Most people who ask how to write a book with AI are asking the wrong question. They want to know which model writes the best prose, or what prompt makes chapters sound human. Those are real questions, but they come late. Books fail far more often in planning than in drafting, and AI assistance does not change that.
The planning architecture in the Book Creation Engine, built by P. Adhil Khan, puts six stages before a single chapter is written. Book mode, research, idea analysis, title lab, blueprint, outline. This article walks through what each one decides, what it produces, and what goes wrong when you skip it. The focus here is on the six, because they are the ones a first-time AI-assisted author is most likely to rush.
Why planning comes first, even when drafting feels urgent
There is a version of AI book writing that feels great for about three days. You open a chat window, describe a premise, and get a compelling first chapter. You ask for a second, then a third. By chapter six the protagonist’s job has changed, a sibling has been renamed, and the central conflict has quietly become a different story.
None of that is a model failure. It is a planning failure. Nothing was locked, so everything drifted.
The engine’s response is a gate before each handoff. The research output must be approved before idea analysis begins. The title must be approved before the blueprint. The blueprint is described in its own documentation as the master design document that governs everything downstream. That phrase is doing real work: it means later stages do not get to invent the premise, theme, or stakes, because those were decided earlier and locked.
Six stages is more overhead than most people expect. The trade is that once you reach drafting, you are executing a design rather than discovering one.
Stage one: book mode decides what kind of book this is
Book mode is the entry point, and it is deliberately boring. It establishes the locked book brief, creates the project workspace, and opens an unresolved-items ledger that will collect every question the project cannot answer yet.
Two things about this stage matter more than they look.
First, the ledger. Most projects lose track of open questions. Someone asks whether a character has a brother, the answer never arrives, and six chapters later the brother appears and then vanishes. The engine routes every unresolved item to a ledger and keeps it there. An item is not deleted and not silently closed. It is resolved or it stays open.
Second, ownership. The engine’s authorship is split in a way worth understanding. The engine documents name P. Adhil Khan as the owner and author of the engine itself. The author of any individual book produced with it is whatever user is running the project. Governance of the system and authorship of the book are separate roles, and the engine keeps them separate on purpose.
If you are running this manually rather than through software, stage one is one page: what kind of book, who it is for, and a written list of everything you do not yet know.
Stage two: research is where accuracy is earned or lost
Research is the stage most beginners skip, and skipping it is the single most expensive shortcut available.
The engine’s research stage ingests nine source types, including your own notes, PDFs, DOCX files, URLs, books, images, audio transcripts, and whole folders of material. Each source gets a registry entry with an ID, an origin, a format, and a status. Then every extracted statement becomes a claim row carrying a source ID and a confidence label: verified, unverified, or open.
The pipeline rules are strict about two things. Duplicates merge rather than double-register. Conflicts do not. If two sources disagree, the engine records two rows with both sources attached and leaves the disagreement visible. It does not pick a winner quietly.
For a beginner, the practical version of this is a two-column file. On the left, every factual claim you intend to make. On the right, where it came from, or the word “unverified.” It takes an afternoon and it prevents the specific disaster of building a chapter on a fact you half-remember.
The engine also produces a metadata research pack here, covering publishing data rather than story content, for the later stages that handle platform information. That is the beginning of the realization that most of a book’s production work is not prose.
Stage three: idea analysis narrows the field before you commit
Idea analysis converts research into recommended directions. It produces a prioritized primary direction plus alternatives, a reader promise statement, and a genre profile.
The most useful output here is the reader promise, because it is the constraint you will be held to later. A promise statement says what the reader gets by the end. Not the theme, not the message. The thing they can describe to a friend.
Directions are scored and recorded. The engine’s rule is that no score is invented or hidden. If you are adopting this pattern without the software, the equivalent is writing down three to five candidate book ideas and, next to each, a single sentence answering “what will the reader be able to do or understand after this that they could not before.” Ideas that cannot fill that sentence are not ready for the next stage.
Stage four: the title lab is an evaluation step, not a naming step
Beginners treat titles as branding. The engine treats them as a scored decision with an approval gate.
The title lab generates three to five candidate titles and three to five subtitles, then evaluates them on five dimensions: clarity, memorability, discoverability, audience fit, and genre fit. It also runs an SEO-friendly evaluation and an originality check against known works, and it defines the unique selling proposition and the hooks for each candidate.
Two decisions in this stage are easy to get wrong.
Discoverability and originality pull in opposite directions. A title that is highly discoverable often resembles titles already selling, which is exactly what the originality check flags. The engine scores both and leaves the tension visible rather than resolving it silently.
Subtitle count matters. Three to five subtitle candidates are generated and evaluated against the titles they pair with, not in isolation. A strong title with a subtitle that repeats it wastes the subtitle slot entirely.
The other thing worth internalizing is that a title is a promise you have to keep. If the title implies a transformation and the book delivers a survey, readers will feel the gap even if they cannot name it. The engine’s title evaluation includes a reader hook and a curiosity hook for each candidate, which forces the question of what specifically the title is promising. Answer it honestly at this stage and the blueprint stage gets easier, because the promise statement and the blueprint’s story promise should say roughly the same thing.
The stage ends with an approval gate. Nothing proceeds to the blueprint until a title is approved, which is the engine’s way of saying that the title is a design commitment rather than decoration added at the end.
Stage five: the blueprint is the document everything else obeys
This is the heaviest planning stage, and the one that pays for itself fastest.
