Claude Code, Cursor, and Codex CLI can write novels from local folders. Compare their DIY rules, ledgers, and diffs with Maliang Writer's editor.
Coding agents are very capable: they read and write across files, make batch edits and run scripts, and technical authors can build very flexible writing workflows. The price is that the folders, rule files, setting ledger and version management all have to be designed and maintained yourself, and the interface is aimed at programmers. Maliang turns these into the cards, table of contents and editor that authors are used to.
Since 2025, more and more people have discovered that coding agents can write novels too: save each chapter as a Markdown file, create folders for settings, outlines and tracking, put a CLAUDE.md or AGENTS.md at the root spelling out the rules, and let Claude Code, Cursor or Codex CLI read the whole folder and write. GitHub already has quite a few open-source templates like this, such as webnovel-writer and oh-story-claudecode. The method does work and is very flexible. But in publicly shared practice, a few kinds of problems keep coming up: settings and requirements are forgotten after long conversations, style and characters drift, usage costs climb, Chinese chapter files sort out of order, and the barrier is not low for authors who don’t write code. In the end, coding agents are for programmers managing code; to write a novel with one, you first have to be half an engineer and build the book’s structure and rules.
On narrow screens, scroll sideways to see all columns.
| Dimension | Maliang Writer | Coding agents | Notes |
|---|---|---|---|
| Entry barrier | ✅ Open it on the web and use it, nothing to install | ⚠️ With the local-file approach you install, sign in and work in a terminal or code editor; some tools also have web or desktop entry points | Codex CLI can be installed with npm or Homebrew; Cursor is a desktop editor; Claude Code also has desktop and web versions |
| Structure of a book | ✅ Volumes, chapters, chapter outlines, characters, settings, storylines and documents are there out of the box | ⚠️ The author designs the folders, such as body text, outline, setting library and tracking files | |
| Writing rules | ✅ Type / to pick a writing method (continue writing, chapter-outline planning, removing AI-sounding language and so on), and one message can combine several | ✅ Write your own rules in CLAUDE.md, AGENTS.md or .cursor/rules | Codex CLI’s AGENTS.md has a combined default cap of 32 KiB |
| Ledger of character states and foreshadowing | ✅ Story archives are organized after a chapter is written; anything that doesn’t match is listed as “pending confirmation”; can be turned off | ⚠️ The author or AI maintains ledger files in a format the author designs | |
| Batch edits across files | ✅ AI can edit characters, settings and the body text of several chapters as you ask, and the changes can be checked and undone | ✅ A strength: it can check and batch-edit across hundreds of files, and can also run scripts | |
| Revision comparison | ✅ Red deletions and green additions on the original in the novel editor, each of which can be kept or restored | ✅ Cursor has a live diff view; Claude Code can show diffs in the VS Code plugin and the desktop app; Codex CLI shows a text diff in the terminal | |
| Undo | ✅ Undo the current round, branch from before the round, or edit and resend | ✅ Claude Code has automatic checkpoints plus /rewind; Cursor checkpoints are stored locally and the official site recommends Git for long-term versions; Codex CLI’s official advice is to make a Git snapshot before and after tasks | |
| Chinese chapter ordering and reading | ✅ Shown in the table of contents in volume-and-chapter order, with the body text typeset in the editor for reading | ⚠️ Kept as files, so the author has to deal with ordering and merging | Some public practice mentions Chinese chapter names coming out of order when sorted alphabetically |
| Availability in Mainland China | ✅ Chinese interface, accessed directly in the browser | ⚠️ Claude Code: Anthropic’s official supported-region list does not include Mainland China; Codex comes with ChatGPT plans, and ChatGPT’s official supported-region list does not include Mainland China; Cursor’s official site says some model providers have regional restrictions | |
| Cost | ✅ Credits are deducted by usage; free sign-up to try | ✅ Claude Code is included from Claude Pro at $20/month; Cursor Hobby is free, Individual $20/month; Codex is included in all ChatGPT plans (checked against the official pricing pages in 2026-10) |
✅ Confirmed in official material or in the product ⚠️ Partly supported, or not stated in official material ❌ Officially not supported · Competitor information comes from their public material.
Volumes, chapters, chapter outlines, characters, settings, foreshadowing, storylines and documents are already sorted into categories. What you see are cards and a table of contents; there are no folders to design and no rule files to write.
The “character timeline and foreshadowing ledger” that authors in public practice maintain themselves is done in Maliang by the story-archive organizing after a chapter is written: character states, relationships, locations and foreshadowing, with anything that doesn’t match listed as “pending confirmation”.
When AI changes the body text, the changes are marked right on the typeset original, and each can be kept or restored, with no code-style file comparison to read.
Coding agents are very capable: Claude Code understands a whole project and runs tasks across files, with automatic checkpoints to roll back at any time; Cursor has a detailed rules system, live diffs and models from many vendors to choose from; Codex CLI is open source and can look at files, edit files and run commands in the terminal, with system-level isolation and approvals. Technical authors can do many things with them that writing software can’t: write scripts to count word frequency across the whole book, batch-rename characters, customize check workflows and manage every draft with Git.
Every product suits some scenarios; the advice here covers only one: serialized long-form fiction in Chinese.
