From a paragraph of intent to a finished manuscript.

Write the Brief

Create a book and describe it: what it is about, who it is for, what it should feel like. This is the single highest-leverage field in the whole application - the title, the structure and every section of prose are generated from it.

Directly under the Brief sits Target audience, which answers a different question: not what the book is about and not how it sounds, but who it is for. Describe the reader rather than the topic - prior knowledge and reading level matter far more than demographics. What you put here governs vocabulary and how often the prose stops to explain something. Leaving it empty is fine: the AI then writes for a general reader.

Set the book's Style & Language at the same time, so register and output language are fixed before anything is written. Both can be improved with AI if you would rather start from a draft than a blank field. You do not have to write the style instruction from memory each time either: a style profile is your own voice measured from work you uploaded - register, sentence rhythm, diction, point of view, imagery, dialogue habits - and applying one copies its description into the book. Change the profile later and no existing book is restyled behind your back.

Above the audience sit two more fields, and they answer what the book is for. Your goal is why you are writing it - to persuade, to explain, to be found as an expert, to sell something - and it chooses where the emphasis falls and what stance the text takes, without ever being named in the text itself. The reader outcome works the other way round: it is what someone should be able to do, decide or feel once the book is closed, and the AI weighs every page against it, so material that does not serve it is cut. Write that second field carefully, because a vague outcome cuts vaguely. Both are optional; leave them empty and the book is written exactly as it was before.

On the same form, Words to avoid takes two lists - single words and whole turns of phrase - that the AI is then forbidden to use anywhere in this book, however well they would fit. A new book starts with the em dash already on that list.

Add source documents - or start from existing work

Upload the material the book should be built on - reports, transcripts, notes, prior drafts. Scribith indexes them and searches them while drafting, which is the difference between prose that is about your subject and prose that is from it.

You can queue those additional documents while you are still filling in the create-book form; they upload as part of creating the book rather than making you save first and come back.

Existing work can go straight into the binder instead: Import a manuscript takes a Word file, a PDF, an EPUB or a zipped Scrivener project and finds the chapters in it - through your heading styles in Word, per chapter file in an EPUB, and in a PDF by stripping running heads, feet and page numbers. It is a four-step wizard rather than a button, and the step that matters is the one in the middle: before anything is created it shows you the structure it found, a line per level and a total, so a bad split is something you reject rather than something you discover afterwards in a book you now have to delete. Then it reads the text for the same properties Analyse a Book infers - title, premise, brief, template, structure, voice, image style - and flags whatever it was unsure of, all editable before you commit. Set the writing language before you analyse: it is the language of the book you are creating, not of the file you uploaded. An import always goes into an empty book, never over existing work.

Two routes start earlier still. Analyse a Book reads a finished manuscript and answers questions about it before you decide what to change - it imports nothing. Research from sources goes the other way: upload up to eight books you have read on the subject, write in one sentence what your book should do, and Scribith distils where they agree, where they contradict each other and what none of them covers. That goal is the lens they are read through - change it and the sources are read again. Both ask what language your new book should be in before the billed run, not after it.

And one route has no text as its source at all. Book from a repository points Scribith at a GitHub repository - owner/repo, with a name and a goal for the project - and works its commits, pull requests and issues up into a book. It does not read your source files, and the page says so where you start. What it finds becomes a knowledge map over five sections - themes, arcs, decisions, turning points and gaps - which you edit before anything is planned or written. Cost is stated before it is spent: the exact batches, groups and summary calls a run will make, a ceiling, and how many batches were already read and so will not be paid for twice. Store your own GitHub token in Settings → Secrets and it reads a private repository too, raising GitHub's ceiling from sixty shared requests an hour to five thousand on your own account.

Settle the plan and the cast

Brainstorm until you know where the book lands, then turn that into The Plan: one page holding the conclusion, the audience, the style, the story structure, and the canon - the fixed facts, timeline, point of view and terminology. Every generator reads it, so what the plan leaves out is what the book gets written without. A later brainstorm can be merged in rather than overwriting what you wrote by hand.

