WebNovel Writer — How AI Writes 2M+ Word Novels Without Forgetting

Writing a 2-million-word web novel with AI sounds impossible — Claude’s context window tops out at a fraction of that, and across hundreds of chapters the AI inevitably forgets character details, contradicts earlier plot, or hallucinates facts. WebNovel Writer solves this with an architecture worth studying even if you never write fiction: RAG-enhanced context retrieval, entity graph tracking, and narrative momentum metrics. The same patterns apply to any long-context AI application.

Why This Matters Beyond Novels

If you care about long-context AI applications — research writing, legal document drafting, code bases that span months — this tool’s architecture solves problems you’ll face too:

The Architecture

┌──────────────────────────────────────────────────────┐
│            WebNovel Writer Workflow                   │
│                                                       │
│  ┌──────────┐     ┌──────────┐     ┌──────────┐     │
│  │ PLAN      │────▶│ WRITE    │────▶│ REVIEW   │     │
│  │          │     │          │     │          │     │
│  │/webnovel-│     │/webnovel-│     │/webnovel-│     │
│  │ plan 1   │     │ write 1  │     │ review   │     │
│  │          │     │          │     │ 1-5      │     │
│  └────┬─────┘     └────┬─────┘     └────┬─────┘     │
│       │                │                │            │
│       ▼                ▼                ▼            │
│  ┌──────────────────────────────────────────┐        │
│  │       Persistent Story Layer              │        │
│  │                                            │        │
│  │  ┌────────────┐  ┌──────────┐  ┌────────┐│        │
│  │  │ Entity     │  │ Chapter   │  │ Plot   ││        │
│  │  │ Graph      │  │ Summaries │  │ Threads││        │
│  │  │(characters,│  │(searchable│  │(active │        │
│  │  │ settings,  │  │ via RAG)  │  │ arcs)  ││        │
│  │  │ relations) │  │           │  │        ││        │
│  │  └────────────┘  └──────────┘  └────────┘│        │
│  │                                            │        │
│  │  ┌────────────────────────────────────┐   │        │
│  │  │ Embeddings + Reranking (Qwen+Jina) │   │        │
│  │  └────────────────────────────────────┘   │        │
│  └──────────────────────────────────────────┘        │
└──────────────────────────────────────────────────────┘

Three-Stage Workflow

1. Plan — Outline structure, define characters, set plot threads
2. Write — For each chapter, retrieve relevant context via RAG, generate with entity graph constraints
3. Review — Check consistency, track momentum metrics, flag contradictions

Key Technical Insights

RAG-over-chapters instead of giant context Don’t try to stuff the whole novel into the prompt. For each new chapter, semantic search retrieves:

• Previous chapters relevant to the current scene
• Character profiles for people appearing in this chapter
• Active plot threads that need progression
• Any “narrative debt” that should be paid off

Entity graph as source of truth Characters have structured profiles: name, physical description, backstory, relationships, current state. Claude writes against this graph, not against fuzzy recollection. When the graph says “Alice is afraid of water,” Alice stays afraid of water in chapter 200.

Reading momentum as a metric The review agent tracks narrative momentum:

• Hooks: cliffhangers, unresolved tension
• Cool-points: memorable scenes, character moments
• Narrative debt: promises made that haven’t been delivered

This is essentially a quality metric for fiction — something that usually requires human editors.

Installation

# Via Claude Code marketplace
/plugin marketplace add lingfengQAQ/webnovel-writer
/plugin install webnovel-writer

# Install Python dependencies
pip install -r requirements.txt

# Initialize project
/webnovel-init "My Novel Name"

# Configure embeddings (example: Qwen + Jina)
# Edit config.yaml with your API keys

# Start writing
/webnovel-plan 1        # Plan chapter 1
/webnovel-write 1       # Write it
/webnovel-review 1-5    # Review chapters 1-5

Optional: /webnovel-dashboard launches a read-only visual dashboard showing project state, entity relationships, and chapter progression.

Model Flexibility

Different agents can use different Claude models based on task complexity:

This is a harness pattern — match intelligence to stage — same principle as the planning/execution/verification phases in coding agents.

What This Teaches About Long-Context AI

Even if you never write fiction, the patterns here apply to any long-running AI project:

1. Don’t fight the context window — bypass it with RAG — Retrieve only what’s relevant, not everything
2. Structured metadata beats unstructured memory — Entity graphs, plot threads, and typed profiles are more reliable than “remember what we talked about”
3. Separate concerns across phases — Plan/write/review instead of one giant prompt
4. Define quality metrics, not just correctness — “Is this chapter engaging?” is the fiction equivalent of “is this code maintainable?” — measure it explicitly
5. Match model to task — Opus for the hard parts, Haiku for the easy parts

These are the same principles behind harness engineering and Karpathy’s LLM Knowledge Bases. The web novel is just a concrete application of the general pattern.

How LearnAI Team Could Use This

• Long-context curriculum — use the project as a concrete example of RAG, entity graphs, and structured memory for multi-session AI work.
• Agent workflow demos — teach plan/write/review loops as a reusable pattern for research, documentation, and coding agents.
• Creative AI labs — let students compare naive long-form generation against retrieval-backed, metadata-constrained generation.
• Quality evaluation practice — adapt the momentum and consistency checks into rubrics for evaluating AI-generated artifacts.

Real-World Use Cases

• Long-form fiction production — maintain continuity across hundreds of chapters and millions of words.
• Research report drafting — retrieve relevant prior notes while preserving entity, claim, and citation consistency.
• Legal or policy drafting — track obligations, actors, definitions, and contradictions across large document sets.
• Game or worldbuilding bibles — keep characters, locations, factions, timelines, and unresolved plot threads consistent.

Case Study: 《探花书房》 — Full Novel in One Session

I used the webnovel-writer skill to build a complete 48,000-word literary novella from concept to polished PDF ebook, mostly overnight.

Setup

The seed was two WeChat screenshots — a Doubao AI prediction about a friend’s future girlfriend: she’d be named Lin Wan, from Suzhou, they’d meet at a bookshop called 探花书房 on Nanshizi Street in late autumn 2026 over a Pingjiang Road postcard.

The challenge: the skill is designed for web novel tropes (systems, power levels, face-slapping). I needed to adapt it for literary fiction with no golden finger, no antagonist, no power system — just two quiet people slowly falling for each other in Suzhou.

What Actually Happened

Architecture Lessons

• Literary fiction stretches the framework: “爽点规划” and “反派分层” are wasted on literary fiction, but the core pipeline (context → draft → review → polish → data) works for any genre
• OOC checker caught real issues: Flagged Lin Wan being too proactive in Ch1 and Qing’s emotional progression exceeding the “noticed, not attracted” boundary — both valid
• Overnight writing via RemoteTrigger: The pipeline is autonomous enough for 11 unattended chapters
• Anti-AI check is critical for literary prose: Template-ness that’s acceptable in web novels is fatal in literary fiction

Final Numbers

Project: weihaoqu/tanhua-bookshop

Links

• GitHub: lingfengQAQ/webnovel-writer
• Case Study: weihaoqu/tanhua-bookshop
• Version: v5.5.4