The blueprint locks fifteen design elements: core premise, theme, story promise, genre mix, audience, tone, point of view, narrative style, core conflict, stakes, story goals, reader transformation, and emotional direction, plus the structural locks that follow. It then applies five structural systems that the engine inherited as frozen allocations.
The first is sealed-loop structure, which fixes how parts of the book close back on their openings. The second is mystery design, which manages active reader theories and holds three to five of them open at once. The third is the foreshadowing ledger, in which every clue records the chapter where it pays off. The fourth is dependency validation, which tags hidden clues to their payoff chapters and asks what breaks if the clue is removed. The fifth is twist design, tested against three criteria: fair, unexpected, and emotional.
| Blueprint lock | What it prevents |
|---|---|
| Sealed-loop structure | Endings that ignore the book’s opening promises |
| Mystery matrix, 3 to 5 active | One-note plots and abandoned questions |
| Foreshadowing ledger | Clues introduced with no recorded payoff |
| Dependency validation | Subplots that can be deleted without consequence |
| Twist tests | Reveals that feel arbitrary or unearned |
The version 2.1 upgrade added two registers here. Theme engineering trackers follow the governing theme, symbols and motifs, moral arc, philosophical question, life lessons, belief changes, and theme consistency across chapters. A reader value register tracks ten separate value types, including life lesson, education, wisdom, healing, values, moral questions, reflection, takeaway, story promise delivery, and transformation alignment.
The practical beginner version of this stage is one page with the fifteen elements written as single sentences. If any of them cannot be stated in one sentence, it is not locked yet.
One warning about the registers. Trackers and value registers are easy to fill in once and never consult. The engine’s usage rule is that entries are appended rather than rewritten, and that consumers reference the register by section instead of copying it. The point of that rule is drift detection: at the end you can read the register top to bottom and see whether the theme you locked in chapter one is still present in chapter twenty. If you fill it in at the start and never look again, you have produced a document, not a control.
Stage six: the outline turns the blueprint into order
The outline converts the approved blueprint into chapter order, story structure, and emotional progression. It is also what the timeline stage consumes next, which is why outline comes before character, world, and timeline rather than after.
Emotional progression is the part beginners underrate. Structure gets attention because acts and turning points are teachable. Emotional progression is harder to see and equally determinative. A book can hit every structural beat and still feel flat because the emotional register never shifts.
The outline is also the last stage where a change is cheap. Once character, world, and timeline have locked, and especially once image planning has produced prompts against those locks, revision costs multiply. That is the practical argument for spending an extra day on the outline. It is the cheapest place in the entire pipeline to be wrong.
At this point the pipeline branches. Character, world, and timeline all run in parallel, each consuming the outline and each producing a set of consistency locks. Only after those three complete does image planning begin, and only after image planning does writing start.
What comes after the six
Chapter drafting is stage eleven of sixteen. It arrives after character, world, timeline, and image planning have all locked. That ordering surprises people who assumed writing was the third or fourth step.
The reason is that a chapter is where every prior decision becomes visible at once. If the character bible is not locked, the writing stage has nothing to hold the character to. If the timeline is not locked, a scene set in autumn can be contradicted by a later one. The engine pushes writing late precisely because it is the stage that would otherwise paper over every unresolved planning question with fluent prose.
After drafting comes rewrite, quality assurance, metadata, publishing, and export. Quality assurance is worth flagging: the engine’s version 2.1 detectors look for plot holes, timeline conflicts, object consistency, scene logic, emotional consistency, dialogue consistency, and AI writing fingerprints. That last one is a signal about audience trust more than craft.
A realistic starting sequence
If you want to begin this week without adopting a whole system, do these six things in this order.
- Write down the kind of book, the reader, and every open question in a single file.
- Build a two-column claims-and-sources file for any factual material.
- Write three to five candidate ideas, each with a one-sentence reader promise.
- Draft three to five titles and score them on clarity, memorability, discoverability, audience fit, and genre fit. Pick one.
- Write the fifteen blueprint elements as single sentences.
- Convert that into chapter order with an emotional note per chapter.
Only then open a writing tool.
The idea analysis stage and the blueprint design locks are documented in more detail in the engine’s own materials, and the author profile lists the wider toolset these engines belong to.
FAQ
Do I need all six planning stages for a short book?
The engine applies them to every project, but the effort scales. For a short nonfiction book the blueprint may be one page. The sequence stays the same because each stage produces an input the next one consumes.
Can AI handle the planning stages for me?
It can draft candidates, but the approval gates exist because a human must evaluate. The engine requires approval between stages. Automating the approval turns the pipeline into an expensive document generator.
How detailed should the blueprint be?
Fifteen elements, each stated as a single sentence, plus the structural locks. Detail belongs in the outline, not the blueprint.
What if my research turns up contradictory sources?
Record both. The engine keeps conflicting claims as separate rows with both sources attached and never resolves them silently. You resolve them, and the record shows that you did.
Why is image planning before writing?
Because images have to match the story, chapter, timeline, location, character, and emotional beat. Planning them after drafting means re-planning once chapters shift.
Is drafting with AI still worth it after all this planning?
Yes, and it is more effective. Once the blueprint and outline are locked, drafting becomes execution against a specification, which is the condition under which generation is most reliable.
The takeaway
The order of operations is the whole answer. Book mode, research, idea, title, blueprint, outline, and only then drafting. Six stages of decisions that constrain what a chapter is allowed to be.
Skip them and AI will happily write you a fluent book that contradicts itself by chapter six. Do them and the model becomes what it is best at: a fast, tireless executor of a design you already own.