A generated plan now closes with two things that carry you further. The first is a proposed chapter list rather than one line per story movement: one line each saying what that chapter has to deliver for the book to reach its conclusion - and the binder is built from that list, so how finely it is drawn decides how much the next step has to invent. The second is the set of questions the brainstorm raised and did not settle. Each gets an answer box below the editor: Save answers keeps your decision on the book without touching the plan, and Update the plan hands them to the AI, which works each answer into the section it belongs to and removes the question it settles. Answers are matched to the text of a question rather than to its position, so re-ordering the list never re-points an answer at the wrong question.

Record the people, places and things in the Story Bible. Each element carries aliases, a short AI summary and your own private notes. When a section mentions one of them, that element - and only that one - is handed to the AI, so a large cast costs nothing and names stay spelled the same way in chapter fourteen as in chapter three. Private notes are never sent and never exported.

Generate the structure

Press ✨ Generate structure and Scribith proposes a chapter and scene outline from the Brief. You control how big that outline gets.

Then it is yours: rename, add, delete and rearrange in the binder until the shape is the book you actually intend to write. Every move can be undone.

The shape can now be judged before a word of it exists, which is exactly where a structural problem is still cheap to fix. The structure critic reads the binder instead of the prose - every section, its synopsis and its place in the order - and judges the shape against your brief, your plan and your audience. It is one small model call, seconds rather than minutes, so you run it whenever the shape changes. Below the verdict there may be a second reading from a model by a different maker, judging the same binder independently with the first verdict withheld from it: where they agree you can act with more confidence than one opinion ever justified, and where they disagree you have found the genuinely arguable part of your structure. Revise the binder turns that into a proposal, and the proposal is a draft you may reshape first - rename, move, add, remove, and put back anything the revision dropped, along with that section's prose, pictures and comments.

A hundred and forty ready-made structures are available if you would rather start from a known skeleton - the list grew from 42 to 140, ten new ones per template - story arcs for fiction, and for working documents a set that fixes the required units and their order before the AI writes anything, so methodology, evidence, risks, accountable actions and review do not disappear because a prompt happened not to ask for them. Business structures cover a business plan, a strategic plan, a project status report and a market research report among others; a government template adds the neutral, evidence-led register and the binder shape for a performance report, a programme evaluation, a regulatory impact assessment, a public consultation report and an audit or inspection report. And if you write the same kind of document repeatedly, Template from document takes one finished example, separates its recurring headings and purposes from its one-off names and numbers, and proposes a reusable structure you edit and then save into the list. Screenplay went from three structures to seventeen and children from three to fifteen; academic gained the doctoral monograph, the article-based thesis, the PRISMA-ordered systematic review, the research proposal and the methodology chapter. The list is filtered by the template you chose, so a longer catalogue is not a longer dropdown - and while you type the title, subtitle, brief or audience, the form already fills in a matching template and structure for you, locally and with no paid AI call, saying underneath that it was a suggestion.

If you already worked the outline out, no AI needs to be involved at all: Build structure from text takes an indented list - or JSON copied straight out of another tool - and turns it literally into the tree. Anything that is not a list line is skipped, so you can paste a whole plan in one go. The same structure also downloads as a file, edits outside Scribith, and applies back without losing the prose already written into it.

To work on the shape of one chapter rather than the whole arc, the Corkboard opens any container as index cards: edit a title or synopsis in place, drag to resequence, or lay the cards out freeform and let Commit order turn that arrangement into real binder order. Fill empty synopses is the only paid action on the page, and it never overwrites one you wrote yourself.

Draft the prose

Write a one-line synopsis on each section - this is what the generator reads. Then draft: one section at a time while you are thinking, or the whole book at once with 📖 Generate whole book when the outline is settled.

Prose length is set per section, and any passage can be revised with an instruction in your own words. Chapter introductions are generated separately, and the button follows the content rather than the depth: any part or chapter with material beneath it can ask for an opening, drafted from the prose and synopses across that whole subtree - so a chapter nested inside a part gets one, and a leaf with nothing below it does not offer the action. Nothing is final; every section keeps its previous versions.

