Prompt chaining means writing your book in separate prompts (plan, then draft, then edit), where each step has a defined input and output that you check before moving on. It adds steps but gives you control: you can find a problem where it started instead of where you noticed it.
Below is a seven-step chain with the handoff for each step, where to put human gates, the evidence for chaining, and the cases where you should skip it.
Why chain prompts instead of using one big one?
Because each stage can be inspected, corrected and re-run on its own. The evidence is modest but consistent. In "AI Chains" (Wu et al., 2021), chaining LLM steps improved task outcomes, transparency and controllability in a 20-person user study. Anthropic's prompting guide says explicit prompt chaining "is still useful when you need to inspect intermediate outputs or enforce a specific pipeline structure", and describes the common pattern as self-correction: generate a draft, have the model review it against criteria, then refine, each as a separate call "so you can log, evaluate, or branch at any point".
The same guide notes that newer models handle much multistep reasoning internally. So the case for chaining a book isn't that the model can't reason; it's that you need checkpoints. A 40,000-word book is not a task you want to discover was wrong after the final step.
What does a seven-step chain for an ebook look like?
Spec, outline, notes, draft, edit, polish, check. Here is the chain with what passes between steps (the structure is my suggested workflow; adapt it):
| # | Step | Input | Output | Your gate |
|---|---|---|---|---|
| 1 | Spec | Your idea, reader, goal | One-page book spec | Is the promise specific? |
| 2 | Outline | Spec | Chapter plan with a job per chapter | Any gaps or overlaps? |
| 3 | Notes pack | Plan + your sources | Facts and examples per chapter | Is every fact sourced? |
| 4 | Draft | Plan + notes + voice sheet | Section drafts | Does it do the section's job? |
| 5 | Edit | Draft + criteria | Revised draft + change list | Did meaning stay intact? |
| 6 | Polish | Edited draft + style sheet | Final text | Voice and flow |
| 7 | Check | Final + ledger | List of problems | Facts, links, formatting |
Step 4 follows the layout described in prompt structure for turning an outline into a draft, and step 5 is covered in how to get AI to edit without rewriting everything.
What should I pass between steps?
Only the artifact the next step needs, in a fixed format. The most common chain failure is pasting the whole previous conversation into the next prompt. It dilutes the request and brings along the old step's mistakes. Instead, define each handoff:
| Handoff | Format | Keep short by |
|---|---|---|
| Book spec | 5-8 labeled lines | Reader, promise, length, tone, exclusions |
| Chapter plan | Table: chapter, job, key points, word budget | One line per point |
| Notes pack | Numbered facts with sources | Only this chapter's facts |
| Draft | Plain text with headings | One section at a time |
| Edit instructions | Criteria list | 3-5 criteria; see below |
For the edit step, the self-correction pattern works best with explicit criteria, for example: "Check the draft against: (1) answers the section's question in the first two sentences; (2) no claim without a source in the notes; (3) sentences average under 20 words. List violations, then fix only those." Be wary of vague review prompts. Research by Huang et al. (2023) found that LLMs struggle to self-correct reasoning without external feedback, and sometimes get worse, so your criteria and your reading are the feedback.
Where should the human gates go?
After the plan, and after every draft section. Errors compound down a chain: a wrong chapter plan produces ten well-written wrong chapters. In my opinion, the plan gate is the highest-value ten minutes in the whole process. Keep a short "gate log" listing what you changed and why; it becomes your reusable checklist for the next book. If facts matter in your genre, make step 3 an explicit stage, as described in how to stop AI making up facts.
This structure is what the tool Interactive eBook Builder is organized around: Agent A Spec Builder writes the Chapter Plan and builds niche packs, Agent B Drafter writes the text, and Agent C Polisher handles clarity, voice, pacing, layout, interactivity and final QA. The Packs Database holds the QA/Gate items alongside Widgets, Constraint Packs, Research Packs and Section Blueprints, so checkpoints are part of the project rather than something you remember to do. The chain above works manually too; the tool just keeps the handoffs in one place.
Worked example: one chapter through the chain
Here is how a single chapter of a guide for first-time freelancers moves through the steps. The content is invented for illustration.
- Plan row (step 2): "Ch. 4, Pricing. Job: reader picks a starting rate and can defend it. Key points: three methods, one worked calculation, one script for pushback. Budget: 3,000 words."
- Notes pack (step 3): you paste your own rate calculation, two client anecdotes and the source for any statistic. Anything unsourced stays out.
- Draft (step 4): five sections of about 600 words, drafted one at a time. After the first one you notice it explains all three methods at once, which the plan reserved for section 2. You fix the request, not the output, and redraft.
- Edit (step 5): criteria are "answers the section question first; no claim without a note number; sentences under 20 words on average". You get a list of violations, such as one sentence making a claim with no matching note, and fix them.
- Polish and check (steps 6-7): a final read for voice, then a pass against the facts ledger.
The useful part is the third item: because the plan and the draft are separate, you can tell whether a problem is a planning problem or a writing problem, and you fix the right one. In one big prompt you would only see a mediocre chapter. The same split is the reason the Seed Prompt in Interactive eBook Builder sets up the overall structure, Niche Spec and chapter structure first, before any drafting happens.
When shouldn't I chain?
When the task is small, urgent, or already reliable in one prompt. A chain has a cost: more steps, more copy-paste, more chances to lose context. Don't chain:
- A 150-word product blurb. One prompt and one edit is enough.
- Brainstorming. Free-form exchange beats a rigid pipeline.
- Anything where intermediate outputs wouldn't change your decisions.
A useful rule (my opinion): chain when a mistake at one stage would be expensive to find at a later stage. Planning-to-drafting qualifies; titles-to-subtitles doesn't. If you want to keep your plan, draft and edit in one overview, AI plan, draft and edit for an ebook walks through the three stages together, and how long a prompt should be helps you size each step's prompt.
Before you hit send
- Does each step have a defined input, output and gate?
- Are you passing the artifact forward, not the whole chat?
- Is the plan approved by you before any drafting?
- Does the review step have written criteria?
- Did you log the changes so the next book starts smarter?
The takeaway: chain for control, not for its own sake. Plan, draft and edit as separate steps, check at each gate, and pass only what the next step needs.