FAQ

The editor with the Reports rail open — the shape of a typical session

The most common questions

Will you train AI on my manuscript?

No. We don't train models ourselves, and the model providers we route through (via OpenRouter) do not train on API traffic. Your book is not shared, sold, aggregated, or indexed. There is no crawler, no "featured manuscripts" gallery, no third-party analytics pipeline that gets a copy of your prose.

The full detail — including what happens on each individual model provider we route to, and how our contractual terms with OpenRouter differ from their consumer chat products — is on the Privacy page. If a specific model's training policy matters to you, pin that model in Settings and read its API terms directly.

Does the AI ever change my text without asking?

No. Every AI-written character in your manuscript goes through the accept-suggestion flow in Write. You see the suggestion in a side sheet with streaming tokens, decide whether to accept, and only then does the text enter your document. There is no "auto-apply" mode, no "silently improve while you type" background job, no chat-to-editor bridge.

Every accepted change is logged to Provenance with the exact before-text, after-text, task kind, and timestamp. So if your document changed and you don't remember making the change, open Provenance and look at the timeline for that document — the answer is there. If Provenance shows no AI event around that time, you either made a manual edit you don't remember or something else is going on. In the second case, email support with the document name and rough timestamp.

What happens if I stop paying mid-book?

Your manuscript stays. Editing keeps working. Grammar checking keeps working. Your Lore Book, Outline, Vault, and Provenance ledger all stay intact. You can open the app, read your book, edit it, export it, and delete accounts — all without a paid subscription.

What changes is the size of your AI token budget. On Professional you have a large weekly quota that comfortably covers heavy inline command usage, template runs, and Reports sweeps. On the Free plan the weekly quota is smaller — enough for exploratory work but not enough to hammer inline commands on a novel-length manuscript in one sitting. You'll break long AI runs into shorter ones or wait for the rolling 7-day window to free up spend. See Plans and Tokens for the specific numbers.

Nothing about your book is deleted or locked away. If you cancel and come back six months later, everything is where you left it.

Can I export my book?

Yes. Open the document in Write and use the Export button in the meta row above the editor. Six formats: DOCX (best for editors), PDF (best for reading and printing), EPUB (best for e-readers), AZW3 (Kindle, needs Calibre for sideloading), Markdown (best for feeding to another tool), and plain text (best for archival).

For a full audit trail — for an agent asking about AI usage, or for your own records — also export the Provenance ledger as JSON (machine-verifiable) or Markdown (human-readable). The two files together give a durable off-platform record of both your prose and its AI provenance.

Can I import an existing draft?

Yes. From the Dashboard, use Upload book. Supported formats: .txt, .md, .pdf, .docx. .docx and .md import cleanest because they preserve heading structure — H1 headings become chapters, H2 headings become scenes. .pdf and .txt come in as flat prose that you'll need to add heading levels to yourself in the editor.

Import assigns a stable ULID to every paragraph, which is what lets Provenance track changes at paragraph granularity from that moment on. Expect 5–30 seconds for a novel-length draft. If the import misidentifies chapters, adjust the heading levels in the editor — the change flows through to the Outline automatically.

What if I don't like the Deep Analysis result?

No refund once the LLM cascade started — Deep Analysis burns roughly 100k tokens across the reader-simulation pipeline, and the compute is genuinely spent whether you loved the output or not. That's why the confirmation dialog is loud about the $55 price before you click through.

You can start a fresh simulation with a new credit. A re-run may surface things the first run missed (the simulated readers are stochastic — the same manuscript gets slightly different reactions each time). And word-count-fail and synopsis-fail prechecks don't consume the credit at all — if your book is too short or your synopsis is missing, you're bounced back before any tokens are spent. See Deep Analysis credits for the full refund policy.

Does grammar checking send my text to a server?

No. Grammar checking runs entirely in your browser via a local WASM library (Harper.js). There is no code path from the grammar check to a network call — we deliberately chose a self-contained library specifically so we could make this guarantee. If you open your browser's network tab while grammar is running, you will see zero requests fire per keystroke.

The WASM binary is roughly 1.5 MB and downloads once, the first time you toggle grammar on in the Write footer. After that it lives in your browser cache. Individual paragraph checks are cached by content hash, so re-checking an unchanged paragraph is free. See Privacy for the wider "what stays on your device" answer.

Can I use the app offline?