Revise and research

Research, notes and sketch panels sit beside the text so context never leaves the screen, and the 🔤 Words tab answers with synonyms in the book's own writing language - for a passage, or for a chapter title, where word choice counts most per word. Focus Mode clears everything else away, and still lets you move between sections without exiting.

For a book that was outlined or imported before you started caring how it read, two whole-book passes rewrite the wording into something a native speaker would have written - one over titles and synopses, one over the prose. Both tell you how many sections they will touch and wait for you to confirm; version history keeps what each section said before.

Where the writing has to cite something, References in the content toolbar keeps the book's sources in one catalogue and inserts a numbered Vancouver-style marker - [1] - at the cursor. The list itself is built from that catalogue when the book is exported, so the numbers in the prose and the entries in the list cannot drift apart, and a book setting decides whether that list sits once at the end of the document or at the end of every chapter.

None of it has to be typed. Every field that takes more than one line - the prose, the research and summary boxes, every chat and every comment - carries a microphone button, and what you dictate lands at the cursor rather than at the end. Every control in that toolbar also explains itself on hover, as does every field with an ⓘ beside it and every page with a ? beside its title.

Edit the draft

Eight reports read the finished text and change nothing in it, and the Editorial Desk hangs them off the six passes a book editor works in: assess, develop, line, copyedit, proof, produce. That order exists to protect your money - polishing sentences in a chapter you then cut is paying for nothing. Start at the end while structural findings are still open and the page says so; it is a question, not a block.

Manuscript Health covers pacing, point of view, continuity and tone. Scenes turns that same report into one row per scene - length, tension, who is named, point of view - plus a timeline of every time marker it found. Spelling & consistency covers spelling, punctuation, grammar and the names and facts that have to match across chapters. The Style report hunts adverbs propping up weak verbs, passive voice, filler words and clichés - it runs instantly and costs nothing, because no AI model is involved. Review gathers every comment thread into one list, with a status and an assignee. And Check a document runs the same checks against any file you upload, with no book involved at all.

Two reports are new and ask a question the other six did not. Natural writing asks whether the book reads as if a native speaker wrote it rather than only as if it is correct - because prose composed by a model in Dutch, German or Italian can be grammatically perfect and still unmistakably foreign: the word order defensible, the rhythm not. Six dimensions: translationese, AI tells, idiom, register, repetition and rhythm. The first four section by section, the last two from a pass over what that read recorded - because a phrase used twice is a habit and the same phrase opening fourteen chapters is a tic. Everything is judged against the norms of the language the book is written in: a construction that would be clumsy in English but is entirely ordinary in your language is not a finding. Each finding quotes the passage and suggests how a native writer would put it.

And the External critic answers the question no report asked: is the book good, and is it the book you said you were writing? It reads every word of the manuscript - not a sample - and answers with a verdict, what works, what should change and what could change. It is deliberately not blind: it has read your brief, your plan, your target audience and your intended style, which is what lets it say that you promised a thriller pace and act two reads like a memoir. The Reader below is the exact opposite, and the two are not interchangeable. You then argue with it in the chat beside the critique - it concedes to an argument, but not merely because you keep pushing - and Build the list turns the critique plus that discussion into improvements: a point you argued down does not become an item, a problem that surfaced only in conversation does. Each item carries why it matters, roughly what the book gains and roughly what the work is, ordered so that what has to be decided first comes first. It is the most expensive single action in Scribith, because it reads the entire book.

None of them edits anything. Acting on a report is a separate step: Fix all findings and Correct all work through the findings one section at a time, showing the current text beside the rewrite, and save nothing until you accept it. That is right when you want to read each rewrite and wrong when a long report has left you with thirty of them, so Manuscript Health also offers Fix all automatically: the same sections in the same order, accepted and saved as each one arrives, with the report updating as it goes. It is the same paid work either way, and an automatic run now works through the whole report rather than ending at the first thing it did not expect: a section that fails is tried twice more and then skipped, a rewrite that may have been truncated is skipped and never saved with its ending missing, and whatever it skipped is listed when it finishes. Each attempt is another paid request, though, so a run with failures in it can cost more than the number in the confirmation. Everything already saved stays saved, and every rewrite is in that section's version history.

