
SuperWriter
What Is This?
Webnovel Writer is a complete authoring pipeline built as a Claude Code plugin, designed for one specific problem: writing serialized novels of one million words or more without the AI losing track of characters, settings, foreshadowing, or tone.
It is not a single prompt or a chapter generator. It is a coordinated system of seven skills, four specialized agents, a contract-driven state engine, a RAG retrieval layer, and a read-only web dashboard — all working together to turn long-form fiction writing from "prompt engineering" into "project engineering."
If you have ever tried to write a long novel with an LLM and watched it forget the protagonist's sword name by chapter 30, invent a brand-new faction out of nowhere by chapter 50, or quietly drift the main character's personality into someone unrecognizable by chapter 80 — this plugin is the answer to all three problems at once.
The Problem It Solves
Large language models, even the best ones, suffer from three structural weaknesses when writing long-form fiction:
| Weakness | Symptom | How Webnovel Writer Fixes It |
|---|---|---|
| Forgetting | Settings from chapter 1 contradict chapter 50 | .story-system/ contracts + entity database + RAG recall |
| Fabrication | New factions, powers, or characters appear out of thin air | Data Agent extracts entities, enforces disambiguation |
| Character Drift | Protagonist gradually becomes Out-Of-Character | Reviewer agent runs OOC checks, blocks bad commits |
| AI Tone | Prose reads mechanical, "AI-flavored" | Anti-AI Force Check in the polishing stage |
| Foreshadowing Loss | Plants are forgotten before payoff | Urgency-sorted foreshadowing tracker with temporal queries |
The fundamental insight behind the system: fiction writing at scale is a software engineering problem, not a creative-writing problem. Once you accept that, the solution becomes obvious — use event sourcing, contracts, projections, and pipelines, the same patterns that keep large software systems coherent.
Architecture Overview
The system is organized in three layers that mirror the lifecycle of every chapter.
Layer 1: Pre-Write Truth (Contracts)
Before a single word is drafted, the system loads contracts from .story-system/:
MASTER_SETTING.json— the writing constitution: genre, tone, taboos, anti-patternsvolumes/volume_NNN.json— volume-level pacing and arc commitmentschapters/chapter_NNN.json— chapter directive: goal, time anchor, mandatory nodesreviews/chapter_NNN.review.json— forbidden zones and review criteria
These contracts are law. The drafting agent cannot deviate from them; the reviewer agent verifies against them; the commit pipeline rejects anything that violates them.
Layer 2: Pipeline (Multi-Agent Orchestration)
When you run /webnovel-write 45, six steps execute in order:
Step 1 context-agent → research + assemble the writing brief
Step 2 Claude (drafting) → generate prose against the brief
Step 3 reviewer agent → six-dimension quality review
Step 4 Claude (polish) → rewrite style, run Anti-AI check
Step 5 data-agent → extract events, deltas, summary
Step 6 chapter-commit → finalize CHAPTER_COMMIT + Git backup
If any step produces a blocking issue, the pipeline halts. No partial work enters the canon.
Layer 3: Post-Write Truth (Commits & Projections)
After a chapter is accepted, the system writes:
- Truth source:
commits/chapter_XXX.commit.json— an immutable record of what actually happened - Projections (auto-generated views, rebuildable from commits):
state.json— current runtime stateindex.db— SQLite entity database for queriessummaries/— chapter summaries for context recallmemory_scratchpad.json— long-term writing patternsvectors.db— RAG embeddings for semantic search
This is classic event sourcing: the commit chain is canonical; everything else can be rebuilt deterministically. If your index.db gets corrupted, regenerate it from commits. If you change your mind about chapter 30, the system handles cascading updates.