The editor works offline. Autosave saves locally every 1.2 seconds; the sync loop catches up when you reconnect (see How your data works). Grammar checking works offline once the WASM binary has been fetched at least once. Your Lore Book, Outline, Vault, and Provenance ledger are readable offline — anything already loaded stays visible.

AI features (which need to reach OpenRouter and then a model provider) don't work offline. If you're on a plane and want to keep drafting, disable the AI commands and lean on your voice fingerprint mentally rather than fetching one. The editor treats you the same either way.

Can I use the app on multiple devices?

Yes. Sign in on your laptop, phone, and desktop. Your manuscript, Lore Book, Outline, chats, and Provenance ledger all sync via the cloud mirror every ~30 seconds. Grammar checking state and per-device UI details (which document you had open, your scroll position) do not sync — those are per-device by design.

The 30-second sync loop is designed to be boring. In practice you'll switch devices, wait about a minute, and everything's there. See How your data works for the full sync architecture, and the multi-device walkthrough on that page for a concrete "phone to laptop" example.

Does the AI write in my voice?

Only if you tell it to. Generate a voice fingerprint from your existing prose (the surface reads up to ~8,000 words from your manuscript and Vault sources and asks a model to produce a structured JSON style profile) and leave it toggled On. Once it's on, inline AI commands in Write, all 40 Templates, Desk drafts (query letter, synopsis, blurb, newsletter), and every Atomize channel inject the profile into their system prompt before generating.

The Brainstorm Pad, Chat, and Flags deliberately do not inject your voice. Pad and Chat are for thinking — inserting your voice would create weird sparring partners. Flags is supposed to be observational — voice would bias the diagnostic pass.

What's the difference between Ask and Chat?

Ask is a full-text search box — fast, deterministic, and free of AI tokens. It uses SQLite FTS5 on our end with BM25 ranking and returns snippet-highlighted excerpts of where your query matched. Use it when you're trying to find a specific passage.

Chat is a conversation with an AI that knows your book (via RAG retrieval on the relevant scope — this scene, whole book, or clean room). It costs tokens and takes seconds per response, but it can reason about what you asked. Use it when "where did I mention the compass?" isn't the right question and "does the compass motif land or feel forced?" is.

Rule of thumb: Ask for facts, Chat for reasoning.

Can the AI draft my whole book at once?

Not in one shot, and that's deliberate. The Wizard drafts chapter by chapter — you accept or reject each paragraph in a diff view. That's the sanctioned drafting path, and every accepted paragraph is logged to Provenance so you know exactly what the AI contributed.

There is no "generate a whole book" button in Templates, Desk, Atomize, or anywhere else. Three templates that did that (AI Book Writer, Complete Story, Storyteller Mode) were cut before shipping. The product decision was: if you can't audit the AI's contribution at paragraph granularity, the book isn't yours in the way authorship should mean.

How do I know how much of my book is AI-written?

Open Provenance. The hero card at the top shows "N% human-typed" for the whole manuscript — a single number derived from the character counts in the before-text vs after-text of every AI-accepted change. Each of your documents also has its own percentage, so you can see which chapter leaned on AI and which was mostly you.

Scroll down for the chronological event log — every entry has the task kind (Rewrite, Expand, Draft, and so on), the timestamp, and a side-by-side before/after diff. Click any entry to expand the full change. Use the Export button at the top of the page to grab a JSON or Markdown copy for an agent, publisher, or personal record.

What models does the app use?

The default is a Google Gemini model chosen for a balance of quality and cost — it costs roughly a quarter of the equivalent Anthropic or OpenAI model at similar quality for prose work. The picker on the Settings page offers 22 models across Anthropic, OpenAI, Google, Meta, DeepSeek, and others.

Different models deduct different amounts from your weekly token budget (a cheap model like Claude 3 Haiku deducts 3 per token generated; a premium model deducts far more). See Models and AI providers for the full catalogue with descriptions and multipliers.

What if I want to write in a language other than English?

The AI features work in many languages — the underlying models are all multilingual to varying degrees. Voice fingerprint reads your prose in whatever language you've written and produces a profile that reflects your voice in that language. Inline AI commands, Templates, and Desk drafts respond in the language of the input.

Grammar checking is currently strongest in English; other languages have varying support in the Harper.js library we use for local checking. If you write in French, German, or Spanish, expect grammar coverage to be lighter than the English experience.

Can I collaborate with another writer?

