Interactive eBook BuilderInteractive eBook Builder
$59 · unlimited access, unlimited eBooksGet it for $59

· By · 11 min read

How to Write Prompts for Long-Form Content

promptsaiwritingebook

Prompts for long-form content work best when you write one chapter or section at a time, put the source material first, keep standing rules in a reusable instruction, and pass a short brief from one stage to the next instead of asking for the whole manuscript in a single request. A novel, an eBook, or a workbook is not one task. It is a sequence of tasks that have to agree with each other: the same reader, the same promises, the same terms, and no invented facts that later chapters will have to unwind.

Vendors are explicit about the limits and the placement. OpenAI’s prompt engineering guide explains that a model’s memory for a request is a context window measured in tokens, that this window includes the text you send and the text the model returns, and that models vary widely in how much they can hold. It also suggests keeping stable, repeated instructions in a consistent place, which is useful when the same book policy appears in every chapter request. Anthropic’s prompting best practices say that for long documents, you should put the long-form data near the top and the query last, and they report that this can improve response quality by up to 30 percent in tests, especially with complex, multi-document inputs. They also recommend asking the model to quote relevant passages before it carries out a task on long sources. Google’s Gemini 3 guidance gives the same placement rule: supply large amounts of context first, and put the specific instructions or questions at the very end. Google also describes breaking complex work into components and chaining prompts so that the output of one step becomes the input of the next.

For eBook and workbook projects, Interactive eBook Builder is a $59 one-time Notion AI template with Spec Builder, Drafter, Polisher and SEO BOT already set up, so the book can move through planning, drafting and editorial refinement without asking one prompt to do all three. Notion AI access and usage are separate from the purchase. The examples below are teaching prompts written for this article, not paid Builder prompts and not recorded model outputs. This article is about prompting a long project, not about how long a single prompt should be, and not about a vendor-by-vendor comparison of Gemini and Claude. Those guides remain useful companions.

Why one giant request fails

A prompt such as “Write my 12-chapter eBook about freelance boundaries” asks the model to invent structure, voice, examples, exercises and transitions in one pass, then hold all of that consistently. Even when a model has a large context window, you still have to review the result, and a 30,000-word miss is much harder to repair than a 300-word one. You also spend the window on a huge reply instead of on the notes, the outline and the previous chapter brief that would have kept the next section honest.

Google’s “token” reminder is a practical one, even as a rough measure: their prompt-design page notes that a token is approximately four characters, and that 100 tokens correspond to roughly 60 to 80 words. Exact tokenization differs by model, so treat that as an approximation, not a budget you can count on to the word. The useful lesson is simpler: long-form prompting is a resource problem as well as a writing problem. Put the window toward the sources and the current section, not toward a request for the entire manuscript.

Keep standing rules, and write one section at a time

Write a standing brief once, and reuse it. That brief is the long-form version of a system prompt: audience, voice, format, source rules, and what to do when a fact is missing. Then write a request for one unit of work—an outline, a section, an exercise, or an edit.

A standing brief can look like this:

You are helping the author of a short eBook for new freelance designers.
Use plain, calm language. Stay inside the notes supplied in each request.
If a fact, rate or legal rule is missing, write "Need from me" instead of
inventing it. Do not add income claims or "industry standards."

Teaching sections: 250 to 350 words, then one short reusable example.
Workbook pages: a one-sentence instruction, fill-in lines, and one backup
question. Keep terms consistent with the glossary in the notes.

A section request can look like this:

## Previously decided
- Chapter job: help the reader state a revision limit in the kickoff email.
- Approved outline headings: [paste four or five headings]
- Term to keep: "revision limit," not "scope creep policy."

## Notes for this section only
[paste the notes]

## Request
Draft the section "Write the sentence you will actually send," 250 to 350
words, then the one-line email with a blank for the number. Do not rewrite
the rest of the chapter.

The standing brief stays stable. The “previously decided” block is how you give the project a memory without pasting the whole book back in. If a later section needs the actual previous text, paste that section, not the manuscript.

