AI Agents Writing Novels: Rules, Skills & 5 Long-Form Gaps

In this article 9 sections

The Conclusion First: The Working Style Is Similar, and Writing Long Form Means Filling Five Gaps

The past year or two has brought many “agents”: Tencent’s WorkBuddy, ByteDance’s Coze, Manus, Kimi Agent, the open-source OpenClaw, and Claude Code, Cursor and Codex CLI, which came out of writing code. Their marketing differs, but checking against public documentation, quite a few of these products work in similar ways:

1. Multi-step execution: the agent reads your request, breaks it into steps on its own, reads and writes files, looks up material, sees the result, and then decides the next step;
2. A project folder (or project, workspace, library): the material to read and the output live here;
3. Rules files and skills: a standing rules file (Codex uses AGENTS.md, Claude Code uses CLAUDE.md), plus skills used on demand, each skill being a set of instructions, sometimes with scripts.

This convergence is visible in the public documentation: WorkBuddy’s project memory file is compatible with AGENTS.md (Projects), and its skills market is compatible with OpenClaw’s skills ecosystem (Skills Market); in Coze projects, sub-conversations share files and agents.md (Coze Projects); OpenClaw’s and Coze’s memory both use Markdown files (OpenClaw memory, Coze memory). Whether products such as Manus and Kimi Agent use the same way of organizing is not fully explained in official materials, so we suggest confirming this on a trial.

So rather than asking which agent writes novels better, ask: to write a book of several hundred thousand characters with this working style, what jobs does someone have to do? We go through them one by one below, and at the end of each section explain what Maliang Writer does.

1. Where to Put the Material: Build the Book’s Structure First

Products organize material differently, with local workspaces, project files, and libraries, and the scope of use needs to be checked for each. When you manage novel material in folders, a character card, chapter 37, and a setting that was voided in volume three are each just a file, and which is which and which one prevails must be labeled clearly by the author. We suggest setting it up like this:

• Keep settings, characters, outline and body text separate;
• One file per chapter of body text, with the file name carrying a sequence number;
• A “current state” field in each character card;
• A separate foreshadowing list.

With a clear directory and explicit state labels, it is easier for the agent to find the material it should read; on a trial run, also check which content it actually cited. For the specific folder structure and how to write the rules file, see Novel Folder and CLAUDE.md Rules for Writing with Coding Tools.

Maliang Writer builds this layer of structure ready-made: volumes, chapters, chapter outlines, characters, settings, foreshadowing and storylines are categorized out of the box, with the table of contents and cards on the left.

2. What to Read for Each Chapter: Context Is Limited

There is a ceiling on what the model can see at each step. For example, the Kimi Help Center states that Agent context is 128K tokens (about 80,000–100,000 Chinese characters, Capabilities and limits); WorkBuddy’s official documentation states a default context of 200K, with GLM-5.3 able to be set to 1M, while cautioning that raising it to 1M “won’t make the model smarter, it only multiplies each round’s consumption” (model notes). Once a book passes several hundred thousand characters, you usually have to pick what to read chapter by chapter.

Before each chapter, check that the relevant material is complete and up to date, and after writing still check that body text and settings are consistent. When writing chapter 120, you usually need:

• This chapter’s chapter outline;
• The cards of characters appearing, especially their current state;
• The body text of the last two or three chapters;
• The foreshadowing and settings related to this chapter;
• A style sample.

Products supply material in different ways: some read per the rules file, some retrieve from a library and inject it (WorkBuddy project material is injected into tasks in RAG form), and some rely on memory. On a trial run, check whether what it read for this chapter is the latest settings, and whether it missed what just happened in the previous chapter. Long conversations have another problem: early content gets compressed or forgotten. The Kimi Help Center has warned that when a conversation has too many turns (for example over 50), early content may be forgotten, especially in long-form novel creation.

3. Who Organizes After Writing: Chapter-End Material Must Be Updated Promptly

Once a chapter is written, the book’s “state” has changed: someone is wounded, someone has learned a secret, an artifact has changed owners, a foreshadowing thread has been paid off. If these changes aren’t written back to the character cards and the foreshadowing list, the next chapter reads outdated information.

Some general agents offer session-memory organization (for example, WorkBuddy organizes the day’s sessions every night, Memory), but a dedicated process for updating novel character states and foreshadowing chapter by chapter is not described in official materials, so we suggest confirming it on a trial. A common practice is to add a line to the rules file, “after finishing each chapter, update the character states and the foreshadowing list”, or to write a skill to do it. On a trial run you can check:

• After each chapter is written, whether the character states and the foreshadowing list were updated;
• Whether newly added records have a basis in the body text;
• How conflicts between new information and old settings are handled, whether the changes preserve the original text, and whether you can see where things were changed.

