chinese-novelist-skill: Turning a Coding Agent into a Chinese Novel Writing Pipeline

The problem chinese-novelist-skill targets: unfinished long-form Chinese fiction

Most prompt collections for fiction writing fail at the same place. They produce a strong opening chapter and then drift, because nothing in the prompt carries state from chapter 3 into chapter 17. The README states the project's premise directly: the hardest part of writing a novel is finishing it, and the skill exists to attack that specific failure.

The target user is not a novelist looking for a better editor. It is someone who already has a coding agent open, wants a 10 to 50 chapter Chinese novel, and would rather answer eight structured questions than babysit a chat window for hours. The README frames the output as suitable for both standalone long fiction and serialized web novel publication, which is a meaningful constraint: web novel readers expect a hook at the end of every chapter, and the skill bakes that expectation into its rules rather than leaving it to taste.

What the project does not claim is literary quality. Nothing in the README promises publishable prose. It promises structure, continuity and completion, which are the three things a raw chat session reliably loses.

How the pipeline works: phases, JSON state and the writing plan file

The README lays out a numbered pipeline. Phase 0 loads user preferences from user-preferences.json and checks whether an unfinished project can be resumed. Phase 1 runs the three-layer questionnaire: Layer 1 has three mandatory questions (genre and premise, protagonist setup, core conflict), Layer 2 has five optional ones (worldbuilding, narrative point of view, theme, reader positioning, chapter count). Every question accepts a random roll, a skip, or a request to use defaults.

Phase 2 generates an outline using a seven-column chapter template, a character file, and a machine-readable writing plan in JSON. That JSON is the interesting design choice. It exists so that parallel writers can coordinate against a shared state file rather than each holding its own context. Phase 2.5 picks one of three modes: serial, sub-agent parallel, or Agent Teams. Phase 3 then runs unattended, and each chapter goes through pre-write analysis, drafting at 3000 to 5000 characters, a de-AI-flavor polish pass, a word count check, and a summary update. Phase 4 validates the whole manuscript for word count and coherence, rewriting failing chapters for up to three rounds.

The repository layout matches this: SKILL.md is the entry point, references/flows/ holds seven phase documents, references/guides/ holds eight craft documents including hook-techniques.md with 13 ending hook types, and scripts/ holds check_chapter_wordcount.py. The output directory structure in the README shows one folder per project named with a timestamp and title, containing the outline, character file, writing plan JSON, and one Markdown file per chapter.

Installing chinese-novelist-skill and running a first novel

The README gives two installation paths. The quick one uses the skills CLI:

npx skills add PenglongHuang/chinese-novelist-skill

After that, the documented trigger is a plain instruction to the agent:

使用 chinese-novelist 帮我写一部小说

The alternative, also documented, is to place the directory into Claude Code's skills folder manually:

~/.claude/skills/chinese-novelist/

The README also mentions installing through the Claude Code skill management interface. Once loaded, the first real use is the Phase 1 questionnaire. Answering the three Layer 1 questions is enough to proceed; the README shows that replying "跳过" or "都用默认" skips the optional Layer 2, and that each question offers a random option.

Expect the agent to print a planning summary before writing anything, with genre, protagonist, conflict, chapter count, point of view and tone, followed by the first five chapter titles and the main character list. The README shows a confirmation step where you reply "确认" to move on. Only after that does Phase 3 begin, and the README states it runs without further confirmation, so the agent will keep going while you are away from the keyboard. If your session dies mid-book, Phase 0 is documented to detect the unfinished project and offer to resume from the breakpoint.

Where the skill breaks down: word counts, coherence and the wrong kind of book

The 3000 to 5000 character target per chapter is a hard rule enforced by a script, and that is also the most obvious failure mode. A chapter that naturally wants to end at 2200 characters will be padded or rewritten until it clears the floor. The README describes a maximum of three rewrite rounds in Phase 4, which means a chapter that fails coherence checks three times is presumably accepted or surfaced as-is; the README does not document what happens after the third failed round. That gap is worth knowing before you start a 50 chapter project.

Coherence validation is the other soft spot. The README says the skill checks coherence, but it does not describe the mechanism, and there is no separate coherence script listed alongside check_chapter_wordcount.py. Word count is mechanically verifiable. Coherence, in practice, is a model judgement, which means it inherits whatever the underlying agent is bad at.

The wrong tool cases are clear enough. If you write in English or any language other than Chinese, the entire reference library, the hook taxonomy, the dialogue guidance and the word count thresholds are built for Chinese prose and will not transfer cleanly. If you want an interactive editor where you revise sentence by sentence, a phase pipeline that drafts whole chapters unattended is the opposite of what you want. And if you need deterministic output for a fixed manuscript, a system whose Phase 3 is explicitly described as fully automatic is not that.

Compared with a plain Claude Code writing prompt

The real alternative for most readers is not another repository. It is a long custom prompt pasted into the same agent, or a general creative writing skill. The difference is state handling.

A pasted prompt has no Phase 0 and no user-preferences.json, so every session starts from zero and the agent re-asks the same questions. It has no writing plan JSON, so parallel drafting is not possible without the writer manually stitching context. It has no resume detection, so an interrupted session means reconstructing where you were. And it has no validation phase, so nothing checks that chapter 14 hit the length target or that the hook at the end of chapter 9 was paid off.

What a plain prompt does better is flexibility. It will happily write a 4000 word short story, a poem, or a chapter in English. chinese-novelist-skill is opinionated about length, language, structure and the presence of a cliffhanger at every chapter end. That opinionation is the product. If your project does not fit the shape, the skill fights you.

Maintenance, licence and what a fork actually costs

The repository is MIT licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. That matters here because the skill is mostly text: if you fork it to change the hook taxonomy or the chapter template, you are editing Markdown under a permissive licence, and there is no legal obstacle to shipping your fork. This is not legal advice; read the LICENSE file for the actual terms.

The maintenance picture is concrete. The last push to the default branch was on 2026-09-06, and the repository is not archived. There are no retrieved releases, though the README references a v2.0 tag and a v1.0 to v2.0 upgrade pull request. The README also states that sponsorship directly affects update speed, which is an honest signal about the project's resourcing: this is a single-maintainer effort with an optional funding link, not a company-backed tool with a support contract.

Upgrade cost is low in the mechanical sense, since installation is copying a directory or running one npx command, but the v1.0 to v2.0 notes describe a substantial restructuring: detailed execution instructions moved out of SKILL.md into references/flows/, and the questionnaire changed from a single five-question flow to the layered model. If you forked v1.0 and customized SKILL.md, that reorganization is a merge you will have to redo by hand.

Editorial conclusion

Adopt chinese-novelist-skill if you already run Claude Code or a compatible coding agent and want a repeatable structure for long Chinese fiction, especially serialized web novel work where chapter hooks and word counts matter. Skip it if you need a standalone writing app, if you write in a language other than Chinese, or if you expect the skill itself to supply prose quality rather than the model behind it. Before committing to a long book, verify that your agent actually reads SKILL.md and the references/flows files, run one short 10-chapter project to confirm the Phase 4 validation loop fires, and check that scripts/check_chapter_wordcount.py executes in your environment. The repository is MIT licensed and was last pushed on 2026-09-06.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. PenglongHuang/chinese-novelist-skill on GitHub
4. README