2026-06-20 · all guides

The Best AI Prompts for Writing a Nonfiction Book (With Examples)

The three ingredients every book prompt needs

Weak prompts share a shape: "write a chapter about meal prep." Strong prompts share a different shape: a specific reader, a specific outcome, and explicit constraints. The reader gives the model a register and an assumed knowledge level. The outcome gives each chapter a job. The constraints, word count, structure, banned filler, are what suppress the padding models produce by default.

A useful template: "You are writing for [specific reader with a specific problem]. By the end of this chapter they should be able to [concrete outcome]. Write [N] words. No motivational filler, no recap of previous chapters, no sentences that restate the heading." Every prompt below builds on this frame.

Outline prompts: do this before any chapter exists

The single highest-leverage prompt is the outline prompt, because structure is where unguided AI books fail: they repeat, drift, and pad. Ask for the skeleton first: "Propose a 12-chapter outline for a book helping [reader] achieve [outcome]. For each chapter give a working title, 3 to 5 specific points it covers, and what the reader can do afterward that they could not do before. No chapter may repeat a point from another chapter."

Then interrogate the outline before accepting it: "Which two chapters are weakest and why? What would a skeptical expert say is missing? Which chapters could merge?" Editing the outline costs minutes; editing a redundant chapter costs hours. Every serious AI book workflow, hand-prompted or tool-driven, generates and locks the outline before drafting a word of prose.

Chapter prompts: context in, padding out

Draft each chapter against the locked outline, and feed the model context so chapters cohere: "Here is the full book outline. Here are one-paragraph summaries of chapters 1 through 3. Now write chapter 4: [title and bullet points from the outline]. 1,800 words. Use concrete examples with specific numbers. Do not re-explain concepts covered in earlier chapters; reference them in one clause and move on."

Two constraint types earn their keep. Specificity forcing: "every claim must include a number, an example, or a named method; delete any sentence that would survive in a book on a different topic." And anti-symmetry: models love producing three parallel bullet points for everything, so "vary paragraph and section structure; do not use the same structure twice in a row" reads less machine-made. Expect to regenerate: the second attempt with a tightened prompt is usually better than the first attempt edited.

Critique prompts, and the honest limits of all of this

The most underused pattern is making the model attack its own draft: "You are a harsh developmental editor. List the ten weakest passages in this chapter: vague claims, repetition, filler, anything factually suspect. Quote each one." Models are meaningfully better at spotting slop than at avoiding it, and this pass gives you an edit list for free. Follow with a targeted rewrite of only the flagged passages, not a full regeneration.

Now the limits. No prompt makes the model verify facts; you must check names, numbers, and claims yourself, because you, not the tool, are responsible for the content under KDP terms. No prompt guarantees the topic has buyers; that is research the prompt cannot do for you. And no prompt chain removes the disclosure question: prompted or not, text the model wrote is AI-generated content on KDP and gets disclosed at upload, which is a private checkbox, not a public label. If you would rather not manage a forty-prompt pipeline by hand, this outline-first, context-fed workflow is what ebookdone automates end to end at /new, with an editable outline before any chapter is generated.

FAQ