Put the sources first, and quote before you transform

When the input is a transcript, a pile of research, or two conflicting outlines, follow the vendor placement rule: documents first, question last. Anthropic’s quote-then-task pattern is especially useful here. Ask the model to pull the relevant lines from the notes, then write the section only from those lines. If a claim cannot be quoted, it should not appear as a fact in the draft.

That habit also protects you at the end of a long project, when a model might quietly “help” by filling gaps. A chapter that sounds complete can still have invented numbers. The quote step makes the invention easier to catch.

For documents, use consistent delimiters so the sources stay distinct. Our guide to using delimiters in prompts walks through Markdown headings and XML-style tags.

Pass a brief between stages instead of starting over

Long-form work is a chain, even if you never use that word. Google describes chained prompts as a sequence in which the output of one prompt becomes the input of the next. Our prompt chaining guide applies that idea to planning, drafting and editing a book. The prompting habit is the same: do not throw the last stage away, and do not paste it forward unedited.

A clean handoff is short and approved:

If you hand the model a messy first draft and say “continue,” you are asking it to treat unapproved text as canon. Read the handoff the way you would read a brief from a co-author.

Chain-of-thought can help inside a planning stage, when conflicts have to be resolved before anyone writes prose. It is usually the wrong default inside the drafting stage of an already approved outline. See chain-of-thought prompting examples for that distinction.

How to keep voice and facts consistent across chapters

Consistency comes from three prompts, not from one magical paragraph:

  1. A standing style brief, with a short approved sample if voice is a problem.
  2. A glossary or terms block for words the book must use the same way.
  3. A Polisher-style check: compare this section with the notes and with the terms, and list mismatches before rewriting.

Few-shot examples help when openings or exercises drift. They should be short and varied, not three near-copies of the same page. See zero-shot versus few-shot prompting. Role and system prompts also have different jobs in a long project: the system prompt holds the policy, and a stage role can say “you are outlining” or “you are checking for consistency.” See role prompting versus system prompts.

Do not rely on the model to remember chapter 2 while writing chapter 8 unless chapter 2, or a brief from it, is in the prompt. Context windows are large, but they are not a substitute for a handoff you have actually approved.

A realistic chapter workflow

  1. Plan the chapter from the notes, and flag conflicts.
  2. Approve or correct the outline.
  3. Draft one section at a time against that outline.
  4. Build the workbook page from the approved teaching text.
  5. Check the draft against the notes and terms, then polish.
  6. Store a short brief for the next chapter.

That is also the shape of Spec Builder, then Drafter, then Polisher in Interactive eBook Builder. After you set up SEO BOT, it can publish supporting articles to the Auto-Blogger blog, each linking back to your eBook, site and brand. That is a natural place for agentic SEO: the book remains the deep resource, and the articles answer narrower questions from the same notes and the same standing rules. You can keep the supplied, editable setup or adapt the audience, voice, format and activities. Review still happens before anything is published, which is the only safe way to write long-form with a model that can still invent a fluent missing piece.

If you later send a book to print, Lulu automation is available in the Builder; the prompting habits above do not depend on that step.

FAQ

Can I paste my whole draft into the prompt for each new chapter?

You can, if it fits and if it is already approved, but it is usually a poor default. The model has more text to wander in, and you have more to review. A short “previously decided” brief plus the current notes is easier to control. Paste a previous section only when the new section must quote or continue it directly.

How do I prompt for a very long chapter?

Split the chapter into sections with their own jobs, and keep the approved outline in every request. If a section still overruns the response, lower the length target or cut the notes to what that section needs. Do not solve overflow by asking the model to “just continue” without restating the remaining headings and the source boundary.

Do I need a special prompt for reasoning models?

You need a clear goal, constraints and sources. OpenAI’s reasoning-model guidance says those models often work best with straightforward prompts, and that extra “think step by step” cues may not help. You should still write one section at a time and put long notes before the request. Check the current documentation for the model you are using.

What if the model starts repeating itself in later chapters?

That is often a handoff problem: the prompt is restating the same example, or the standing brief has grown a leftover. Trim the sample, vary the chapter notes, and add an instruction such as “Do not repeat the email script from chapter 3; point back to it in one sentence if needed.” Then inspect the next section against that rule.

Before you hit send

Ship niche eBooks with agents already built.

Get the Interactive eBook Builder for $59