PenglongHuang/chinese-novelist-skill

What it is

chinese-novelist-skill is a prompt-and-workflow skill for coding and personal agents that turns a brief interactive Q&A into a finished Chinese novel of roughly 10–50 chapters. It handles planning, drafting, polishing, and checking, so the writer sets premise and direction while the agent produces the manuscript.

The skill runs in phases: initialize and load saved preferences, ask three layers of questions (core positioning, optional deeper customization), produce character files, an outline, and a JSON writing plan for confirmation, then write chapters without asking again. Each chapter is targeted at 3,000–5,000 characters and passes through pre-write analysis, drafting, de-AI-ifying polish, and a word-count check. A final validation phase repairs failing chapters for up to three rounds; continuity and prose quality still need the author's review.

It is aimed at people who want long-form Chinese fiction, including serialized web novels, and who already work inside an agent such as Claude Code. Preferences (genre, point of view, style) are remembered across projects, and interrupted runs resume from saved progress. Three writing modes are offered: serial, parallel sub-agents, and Agent Teams for large works.

Getting started

Preparation: an agent environment that can install skills, and a configured model. The README is written around Claude Code.

1. Install the skill:

npx skills add PenglongHuang/chinese-novelist-skill

2. In the agent conversation, request a novel:

使用 chinese-novelist 帮我写一部小说

3. Answer the Layer 1 questions (genre, protagonist, core conflict), then skip or answer Layer 2 (world, viewpoint, theme, audience, chapter count). Every question offers a random option, and "skip" or "use defaults" is accepted.

4. Review the generated outline, character list, and chapter plan, then reply 确认. Pick a writing mode: serial (default), parallel sub-agents (faster), or Agent Teams (large works).

5. The agent writes all chapters and runs validation. Output lands in a project folder such as 20260412-143000-午夜列车/, containing the outline, character file, writing plan, style baseline, chapter summaries, and one Markdown file per chapter. Preferences are shared via user-preferences.json.

Full phase documentation is in references/flows/, and writing guides (hooks, dialogue, character building, plot structures) are in references/guides/.

Constraints and setup notes

• The skill depends on an agent runtime and model; the parallel modes specifically require sub-agent or Agent Teams support, which the README associates with Claude Code. Serial mode is the recommended default.
• Long novels consume substantial model quota. The README includes a sponsored API provider promotion, but the skill itself does not require that service.
• Automatic validation checks chapter files and word counts. Coherence and literary quality are judged by the agent and must still be reviewed by the author; this is not a guaranteed quality pass.
• Preferences persist in user-preferences.json, so repeated runs will reuse stored genre, viewpoint, and style choices unless changed.
• Licensed MIT.

HotRepo's take

The project appeared on the JavaScript monthly ranking for 2026-11-01 at position 15, with 1,298 stars gained that month against 4,513 total. That is a single monthly snapshot, so it supports the observation that the skill drew concentrated attention in one month; it does not show a sustained trend across periods or languages. What stands out technically is the choice to ship the writing craft itself — flow documents, hook techniques, chapter templates, quality checks — as inspectable skill files rather than as a hosted service, which suits agent users who want to adjust the pipeline. The README's own framing is realistic about limits: scripted checks cover file presence and length, while the judgment of continuity and prose remains with the author.

Project source and documentation