sumeru-write

Web Novel Chapter Writing

Trigger Keywords

help me write a chapter of a novel, continue writing the following content, generate XX plot, batch write web novel chapters, expand/rewrite this passage, help me write an XX plot, continue the novel, expand this passage, rewrite this chapter, batch generate novel chapters, write an opening chapter, write a climax plot, novel content generation, help me write novel content, web novel chapter generation, generate chapters from detailed outline, write a novel according to the detailed outline, batch generate all chapters, detailed outline-driven writing

Core Features

Generate based on complete detailed outline: automatically read .sumeru/outline/chapter-outlines.json, batch generate chapters according to the detailed outline
Intelligent detailed outline matching: supports generation by chapter number, volume number, or all chapters
Generate complete chapter content based on the outline and detailed outline
Automatically adapt to web novel pacing: grab attention at the beginning, include conflict in the middle, leave suspense at the end
Maintain consistency of character personalities and plot logic
Support custom chapter length (default 4000-5000 words/chapter)
Support continuing, modifying, and adjusting existing chapter content

Continuation Rules

Continuation Mode

Support multiple modes such as continuation, rewriting, expansion, and condensation

Continuation Notes

Maintain consistency of character personalities, no OOC (Out Of Character)
Maintain the power system and worldview established in the previous text without collapse
Foreshadowing recovery should be natural, not abrupt
Keep the language style consistent with the previous text
Continue the previous plot and open up foreshadowing for the following text
If existing chapter content is incomplete, prioritize completing it

Multi-Chapter Generation

Support continuous generation of multiple chapters, automatically generated in chapter order

Detailed Outline-Driven Batch Generation

Automatic Detailed Outline Mode Reading

When .sumeru/outline/chapter-outlines.json exists, automatically enable detailed outline-driven mode:

# 生成所有章节（读取细纲，自动并行）
/sumeru-write 全部章节

# 生成指定范围章节
/sumeru-write 第1-50章

# 生成特定卷的所有章节
/sumeru-write 第1卷

# 生成特定章节
/sumeru-write 第3章,第5章,第10章

# 批量并行创作指定范围
/sumeru-write 第1-100章 批量并行

Detailed Outline Input Format Support

Support directly passing in a single-chapter detailed outline:

# 使用自然语言描述生成单章
/sumeru-write 第3章 "主角在拍卖会上获得神秘功法"

# 基于已有细纲生成
/sumeru-write 第3章 按细纲生成

Detailed Outline Data Structure Validation

Automatically validate detailed outline completeness before generation:

Check whether required fields exist
Validate whether character names are defined in characters.json
Check whether scene locations are defined in world.json
Provide supplementary suggestions for missing information

Sub-agent Parallel Batch Writing (recommended for large numbers of chapters)

When a large number of chapters (>3 chapters) need to be generated at once or detailed outline-driven mode is used, automatically enable sub-agent mode:

⚠️ Follow global constraints: each sub-Agent is responsible for at most 3 chapters(see AGENTS.md "Sub-Agent Parallel Processing Rules" for details)

Required number of Agents = ceil(total chapters / 3), automatically calculated by the scheduler
Adjacent chapters are assigned to the same Agent to maintain contextual coherence

Core Advantages

✅ Detailed outline isolation: Each sub-agent only obtains the detailed outline of the chapters it is responsible for, avoiding context overflow
✅ 3-chapter cap guarantee: Each Agent has at most 3 chapters, ensuring generation quality and consistency
✅ Context isolation: Each sub-agent does not carry historical chapter content, thoroughly solving the long-context compression/overflow problem
✅ Speed improvement: Multiple parallel writing, speed is N times that of serial
✅ Error isolation: Failure in generating a single chapter does not affect other chapters, failed chapters are automatically retried
✅ Memory optimization: Sub-agents are automatically destroyed after completion, releasing memory resources
✅ Incremental writing: Each chapter is saved immediately after being written to chapters/, no need to wait for all to complete
✅ Progress visualization: Displays in real time the status of completed/in-progress/pending chapters

