The AI Assistant is a conversational collaborator built for long-form writing. This article covers what it can see, how to steer its voice, what it will and won't touch in your text, and how credits are consumed when you use it.

It suggests, you decide

Every response the assistant produces is a proposal. It does not apply changes to your manuscript on its own. When it suggests an edit, you get three buttons:

• Accept. Writes the change into your text.
• Accept with diff. Shows a coloured preview before committing.
• Reject. Closes the suggestion and leaves the text as-is.

This is a design decision, not a limitation. Your book is your responsibility. The assistant exists to catch things, propose alternatives, and keep you moving through hard chapters, not to silently rewrite your pages.

Opening a conversation

Press ⌘J anywhere in the editor to open the assistant panel. The panel opens on the right by default and can be detached or moved to the bottom from the panel menu.

Three ways to start a conversation:

• Unscoped question. Type a question with nothing selected. The assistant uses the current chapter as default context.
• Selection question. Highlight a passage first, then ⌘J. That passage becomes the focus of the conversation.
• Book-wide question. Type @book at the start of your message to force a whole-book read. This costs more credits and takes longer.

A conversation stays open until you close it. Each chapter remembers its last conversation.

What the assistant can see

By default, the assistant sees:

• The text of the current chapter in full
• A short summary of the preceding and following chapters
• Your Personal preferences from General Settings
• Your active style rules (from Settings → Style rules)

It does not see other books in your account, other users' manuscripts, or anything you have not given it access to.

For book-wide questions (plot consistency, pacing across the arc, character continuity), the assistant reads a summary index it builds and maintains in the background. It does not reread the full text of every chapter on every question. The index rebuilds as you write.

Shaping its voice

Three things shape how the assistant sounds:

Personal preferences

The textarea under Settings → General → Personal preferences is re-read at the start of every new conversation. Write in first person and be specific. Examples that work:

• "I write cozy mysteries set in coastal Maine. Keep suggestions conservative on profanity."
• "Flag passive voice in action scenes. I tend to fall into it."
• "I overuse 'just' and 'really.' Catch them."
• "Comparable authors: Kate Atkinson, Tana French."

The prompt chips next to the textarea help if you don't know where to start.

Tone and length sliders

Under Settings → AI Assistant:

• Volume. How often the assistant speaks up without being asked.
• Tone. Gentle to frank.
• Length. Terse, balanced, or detailed.

These take effect on the next message. You can also override them per conversation by typing /frank or /terse at the start of a message.

Style rules

Under Settings → Style rules, you can pin specific rules the assistant will respect across every conversation in every book:

• Preferred spellings ("grey" not "gray")
• Serial comma on or off
• Specific dialogue-tag conventions
• Phrasings or character names that should never be altered

Style rules outrank anything the assistant might suggest.

What it won't touch

Even on aggressive edits, the assistant leaves the following alone by default:

• Character names, place names, and invented terminology. Preserved across rewrites.
• Dialogue tags you've set as intentional in Style rules.
• Passages inside {{ }} braces. Use these as "hands off" markers when you want a sentence preserved exactly.
• Chapter and section titles unless you ask for them to be rewritten.

What it will do without asking

The assistant will always:

• Flag sentences that break your own style rules, even without prompting
• Suggest fixes for typos it is confident about
• Add light punctuation corrections inside dialogue

You can turn off inline suggestions entirely under Settings → AI Assistant → Behaviour.

Undo, history, and auditing

Every accepted suggestion is a restore point. Open File → Version history or press ⌘⌥H to scroll back. Assistant-made changes are tagged in the history so you can see which edits were yours and which were accepted suggestions.

If you reject a suggestion and change your mind, the last five rejected suggestions per chapter are kept in the panel under Recent suggestions for five minutes.

Credits and usage

The assistant consumes credits on every message, scaled by work:

• Short questions, inline suggestions. Very low cost.
• Chapter-scope edits. Moderate.
• Book-wide reads (@book queries). High.
• Full manuscript passes (proofread, developmental pass, line edit). Highest.

Your Usage page shows the per-feature breakdown in real time. If you hit your monthly credit cap, inline suggestions continue to work on a reduced budget until your next reset.

Privacy

Your manuscripts are not used to train the assistant or any model by default. You can verify this under Settings → Privacy, where AI training is off.

If you opt in:

• Only manuscripts you explicitly mark as shareable contribute to training.
• Assistant conversations never contribute, regardless of opt-in status.
• You can opt out at any time. Anything previously opted in can be revoked. It takes about 72 hours for revocation to propagate through the training pipeline.

Full details live in the AI training and your writing article.

Keyboard-driven use

For power users:

• ⌘J. Open or focus the assistant panel.
• ⌘K then "ask". Opens the command menu, navigates to ask mode.
• Enter. Send message.
• Shift+Enter. New line within message (for multi-paragraph questions).
• ⌘Enter. Accept the current proposed suggestion.
• ⌘Delete. Reject the current proposed suggestion.
• ⌘↑ / ⌘↓. Navigate between suggestions when multiple are visible.

The panel is designed so you never have to reach for the mouse during a writing session.

When to use the assistant vs a human editor

The assistant replaces the "first editorial pass" most authors run on their own drafts. It catches typos, inconsistencies, awkward sentence constructions, and repetition.

It does not replace:

• A developmental editor reading your manuscript at book-level
• A copy editor catching subtle grammar and style issues across hundreds of pages with continuity
• A sensitivity reader for books that touch on experiences outside your own
• A beta reader who tells you whether the story works

Use the assistant as a tool in your writing session. Use humans for the editorial passes where their judgment matters most.

Common edge cases

The assistant is suggesting things that break my style. Re-check Settings → Style rules and Personal preferences. If those look right and it still ignores them, copy the specific suggestion text and contact support; it is a bug.

I keep getting generic suggestions. The assistant leans on your Personal preferences for voice-specific feedback. If that field is empty or vague, suggestions default to generic best practice. Rewrite it more specifically.

The panel won't open. Press ⌘J. If nothing happens, the assistant may be disabled at the organisation level (Settings → Account → AI features). On Starter plans during extreme load, the assistant may be temporarily queued; the panel shows a wait time.

My suggestions are in the wrong language. The assistant uses the book's language setting (File → Book settings → Language). Change it there, not per conversation.