webnovel-writer
A long-form webnovel assisted creation system based on Claude Code, solves forgetting and hallucination in AI writing, supports up to 2 million words serialized creation

Suggested prompts
About this skill
Webnovel Writer
A long-form web fiction writing plugin. It strings the whole writing flow together: initialising the setting, planning volume outlines, writing chapters, reviewing, persisting memory, querying state, and a read-only visual dashboard.
It solves one problem: keeping the AI's memory of the setting, the foreshadowing and the outline intact after several hundred chapters. This is a consistency system for serialised long-form fiction, not a one-shot generator that forgets what it wrote.
When to Use It
The hardest part of long-form writing is not chapter one — it is keeping things straight by chapter 80 or 200: character motivation that does not drift; power levels, timeline, locations and world rules that do not contradict each other; foreshadowing that is registered, advanced and paid off; a steady rhythm of payoffs, romance and worldbuilding; and facts from every finished chapter settling into a searchable state system. The system turns those "must remember, must not break" constraints into steps that run automatically: it checks references before writing, records what newly happened afterwards, runs a consistency review, and syncs the latest state into the retrieval index, chapter summaries, long-term memory and the dashboard.
Core Capabilities
| Capability | Command | Description |
|---|---|---|
| Deep initialisation | /webnovel-init |
Multi-stage Q&A that builds the book's skeleton, setting, master outline and initial state |
| Volume planning | /webnovel-plan |
Splits the master outline into volumes and chapters, fills in the timeline, and writes new settings back |
| Chapter writing | /webnovel-write |
Writes a chapter end to end: gathers context, drafts, reviews, polishes, records facts, backs up |
| Quality review | /webnovel-review |
Reviews chapters across payoffs, consistency, pacing, out-of-character risk, continuity and hook strength |
| State query | /webnovel-query |
Queries characters, foreshadowing, pacing, entity relationships and runtime information |
| Project learning | /webnovel-learn |
Records writing techniques that worked in this book into the project's long-term memory |
| Dashboard | /webnovel-dashboard |
Read-only browsing of project state, entity graph, chapter content and hook-strength data |
| Project health check | /webnovel-doctor |
Stage-aware check of directories, files, database, retrieval and dependency state |
Built-in Genres
37 Chinese web-fiction genre templates are built in, and several genres can be combined. A partial list:
| Group | Examples |
|---|---|
| Cultivation and fantasy | Cultivation, system, high-martial, western fantasy, infinite flow, apocalypse, sci-fi |
| Urban and modern | Urban powers, urban slice-of-life, urban brain-teasers, realism, e-sports, livestream fiction |
| Romance | Historical romance, palace and household intrigue, youth romance, CEO romance, melodrama, stand-in fiction, farming life |
| Special | Rules-based horror, mystery brain-teasers, paranormal mystery, historical, wartime espionage, Zhihu short fiction, Cthulhu |
How It Works
Writing a chapter is not a single generation pass; it is a gated pipeline:
- Preflight the project root, placeholders and system health
- Refresh the runtime requirements for this chapter
- Generate the writing brief for the chapter
- Draft the prose from that brief
- Run a multi-dimensional review; blocking issues stop the flow
- Polish, typeset and run a final AI-trace check
- Extract and register the facts that occurred in this chapter
- Project those facts into state, index, summaries, long-term memory and retrieval vectors
- Back up the chapter
The design separates "how it is written" from "what happened": prose and pacing are free to vary, but facts must be registered, reviewed and archived.
Initialisation, planning, writing and review each end with an author-facing report instead of raw JSON or long command logs. It opens with an overall status:
- Completed: the target artefacts and key checks passed; you can move to the next step.
- Partially completed: main artefacts are kept, but something was skipped, auto-handled, or left for confirmation.
- Needs your input: the system has stopped in a safe place and needs you to decide the creative direction, which facts to keep, whether to overwrite files, or how to handle a blocking issue.
- Not completed: key artefacts were not credibly produced; follow the recovery advice in the report.
The report always has three parts: files produced and their status, problems encountered and unusual timings, and suggested next steps. It only interrupts you for creative direction, fact consistency, file-overwrite risk or blocking issues.
What You Provide
- Book title, genre and creative direction: settled through multi-stage Q&A during initialisation
- Command parameters: volume number, chapter number or review range, for example
/webnovel-plan 1,/webnovel-write 45,/webnovel-review 1-5 - Optional retrieval configuration: an endpoint, model name and API key for Embedding and for Rerank; any OpenAI-compatible interface works
- Without an Embedding key it still runs — the system falls back to BM25 keyword retrieval, with weaker semantic recall
What You Get
- A complete book project directory containing contracts and chapter commits, state and index, prose, outlines, the setting bible and review reports
- For every chapter: the prose, a multi-dimensional review report, and the registered facts
- A read-only dashboard for project state, entity graph, chapter content, foreshadowing and hook-strength data
- Project-level long-term memory and captured techniques for later chapters
Usage Examples
/webnovel-init # initialise a book
/webnovel-plan 1 # plan volume 1
/webnovel-write 1 # write chapter 1
/webnovel-review 1-5 # review chapters 1-5
/webnovel-query 伏笔 # query project state
/webnovel-dashboard # open the read-only dashboard
Notes
- Commands are invoked as
/webnovel-*. - Python dependencies must be installed.
- The dashboard is read-only, and its frontend ships pre-built with the plugin, so no local frontend build is needed.
- Re-running the same chapter command checks for a trusted checkpoint first and resumes from the failure point where possible, without rewriting prose, reviews, commits or backups that are already trusted.
- Only unrecoverable failures point you to the run log.