Whole-book passes are the longest and most expensive thing Scribith does, and they no longer depend on your browser staying connected: the run's state lives on the server, so a dropped connection reconnects to the run that is already going rather than starting a second, separately billed one. Manuscript Health checkpoints after each stretch of the book it has read, so a run that failed part way continues from there instead of re-reading what you already paid for. And Keep working lowers the progress window into a small control in the corner so you can carry on elsewhere while the check runs - and that now covers generating a whole book, illustrating it, filling in every synopsis and translating a book as well.

One model judging your book is one opinion, and nothing on the page distinguished a real structural problem from a confident invention. Manuscript Health can therefore be read back by a second reviewer - a model from a different maker, so a genuinely separate reading - which goes over the findings with the passages they are about. Agreement is silent: only disagreement is marked, because a badge on nine findings in ten would be noise rather than information, and a tick takes the disputed ones off the page entirely if you want. It is a separate paid call, so a new book starts with it switched off; you turn it on per book on the book's AI tab.

Each pass carries one overall score with a line across the previous runs, because "what is wrong" and "is it getting better" are two different questions. Sixty individually anchored findings can also be twelve problems, and a list cannot say that - so Write the letter gathers the open findings into a handful of named problems and orders the work by dependency rather than by severity, because repairing the structure underneath dissolves half the notes sitting on top of it. That letter converts, with no further AI call, into a guided revision plan: ordered packages that each keep their reason, location, priority and linked findings, marked To do, In progress, Done or Deferred, with notes and a completion bar - and the book's strengths listed separately, so revision does not erase what already works. Change the manuscript and the plan is marked stale rather than quietly pretending it still describes the current draft.

The desk keeps the paperwork too. The editorial brief records the scope of the edit - level, deadline, language, medium, audience, what is in and out, reference guide, deliverable, feedback style, non-negotiables, budget. Global editorial queries hold what applies across the manuscript rather than to one passage - a question, an instruction, an author decision, a fact check, a permissions or production issue, a consistency ruling - each with an owner, moving from Open to Answered to Resolved. Your book's style sheet starts from the writing language's own defaults or from Chicago, Oxford or AP, and you then record only where this book departs from it; every effective value shows its source, so an inherited rule never poses as a decision an editor made. Every report downloads as a Markdown remarks file for an editor who does not use Scribith, and so do the letter and the plan.

One more read is worth having before you call it done, and it is not a report. Reader is a beta reader persona that speaks in the first person about reading your book - where interest dropped, where it got confused and why, what it expected to happen next, who it cared about. It is told nothing beyond the title, subtitle and language, deliberately: a model that has read your plan answers as an informed insider and quietly explains away the very confusion you are trying to catch.

Illustrate and dress the book

Image sections render what you describe; infographic sections read the surrounding chapters and turn what the book actually says into a labelled diagram - using the figures your text contains, never invented ones. Both follow the book's image style, so the art looks like it belongs together.

Illustrating a whole book at once is the wrong run when only a few passages need a picture. Each written section carries a single choice - Visual: off, Image or Infographic - which saves immediately, costs nothing and changes no prose. When the plan is settled, Generate selected visuals works through the queue in reading order, tells you how many nodes it will add before it starts, and clears each selection as it completes it - so a rerun retries only the visuals that failed.

On the Publish page, give the book a front cover, a back cover and a details page. Cover text is spell-checked in your language before it is drawn, and an existing cover can be adjusted in your own words rather than redrawn from scratch. The Colophon holds everything that belongs on a copyright page: rights and licence, imprint and edition, three ISBNs, credits and trade data.