Maliang Writer makes this step a fixed process: after a chapter is written and this round ends, the archivist organizes the chapter’s character states, relationships, locations, items and foreshadowing; places that don’t match are listed as “pending confirmation” for the author to check and handle.

4. How to Review Changes: Which Sentences Changed Among Dozens of Files

An agent revising a draft often touches several files at once. Tools that came out of writing code have ready-made ways: Claude Code and Cursor have checkpoints, Codex CLI officially recommends using Git for checkpoints, and changes can be compared and rolled back per file. General agents each do it differently; for example, in WorkBuddy’s library, the AI’s edits to Markdown documents don’t overwrite the original and are presented in a pending-confirmation state (Library).

For someone writing a novel, what needs viewing isn’t file differences but “how did this passage read before, and how does it read now”. Maliang Writer puts the AI’s edits inside the typeset body text, with red deletions and green additions, to keep or revert one by one; if you’re unhappy with the whole round you can undo this round, or branch from before this round.

5. When to Put Old Settings Away: Once You Reach Volume Three

By volume three, some of the supporting characters, places and foreshadowing from the first two volumes have left the current plot, but the material is still there. If it’s included when reading the character table or retrieving material, it takes up context and may also get exited characters written back in.

When writing a book with a general agent, the common practice is for the author to move old material into an archive directory themselves, or to state in the rules which items are no longer referenced, and to check what material is actually cited when writing new chapters. Maliang Writer handles this by volume: when you reach a new volume, characters, places and foreshadowing that appeared only in earlier volumes are moved into the archive and no longer referenced when writing new chapters; when a character returns, they are put back into active use. For the specific approach, see Writing a New Volume: What to Do with Old Settings.

Five Trial-Run Checks for Writing Long Form with a General Agent

If you’re already writing a book with WorkBuddy, Coze, OpenClaw or a coding agent, you can check against the five items below. Whether a given product supports rules files and custom skills is subject to its official documentation, and where nothing is stated we suggest confirming on a trial:

1. Material structure: whether settings, characters, outline and body text can be kept separate, and whether body text can be kept as one file per chapter.
2. Rule settings: whether you can state which material to read before each chapter, style requirements, and which settings must not be changed.
3. Chapter-end organization: after each chapter is written, how character states and the foreshadowing list are updated, and how new records and conflicts are handed to you for review.
4. Change review: whether the product’s built-in versions, pending-confirmation features or Git let you see clearly which files and passages were changed.
5. Archiving by volume: after finishing a volume, whether exited characters and paid-off foreshadowing can be moved out of the commonly used material.

These five items can serve as a checklist for trial runs, and also show you clearly which jobs the author has to take on themselves.

Choosing a Tool by Writing Workflow

For comparisons of the various tools, see Six Kinds of Novel-Writing Software and General Agent vs Novel-Writing Software: How to Choose.

FAQ

Q: How does an agent in novel-writing software differ from a general agent?
Both let the AI read and write a book’s material and complete tasks in multiple steps. Taking Maliang Writer as an example, the volume and chapter directory, character cards, the archivist’s chapter-end organization and pending confirmation, red deletions and green additions in body text to keep or revert one by one, undo this round, branching, and archiving and returns by volume are all ready-made features, and content listed as pending confirmation is still checked and handled by the author.

Q: If I install a novel skill into a general agent, is it the same?
A skill can solve part of “how to write”, such as the book-start workflow, chapter outline format and style requirements. But who updates character states, how conflicts are handled, and how changes are reviewed still depend on whether the product has a corresponding interface and process. Look at a skill’s scripts before installing it; WorkBuddy’s official documentation has also warned that unofficial skills carry injection risks.

Q: If I set context to 1M, can the whole book fit in?
More content fits, but each round’s consumption also rises. Taking WorkBuddy as an example, the official notes say GLM-5.3 can be set to 1M, and caution that it “won’t make the model smarter, it only multiplies each round’s consumption”. Whether the whole book fits at once also depends on other context usage. Picking out the material you need chapter by chapter is also a way to control consumption.

Q: What model does Maliang Writer use?
You can choose the model, and credits are deducted according to the chosen model and usage; go by what the product displays for specifics.