Core Capabilities
1. Structured Project Initialization
/webnovel-init is a guided, multi-phase interrogation. The system refuses to generate anything until you have committed to:
- Story Core: genre (supports A+B fusion like "Cultivation × Infinite Worlds"), word/chapter count, one-line premise, target reader
- Characters: protagonist's desire, flaw (mandatory — flat heroes get rejected), antagonist tiers and mirror dynamics
- Golden Finger (the power gimmick): type, style, visibility, and crucially, irreversible cost — the system refuses "power without consequence"
- Worldbuilding: scale, power system, factions, economy
- Creativity Constraints: anti-trope rules, hard constraints, opening hook
It ships with 37 built-in genre templates spanning Chinese webnovel categories (cultivation, system-driven, urban fantasy, palace intrigue, regression, rule-based horror, Cthulhu, and more). Each template includes anti-trope libraries and naming conventions specific to the genre.
You can also feed the system a reference book — the deconstruction-agent will analyze its golden three chapters, cool-point loops, and protagonist patterns, then deliver them as inspiration (never as material to copy verbatim).
2. Multi-Level Outline Planning
/webnovel-plan 1 decomposes the master outline into:
- Volume pacing: chapter ranges, climax positioning, sub-arc beats
- Hard timeline constraints: every chapter carries a time field, monotonically increasing; backward jumps must be explicitly flagged as flashbacks
- Structured chapter outlines: each chapter is broken into CBN (Chapter Beginning Node) → CPN (Chapter Pivot Nodes) → CEN (Chapter End Node), with concrete actions, not vague summaries
- Incremental setting write-back: new characters or factions introduced in planning are auto-added to the settings collection
3. End-to-End Chapter Production
The crown jewel. /webnovel-write 45 runs an autonomous six-step pipeline with three modes:
| Mode | Steps | Typical Time |
|---|---|---|
| Default | All 6 steps | 3-5 min/chapter |
--fast |
Lightweight review | ~2 min/chapter |
--minimal |
Format only, skip review | <1 min/chapter |
The pipeline is strict by design. The orchestrating skill is explicitly forbidden from "summarizing" what the agents would output — every agent must be invoked through the Agent() tool, and their JSON outputs must be persisted to disk before the next step proceeds. No shortcuts.
4. Six-Dimension Quality Review
The reviewer agent performs an integrated review across six dimensions:
| Dimension | Focus |
|---|---|
| High-point Checker | Density and quality of payoff moments |
| Consistency Checker | Power levels, geography, timeline |
| Pacing Checker | Quest/Fire/Constellation strand balance |
| OOC Checker | Character behavior fidelity |
| Continuity Checker | Scene transitions and narrative flow |
| Reader-Pull Checker | Hook strength, expectation management, retention |
Each issue must include cited evidence from the prose; the reviewer is not allowed to assert problems without quotes. Issues marked blocking: true halt the pipeline.
The system enforces a target strand weave ratio of 60% Quest (main plot) / 20% Fire (relationships) / 20% Constellation (worldbuilding expansion), with hard rules — Quest may not run more than 5 chapters straight, Fire may not vanish for more than 10, Constellation for more than 15.
5. Temporal State Queries
/webnovel-query answers questions across the entire history of your novel:
/webnovel-query 萧炎 # current state of a character
/webnovel-query foreshadowing # unresolved plants, urgency-sorted
/webnovel-query pacing # strand balance analysis
For temporal queries, it can answer "what was character X's state at chapter 30?" or "who was allied with whom at chapter 50?" by reconstructing state from the commit chain. This makes mid-novel continuity checks trivial.
6. Read-Only Web Dashboard
/webnovel-dashboard launches a local FastAPI + Vue.js dashboard showing:
- Chapter list with completion status and review scores
- Entity relationship graph (interactive)
- Foreshadowing tracker with urgency heat map
- Reader-pull curve over the novel
- Story Runtime health metrics
- Setting collection browser
The dashboard is strictly read-only and confined to the project root — it cannot modify any files.
7. Pattern Learning
/webnovel-learn "this chapter's crisis hook landed well" extracts and stores successful patterns to project_memory.json. On subsequent chapters, the context-agent automatically recalls these patterns, so the system gets better at writing your specific book the more you write.