For this page, record in particular how much work the author has to do to prepare folders, writing rules, material updates and revision checks. Don't compare only the prose of a single generation, and don't assume in advance that folder-based tools must be harder to use.
Below is a same-task trial checklist, not a finished comparison experiment. You can use the original teaching material “Jiehuo City” or a manuscript you have the right to handle; record the preparation, the output and the author’s edits for each tool.
Give both tools the same protagonist, core rules and reading preferences; compare the opening directions first, then organize the master outline, the first volume outline and the outlines for the first three chapters.
Check: Can it tell the roles of the three outline levels apart? Can the author confirm a plan before it continues? Were settings that hadn't been settled yet quietly written in as fact?
Use the same book-opening kitUse the same old manuscript you have the right to handle, specify how the last chapter ends and the goal of the next chapter; try writing one scene first, without changing the previous twenty chapters.
Check: What material did you prepare? Did character positions, items and open threads carry over? How are candidate drafts told apart from adopted ones?
See the import-and-continue task checklistPick a real rule that affects earlier chapters; first ask it to list the affected chapters and the evidence, then let the author choose the scope of the change.
Check: Does it say which parts it did not check? Can you review the changes first? Did it wrongly edit text outside the task or write an unadopted plan into the material?
See how to change a settingActually try keeping the text, the order of volumes and chapters, character profiles and world rules, and check whether the output is suitable for further editing; list what cannot be taken along.
Check: Are the format and order recognizable? What material is inside the platform account, what is in local files, and what is in the export? Do not infer usability from a marketing page alone.
Copy the task and result record cardRecord the models, plans or account conditions you actually used, and the time spent preparing materials and editing the text. Leave blank anything you have not tried or cannot confirm; do not let a feature name or a single generation stand in for the full experience.
Yes. The common approach is one Markdown file per chapter, separate settings, outline and tracking folders, rules in CLAUDE.md, and Claude Code reading and writing the whole folder. Note that Anthropic’s official supported-region list does not include Mainland China.
Cursor lets you reference files with @, set rules and see live diffs, and technical authors find it very flexible. Authors in public practice also mention problems such as forgetting settings after long conversations and Chinese chapters sorting out of order, which you have to solve by designing your own folders and rules.
You need to install Codex CLI, sign in with a ChatGPT account or use an OpenAI API key, and write AGENTS.md rules in the project. The official CLI docs recommend making Git snapshots before and after tasks so you can roll back.
A common split in public templates is four kinds: body text, outline (master outline, volume outlines, chapter outlines), setting library and tracking (status and foreshadowing ledgers), with a rule file at the root and Git for versions. In Maliang these categories are there out of the box, and the archive is organized automatically after a chapter is written.
Generally style requirements, hard character settings, prohibitions and which files to read for each chapter. The more rules there are, the more they need trimming from time to time, or they get ignored or conflict with each other. Maliang uses writing methods in place of rule files; type / to pick one.
You can learn, but you first need to get comfortable with the terminal, files and folders, and also be able to read file comparisons. If you just want to concentrate on writing the novel, web-based writing software will save you a lot of trouble.
Coding agents are mostly included in subscriptions, such as Claude Pro from $20/month and Cursor Individual at $20/month (checked 2026-10); long conversations and batch operations use more quota. Maliang deducts credits by usage, with each round’s cost shown under the reply, and you can sign up for free and compare.
You can paste the setting files and chapter content to Maliang’s AI and have it organize them into character cards, setting cards and a chapter table of contents; check them and keep writing.
Claude Code can write novels, and it is very capable: it reads and writes a whole book’s folder, checks continuity across files and makes batch edits. In public practice, authors design the folders, write CLAUDE.md, maintain a setting ledger and work in a terminal or editor plugin, which suits authors who are comfortable on the command line. Maliang builds the table of contents, character cards, story archives and revision comparison into a web editor, and the author starts writing from the chat.
Read the comparison →Cursor can write novels: the editor lets you see files and live diffs, the Rules system is detailed, and the choice of models is wide. In public practice, authors build their own chapter folders, write rules and maintain settings and a ledger, and once there are many chapters they also have to deal with file ordering and merging. Maliang builds the volume-and-chapter table of contents, character cards, story archives and revision comparison into a web editor.
Read the comparison →Codex CLI is open source and runs in a local terminal; it reads and writes a whole project and runs scripts, with system-level isolation and approvals. Writing a novel with it, the author builds the folders, writes AGENTS.md, maintains a ledger and uses Git for versions, and views changes as a text diff in the terminal. Maliang builds these into a web editor: a volume-and-chapter table of contents, character cards, story archives and revision comparison you review one by one.
Read the comparison →Folders plus Markdown are flexible and transparent, with the files in your own hands, which suits authors who like to define their own formats; the drawback is that once there are many files, which file is the latest and which other places need syncing after one edit are all yours to remember. Character cards plus story archives fix each character and setting as a card and organize states automatically after a chapter is written, which suits authors who don’t want to maintain files. Maliang has both: characters and settings are cards, the master outline is stored separately, and style samples and book-opening material are documents.
Read the comparison →To start with the kinds of novel-writing software and what each is good at, read How to choose an AI novel writing tool.
After signing up, create a book or import existing chapters, tell the AI which chapter to continue, then look at what it wrote, what it changed and what it recorded in the archive, and decide whether to move your long novel over.