Choose the interior, then publish

Before you export, decide how the book is typeset. An interior design fixes the typography - what the body text is set in, how far paragraphs indent, how a chapter opens, what marks a scene break, what runs along the top of each page - and you choose it by looking at it, from a gallery of miniature pages that reflect each design's type family, chapter alignment and sink, numbering, ornament, paragraph shape, opening initial and page furniture. Ten ship: Classic, Literary, Modern, Manuscript, Accessible, Romance, Suspense, Fantasy, Practical nonfiction and Academic, and your administrator can add more. Around thirty controls sit under them, each resettable to the design's own value, with a preview beside them that redraws immediately - a real typeset page of your book, not a mockup. Choosing a design keeps the overrides you already made, and a book that never picked one keeps exactly the layout it always exported with.

Nine export formats, each covering the whole book: Markdown, Word, EPUB with chapter navigation, an A4 PDF, a print-ready paperback interior at the trim size you choose, a Scrivener project that imports back, the image set, the covers, and a full backup. Section numbering and a table of contents are toggles, and where the book carries citations its numbered reference list is built from the book's own source catalogue - once at the end of the document, or once per chapter, as you set it. For the paperback you first pick the printer - Amazon KDP or Pumbo (NL) - because they share no specification at all; with Pumbo you also get a cover wrap PDF carrying back, spine and front on one press-ready page.

The print pass ends in a sign-off rather than a checkbox. Proof renders the real print interior once for the printer and trim you chose, keeps that exact file, and checks its page geometry, fonts, page furniture and composition before offering you the download. An objective failure - a wrong trim, an unembedded font - blocks sign-off; composition concerns stay warnings for a human to judge. The approval belongs to the PDF you downloaded, not to the abstract book, so changing the manuscript, its images, the printer, the trim or the interior afterwards makes the proof stale and reopens the gate. Reading the proof is still yours: a mechanical preflight cannot decide whether a page looks good.

The Full backup is the one that is not a manuscript: it holds the binder, prose, images, story bible, comments, plan and colophon in a single file, and Restore turns it back into a working book - always a new one, never merged over an existing book. The same manuscript can also become derived work: blog posts written inside Scribith - which now also export as an editable Word document, with real headings rather than a pasted picture of the preview - a document at a real page size, or a presentation built from the book or part of it. And on the same page sits the Chatbot with the entire manuscript open in front of it, for the questions you only ask at the end - does the ending deliver what the opening promised? What that question costs is stated before you ask it.

For anyone publishing on Amazon it does not stop at the file. KDP asks for seven keyword fields of fifty characters and a product-page description of up to four thousand, and the Amazon (KDP) tab writes both, in the language the book is written in, from the book's own metadata - title, subtitle, genre, brief, plan, audience and chapter titles. From the plan, that is, not the manuscript: sending the prose would multiply the cost for material the model does not use. The seven keywords appear as seven separate fields, the shape of KDP's own form, each with its own copy button and always padded to seven. Amazon's own rules are part of the instruction - no title or author name inside a keyword, no bestseller claims, no review quotes or links - and the character limits are enforced when the text is generated rather than left to the model to respect. Nothing is saved: leaving the tab discards it, so there is no draft to go stale against a book that has moved on.

What the book has cost is on that page too. Support Scribith reads only this book's recorded requests and shows its input, cached and output tokens, its request count and an estimated euro cost. Where the deployment has configured them there are optional iDEAL and PayPal links beside it - an invitation, not a charge, and with none configured the usage and cost simply remain visible.

And optionally: the same book in another language

Translate creates a second book in the language you choose and leaves the original exactly as it is - so a translation that goes wrong costs you the translation, never the manuscript. First you settle the glossary: every name and term from your Story Bible gets keep as is, translate, or your own wording. That list is not a formality, because each section is translated separately and without it the same character can pick up three different names across three chapters.

The run works section by section and names each one as it goes. You can stop it; what is done is kept, and resuming never pays twice for the same work. Sections that failed keep their original text and can be retranslated one at a time. For a manuscript that is not a book here yet, Translate a manuscript does it in four steps: upload, check the import, pull out the names and terms, translate.