claim that a single prompt produces a finished novel.
Continue the project, not just the conversation On the next session, return to the same writing folder and ask what is already established and what comes next. Point to the exact chapter or passage when requesting revisions. Keep author preferences separate from facts about this book, and ask the agent to update the relevant project records rather than only acknowledging a correction in chat. For DeepSeek Harness, use the separate DSH writing guide.

先判断一句提示词是否已经够用 只是“想五个书名”或临时修改一段文字，一条清楚的提示词往往就够了；提示词也可以保存成可复用模板。如果你反复粘贴同一套步骤、反复提供同一批参考资料，或需要让拆文、列纲、修订之间有固定交接，再考虑 skill。打包让流程可复用，不等于每次回答都会更好。
分清指令与技能包 一个 skill 以 SKILL.md 为入口，包含元数据和任务说明，也可以附参考资料、模板或脚本。兼容宿主发现它后，按需加载相关说明。“Skill pack / 技能包”通常是一组相关 skills 的非正式叫法，不是 Agent Skills 标准另外定义的对象。Oh Story 就把写作、导入、拆解、修订等任务组织成一组技能。
把模型、宿主、插件和 MCP 分开看 模型负责生成回答；宿主负责运行会话、管理工具与权限。在 Claude Code 中，plugin 是分发容器，可以一起装入 skills、agents、hooks 和 MCP servers。MCP 是让 AI 应用连接外部数据和工具的协议，不是写作模型，也不是 skill 的同义词。这些层可以配合使用；装一个 skill 并不等于训练了一个新模型。
按手头材料选择入口 只有点子：长篇用 story-long-write 规划，短篇用 story-short-write。已有小说：先用 story-import 导入，而不是在空目录里直接要求写第 21 章。已有正文想听编辑意见：用 story-review；只想改掉某段的套话和机械表达：用 story-deslop。不知道该选哪个，就让 story 路由判断。Oh Story 侧重中文网文流程，它的评审尺度不是所有小说的通用标准。
选一种匹配的安装方式，在自己的书稿目录里工作 Oh Story 项目页提供 npx skills 安装方式。在 Claude Code 里，这条路径使用 /story-setup，完成后新开会话，再用 /story。Claude Code marketplace 插件则使用 /oh-story:story-setup 这样的命名空间命令。按实际安装路径使用命令，不要混用；其他宿主的语法以及 hooks、专用 agent 支持也可能不同。
交付要求写清楚，保留作者决策 说明题材、读者体验、视角、限制和交付物。下面的首次请求只要规划，不要正文。先读细纲、解决故事选择，再明确要求写第一章。示例是一份可改写的创作简报，不是模型实测产物，也不表示一句话就能得到可发表的整本小说。
续的是同一个项目，不只是同一段聊天 下次打开时，回到同一个写作目录，先询问已确定的内容与下一步。修订时指出具体章节或段落；把作者习惯与这本书的故事事实区分开，要求更新相关项目记录，而不是只在聊天里答应“记住了”。如果使用 DeepSeek Harness，接着看独立的DSH 写作指南。

Example · 示例

In Claude Code, after the npx skills installation and /story-setup, open a fresh session in your writing folder. These are suggested requests, not recorded model output.

/story-long-write 帮我开书
Plan an original Chinese-language coastal mystery serial.
The protagonist repairs old radios and avoids confronting her family.
Use close third person. Clues must be observable; no supernatural solution.
First deliver the premise, character constraints, volume outline and first ten chapter outlines.
Stop before prose. List story decisions that need my input.

After reviewing and revising the plan:
/story-long-write 写第1章
Follow the agreed outline. Let the conflict emerge through dialogue and action.
Write only chapter 1, then summarize the changed project records.

在 Claude Code 中通过 npx skills 安装并完成 /story-setup 后，在写作目录新开会话。下面是建议请求，不是模型运行记录。

/story-long-write 帮我开书
规划一部原创的中文海港悬疑连载。
主角修理旧收音机，遇到家庭矛盾时习惯回避。
第三人称限知视角，线索必须可观察，不用超自然力量解谜。
先交付核心卖点、角色约束、卷纲和前十章细纲。
不要写正文；列出需要我决定的故事问题。

读完并修改规划后：
/story-long-write 写第1章
按已确认的细纲推进，用对话和行动呈现冲突。
只写第一章，最后说明本次更新了哪些项目记录。

Expected files · 预期文件

Setup creates the host-specific skill/configuration files for the chosen installation. Those are the working environment, not the manuscript.
Planning produces book settings and outlines; an explicit chapter request produces prose. For an imported book, inspect the reconstructed project before asking for continuation.
Chapter continuity uses project records such as tracking state, not a promise that the model remembers every message. Preserve the writing folder and a backup of material you want to keep.

环境部署生成所选宿主的技能与配置文件。这些是工作环境，不是书稿本身。
规划阶段交付设定与大纲，明确要求写章后才交付正文。已有小说导入后，先读重建的项目材料，再要求续写。
章节连续性依靠追踪状态等项目记录，不靠“模型记得每句话”的承诺。保留写作目录，并备份需要保存的材料。

Verify the result · 验证结果与边界

Read for the intended scene, character choices and continuity before accepting a chapter. A completed workflow is not a guarantee of literary quality, detector scores or publication acceptance.
Review installation trust and model/service configuration separately. Open-source skills are not free model access, and files on your computer do not imply offline inference.

接纳一章前，先读场景是否成立、人物选择是否可信、前后事实是否一致。流程完成不保证文笔、检测器分数或投稿通过。
安装信任与模型、服务配置需要分别判断。开源技能不等于免费模型额度，文件放在本机也不等于推理离线。