本文目录
- Why the chapter is the right unit
- The three inputs every chapter needs
- 1. A goal written as a change
- 2. The book’s stored material
- 3. The recap of the chapter before
- Writing the chapter
- What breaks a chapter generator
- A chapter workflow that survives a long book
- The part that makes it your book
- The bottom line
There is a tempting version of AI writing where you describe a book, press a button, and come back to a manuscript. It demos well. It fails at chapter twenty, and it fails in a way that is expensive to repair: the numbers have drifted, a story the author swears is true has acquired new details, a promise made in the introduction has quietly vanished.
The alternative is less cinematic and much more useful. You generate one chapter at a time, and you treat the chapter as a unit with a job. The model does the typing; the unit does the remembering.
Why the chapter is the right unit
A chapter is small enough to hold in one context and large enough to contain a change. That combination is what makes it work.
- Small enough: a chapter prompt can carry the goal, the relevant notes and the previous recap without turning into a summary of the whole book.
- Large enough: a chapter can deliver a lesson, a case, a reveal or a decision — the things a reader actually remembers.
- Checkable: you can read one chapter and answer a concrete question. Did the goal happen? Did the voice hold? Did anything contradict the facts?
A paragraph is too small to check for continuity. A whole book is too large to check at all. The chapter is the smallest unit that can be verified.
The three inputs every chapter needs
1. A goal written as a change
Most chapter plans are lists of topics, and topic lists produce books that inform without moving. Write each chapter’s goal as a before-and-after. Take a coach writing a shooting manual:
- Topic: “A chapter about the elbow.”
- Change: “By the end, a parent can spot a flat-set elbow in five seconds and fix it with three drills, one of which needs no equipment.”
The second version gives the model a starting state and an ending state, so the chapter has somewhere to land. It also gives you a way to audit the plan before writing it: read the goals in order and ask whether the reader’s ability rises. If chapter twelve leaves the reader where chapter four did, the middle will sag, and no prose style will save it.
2. The book’s stored material
“Keep it consistent” is an adjective. A fact is something a chapter can contradict.
There are two layers, and most tools only offer one.
The first layer is facts — short, labeled claims injected into every chapter:
| Label | Value |
|---|---|
| First gym | Huaihai Road, twelve courts, no heating, opened 2015 |
| Free-throw routine | Three dribbles, one breath, no talking |
| Coaching rule | Never teach the fadeaway before the free throw |
| Case | Liu, 16, fixed his elbow in six weeks with the wall drill |
| Timeline | The book covers years one to ten; chapter one is year one |
Three people or cases with five facts each is fifteen lines. It will do more for continuity than any instruction to “keep it consistent,” because it gives the chapter something to check against instead of an adjective to obey.
The second layer is the knowledge base — the long material your experience actually lives in. Training logs, client notes, drill diagrams, old lesson plans. In our editor you drop these in as markdown or text, grouped into a library, and mount that library to the book. While writing, you choose which library the chapter may read from: the chapter on injuries consults the injury notes, the chapter on tryouts consults the tryout files, and a different book’s notes never leak in.
Already covered handles the other half of continuity: points the book has explained. Later chapters are told not to re-teach them, which is how a book avoids explaining the same basic idea five times.
Voice & structure carries the rules that should never drift: tone, person, sentence rhythm, formatting. It is injected verbatim, so it behaves like a style sheet rather than a suggestion.
3. The recap of the chapter before
After a chapter is saved, the editor writes a short recap: what happened, what changed, which points were established. The next chapter is drafted against that recap instead of against the entire manuscript.
This is the part writers skip and then blame the model for forgetting.
The recap does two jobs. It keeps the context small, so the chapter can focus on its own goal instead of re-reading the book. And it gives you a weekly check: read the recaps in order and look for the three mistakes that compound fastest.
- A recap that says the book covered something it actually only mentioned.
- A recap that turns a side case into a main argument.
- A recap that resolves a question you meant to keep open.
Fixing a recap costs a sentence. Fixing the chapter costs a rewrite.
Writing the chapter
With the goal, the material and the recap in place, generation becomes a small decision rather than a gamble.
- Open the chapter. Read the goal and the recap. If you cannot state what changes by the end, the goal is still a topic.