Not yet. Manuscripts.ai is a single-author workspace at the moment. Each account is its own account, and manuscripts don't have a "share with" flow or a co-editing mode. If two people want to work on the same book, they'd need to share account credentials — which we don't recommend for security reasons.

Collaboration is on the wishlist and we hear the demand. When it ships it will be built on the same provenance + sync foundation as the single-author flow, so it works cleanly across devices and preserves per-author authorship records.

Does the app work on mobile?

Yes, in a mobile browser. The editor is designed primarily for keyboard writing but is readable and editable on a phone or tablet. Sync works the same way, so a draft you started on your laptop opens on your phone with the last cursor position intact.

Some rail-heavy pages (Reports, Provenance, Deep Analysis) are more comfortable on a wide screen than a phone, but nothing is locked behind screen size. There's no native app; every device signs in through a browser.

What's the Brainstorm Pad for?

Exploration and freewriting, not drafting. Nothing typed in the Brainstorm Pad can cross into your manuscript — that's structural. Three enforcements: a system prompt that refuses to draft manuscript prose, DOM-level copy-out suppression on the Pad's output, and no "send to editor" button anywhere in the UI.

Use it to think out loud with an AI without any risk of the language creeping into your book. If you find yourself wanting the AI to draft manuscript prose, that's what the Wizard is for — that's the sanctioned drafting flow, and every accepted paragraph is logged to Provenance.

Is my data encrypted at rest?

Yes. The cloud mirror your manuscript syncs to is a managed Postgres database (Supabase) with encryption at rest handled by the cloud provider. In-flight, everything moves over HTTPS. On our servers, the local database that holds your live edits sits behind session-scoped access controls — only your account's session cookie can read your rows.

Password hashes use scrypt (Node's crypto.scryptSync, 64-byte key, per-user salt) — even we can't recover the original password if you forget it. Session cookies are HttpOnly with a 30-day sliding TTL. See Privacy for the full posture.

Yes. You can already self-export the two biggest artifacts: your manuscript (from Write, the Export button — DOCX, PDF, EPUB, AZW3, Markdown, or plain text) and your Provenance ledger (from Provenance, Export as JSON or Markdown). Those cover the majority of what a data-request answer usually needs.

For anything not covered by those exports — chat threads, lore entries, outline nodes, Pad sessions — email support with the subject line "Data request" and we'll produce a full account dump within a reasonable window (typically inside a business week). Deletion is self-service from your account settings; see Your account.

What happens if OpenRouter goes down?

AI features stop working until OpenRouter (or the specific downstream provider you were routed to) recovers. Non-AI features — the editor, autosave, grammar checking, Ask, the Lore Book, the Outline, the Vault, Reports that use our own analyzers — keep working normally. Sync keeps working too, so your writing is safe and mirrors to the cloud on the usual schedule.

When AI comes back up, retry the run you were trying to do. If you were mid-suggestion and cancelled, nothing was billed against your token budget for the incomplete work — the token accounting only fires on completed responses. See Troubleshooting for the diagnostic pass.

What happens if my token quota resets mid-chapter?

The budget is a rolling 7-day window, not a scheduled midnight reset. Older calls roll out of the window continuously as time passes, freeing up budget in the background. In practice you'd almost never hit a hard "you're out" moment mid-chapter unless you burned the whole week's budget in one sitting.

If you do run out mid-run, the current call finishes with what it's already spent (nothing is billed for tokens not generated), and any next call shows a "budget exhausted" message with the earliest time your budget will be back above the threshold. Non-AI features keep working normally — you can keep writing, keep checking grammar, keep browsing the Lore Book. See Tokens for the mechanics.

Where do I contact support?

Email us. The address is on your account settings page (via the avatar dropdown in the top-right). Include your account handle and — if the question involves a specific chapter or run — the manuscript name and rough timestamp of when the issue happened.

We answer emails ourselves. There's no chatbot triage layer and no ticket queue that goes stale. A typical response is inside one business day.

What if I find a bug?

Email support with steps to reproduce, the URL you were on, and any error message you saw (a screenshot is fine). If it's a data-loss issue — a paragraph vanished, a chapter reordered itself, autosave stopped firing — don't refresh the tab first. Open DevTools (F12) and grab any errors from the Console tab and any failed requests from the Network tab. Then email us before doing anything that might discard your local state.

The local database on our server side still holds every autosave, so even if the visible editor is confused, the underlying data is usually recoverable. But we can only reconstruct what happened if you tell us before you clear the state.