# chinese-novelist

> A coding-agent skill that plans and writes a complete Chinese novel, chapter by chapter

- Category: AI Productivity & Writing / Writing & Marketing
- Language: Python
- Licence: MIT
- Repository: https://github.com/PenglongHuang/chinese-novelist-skill
- Canonical page: https://ai-tldr.dev/tools/chinese-novelist/
- Updated: 2026-10-20

## In short

chinese-novelist is an Agent Skill for long-form Chinese fiction. It sits in the Writing & Marketing category of the AI Productivity & Writing tools and is open-source, written in Python and licensed MIT. To get started: install the skill.

## Overview

chinese-novelist is an Agent Skill for long-form Chinese fiction. Installed into Claude Code or another coding agent that reads `SKILL.md`, it turns "使用 chinese-novelist 帮我写一部小说" ("use chinese-novelist to write me a novel") into a staged project: a short interview about genre, protagonists and core conflict, a planning pass that produces an outline, character files and a machine-readable writing plan, and then an unattended run that writes every chapter to its own Markdown file. The repository describes the target as a complete 10–50 chapter novel, with 3,000–5,000 characters per chapter.

The work is split into numbered phases, each with its own instruction file under `references/flows/`. Phase 0 loads the writer's saved preferences and checks for an unfinished project to resume. Phase 1 asks three required questions, then optional deeper ones (world, point of view, theme, audience, chapter count) and proposes a title; every question can be skipped or randomised. Phase 2 shows the plan for confirmation, and the writer picks a mode: one agent writing chapter after chapter, parallel sub-agents writing in batches, or Claude Code Agent Teams for longer books.

Phase 3 writes without further prompts. For each chapter the skill analyses what the chapter needs, drafts it, runs a polishing pass meant to strip stock "AI-flavoured" phrasing, checks the length and appends a summary to a running chapter digest that later chapters read for continuity. Phase 4 then checks the whole manuscript for word count and coherence and rewrites failing chapters, up to three rounds. Version 2.3 added a style-baseline file (a style fingerprint, reference paragraphs and a list of phrasings to avoid) and a glossary to stop names and terms drifting between chapters; v2.3.1 added a fact-ledger rule and a quantified AI-flavour check.

Everything the skill produces is plain files in a dated project folder (outline, character profiles, writing plan JSON, style baseline, chapter summaries and one file per chapter), so a draft can be read, edited or continued in any editor. A small Python script, `scripts/check_chapter_wordcount.py`, does the length checks. The README and prompts are in Simplified Chinese, and the skill writes Chinese fiction only.

## What it does

- Layered interview: three required questions (genre, protagonists, core conflict), optional deeper customisation and a generated title, with skip or random answers at every step
- Planning artefacts before any prose: a seven-column chapter outline, character profiles and a JSON writing plan that tracks each chapter's status, word count and retries
- Three writing modes: one agent writing chapter after chapter, batched parallel sub-agents coordinated through the plan file, or Claude Code Agent Teams
- Per-chapter loop of pre-writing analysis, a 3,000–5,000 character draft, a de-"AI-flavour" polish, a length check and a running summary for continuity
- Automatic final validation of length and coherence, rewriting failing chapters for up to three rounds
- Cross-session memory of your preferred genres, viewpoint, tone and chapter count, plus detection and resumption of unfinished projects

## Getting started

chinese-novelist is a skill folder, not an app: it runs inside a coding agent that reads `SKILL.md` (the README targets Claude Code). Prompts and output are Simplified Chinese.

1. Install the skill

   The skills CLI fetches the repository and asks which agents to add it to.

```bash
npx skills add PenglongHuang/chinese-novelist-skill
```

2. Or copy it into Claude Code's skills directory

   The README's manual route is to place the folder at `~/.claude/skills/chinese-novelist/`.

```bash
git clone https://github.com/PenglongHuang/chinese-novelist-skill ~/.claude/skills/chinese-novelist
```

3. Ask for a novel

   In a new session, give the trigger sentence from the README ("use chinese-novelist to write me a novel"). The skill loads your saved preferences and offers to resume any unfinished project.

```text
使用 chinese-novelist 帮我写一部小说
```

4. Answer the interview

   Pick a genre, protagonist set-up and core conflict, then continue to the optional questions or say "跳过" (skip) or "都用默认" (use the defaults).

5. Confirm the plan and choose a writing mode

   Reply "确认" (confirm) to the outline and character summary, then choose serial, parallel sub-agent or Agent Teams writing. From here the skill writes without asking again.

6. Read the manuscript

   Output lands in a dated project folder next to a shared `user-preferences.json`: `00-人物档案.md` (characters), `01-大纲.md` (outline), `02-写作计划.json` (plan), `03-文风基准.md` (style baseline), `99-章节摘要.md` (chapter digest) and one `第NN章-<title>.md` per chapter.

## When to use it

- Drafting a complete Chinese web novel (mystery, romance, fantasy, sci-fi, wuxia, historical) from a one-line idea
- Producing an outline, cast and chapter plan to revise by hand before any prose is written
- Running a long serial unattended with parallel sub-agents, then resuming it in a later session
- Studying how a multi-phase writing workflow can be encoded as a skill with plain Markdown and JSON state

## Version history

- 2026-10-20 v2.3.1: Added a fact-ledger rule to keep facts consistent across chapters and a quantified check for AI-flavoured prose, and centralised the chapter checklist.