Scheduling logic (detailed outline-driven)

flowchart LR
    A[批量写作任务] --> B[读取chapter-outlines.json]
    B --> C[验证细纲完整性]
    C --> D[创建任务队列，按每Agent最多3章分配]
    D --> E[计算所需Agent数 = ceil/总章数/3/]
    E --> F[启动N个并行子agent]
    F --> G[子Agent拉取任务 → 获取对应章节细纲 → 生成章节 → 保存文件]
    G --> H{队列是否为空?}
    H -->|否| G
    H -->|是| I[汇总进度，生成完成报告]

Chapter allocation rules

Allocated consecutively in chapter order, e.g., Agent1 is responsible for chapters 1-3, Agent2 is responsible for chapters 4-6, and so on
Agents with fewer than 3 chapters remaining at the end are allocated according to the actual number of remaining chapters
Adjacent chapters are allocated to the same Agent to maintain contextual continuity

Sub-agent input context

Complete chapter content, conforming to the specified style and pacing
Preview/idea suggestions for the next chapter's content
Summary of key plot points in this chapter
Hints for foreshadowing planted in this chapter (optional)
Summary of character growth/changes (optional)

Chapter file naming convention

Mandatory format:{三位章节号}-{章节标题}.md

Rule	Description	Example
Chapter number	Three-digit zero-padded	`001`、`003`、`042`、`128`
Separator	English hyphen `-`	`-`
Title	Original chapter title, without special characters	`第一次解析`、`废物觉醒系统`
Extension	`.md`	`.md`

Naming example:

003-第一次解析.md
001-废物觉醒系统.md
042-时间线冲突.md
128-终极决战.md

❌ Incorrect naming:

第3章-第一次解析.md    # 不要加"第X章"前缀
3-第一次解析.md         # 章节号必须三位零填充
003_第一次解析.md       # 分隔符是短横线，不是下划线
003-第一次解析.txt      # 扩展名必须是.md
003第一次解析.md        # 必须有短横线分隔符

Data persistence

Formal output (visible to users)

Generated chapters are saved by default under the current working directory's chapters/ , strictly following the above naming convention
Chapter files are pure body content, without any intermediate markers or metadata, and users can read and edit them directly
Supports custom chapter output directories
Batch generation progress report：chapters/WRITE_PROGRESS.md(updated in real time)

Intermediate process data (for internal system use only)

All intermediate states, metadata, and progress information are uniformly saved to the .sumeru/write/ directory:

progress.json: Creation progress tracking, including completed chapters, word count statistics, and the status of each chapter
chapter-meta.json: Metadata for each chapter, including core events, characters appearing, exciting point locations, and foreshadowing records
character-state.json: Dynamic tracking of character status, recording changes in each character's abilities, relationships, and status at each point in time
used-outlines.json: Records of used chapter detailed outlines, supporting incremental generation
original/: Backup directory for original chapter files. When review or polish modifies chapters/ files, the original chapter files are first automatically backed up to the .sumeru/write/original/ directory, ensuring rollback is possible

Coordination with other Skills

Preceding Skill: Automatically reads the outline data in the .sumeru/outline/ directory
- Use characters.json to maintain character personality consistency
- Use chapter-outlines.json incomplete chapter outlineto drive batch generation
- Use world.json to maintain worldbuilding consistency
subsequent Skills: the generated chapter data can be used by sumeru-review、sumeru-polish、sumeru-finalize use

resume from checkpoint

On each task startup, read chapters/ existing chapter files in the directory and .sumeru/write/progress.json progress
Read .sumeru/outline/chapter-outlines.json to get the complete outline
Continue from the latest unfinished chapter, automatically skipping chapters that have already been generated
Support resuming creation from a specified chapter
Support generating only missing chapters (incremental mode)