- Generate. One chapter, at the length set for the book. The default is 1,200 words; anything from 400 to 4,000 is allowed.
- Read for the goal first. Prose problems are cheap to fix later. A chapter that does not deliver its change is not a prose problem.
- Check the facts. Did anyone contradict a fact or a case? If the chapter is right and the fact is wrong, fix the fact before continuing — the next chapter inherits it.
- Save, then read the recap. If the recap misstates the book, fix it now. It is the context for everything that follows.
- Move on. Do not polish chapter nine while chapter ten is unwritten. Polish is a second pass.
When a book has many empty units and the plan is solid, Write all fills the empty chapters in one run. It is the right tool for a first draft of a guide with a firm outline. It is the wrong tool for the chapter where the whole book turns. Write that one alone.
What breaks a chapter generator
- No goal. The model writes a chapter that is competent and inert. Nothing changes, so the next chapter has nothing to build on.
- No stored material. The details drift and the voice flattens into the average. Ten chapters later the drift is structural.
- No recap. The model re-explains what the reader already knows, or forgets a case it introduced two chapters ago.
- No stop condition. Continuing a chapter that has already delivered its lesson produces a coda, then another coda. If the change the goal asked for has happened, end the chapter.
- Editing the plan and the prose at once. Fix the plan first. Rewriting prose against a broken goal is wasted effort.
A chapter workflow that survives a long book
- Premise in one paragraph: who the book is for, what they will be able to do, what staying stuck costs them.
- Facts: five lines per key case, the terms of your trade, the timeline’s zero point.
- Knowledge base: upload the notes the book is allowed to use, grouped by subject, one library per book.
- Plan: one-line goals as changes, read in order for escalation.
- Draft three chapters, then stop and read them together. Fix the facts before continuing.
- Draft in blocks of five. Read the recaps after each block.
- Rewrite the first chapter last, when you know what the book is about.
At 1,200 words a chapter, a 25,000-word guide is about 21 chapters and a 70,000-word memoir is about 58. A monthly plan covers more chapters than one long book, so the constraint is not generation. It is the reading and the rewriting — exactly where your time should go.
The part that makes it your book
A chapter generator can produce a chapter that is technically correct and completely forgettable. That is what an average looks like when it is wearing your outline. The fix is the material: the drills you actually ran, the client who actually quit, the number in the ledger that surprised you.
That is why the knowledge base matters more than the prompt. The prompt tells the model what to do with the chapter. The knowledge tells it what only you know.
The bottom line
A chapter generator is not a button that produces books. It is a unit of work that can be planned, generated and checked. Give each chapter a goal written as a change, the material the book must never contradict, and a recap of the chapter before. Then one chapter at a time is not a limitation. It is the reason the book holds together at all.
Key takeaways
- 1Generate one chapter at a time: it is the smallest unit you can actually verify.
- 2Write chapter goals as changes, not topics, so the plan shows escalation.
- 3Keep two layers: short facts the chapter cannot contradict, and a knowledge base of your own notes.
- 4Mount one library per book so the chapter reads only the material that belongs to it.
- 5Read the recap after every chapter; it is the context for everything that follows.
常见问题
Can AI write a whole chapter?
Yes, at a length you set. In our editor chapters default to 1,200 words and can run from 400 to 4,000. The hard part is not length; it is giving the chapter a goal and the context to fulfill it without contradicting earlier chapters.
Why not generate the whole book at once?
Because consistency degrades with distance. A single pass has to hold the whole book in one context, and it cannot check itself against chapter three. Generating chapter by chapter lets each unit be written against stored material and a recap.
How long should a chapter be?
The subject and the pacing decide. 1,200 words is a workable default for both guides and fiction; practical chapters often run shorter. What matters is a consistent unit, because a steady length makes the plan and the reading time predictable.
What is a recap for?
The recap is a short summary written after a chapter is saved: what happened, what changed, which points were established. The next chapter is drafted against it instead of against the entire manuscript, which keeps the context small and focused.
参考资料
关于作者
每两周,一封有用的信
关于写作、发布,以及产品上线前后那些不体面的细节。没有热点合集,也没有增长黑客。





