Writing with a model

Three modes. Pick the one the work needs.

A model is not equally useful at every stage of every document, and the damage it does when misused is not the same either. Byline gives the chat three modes and makes you pick one per turn, so what it is allowed to touch is a decision rather than an accident.

The three modes

Edit It changes the files. Your work is snapshotted first, its turn is committed after, and the whole turn can be undone as one unit. This is the only mode that writes, and its writing is the writing that gets disclosed.

Ask A read-only side question in its own session. Nothing is touched and nothing is added to the editing conversation except the answer. Use it for how-to, a reference check, a consistency check.

Ideas Outlines and brainstorms, saved to a research folder with their sources named, never into the manuscript. Structure and claims come back; the prose is still yours to write.

A book

Talk the chapter through, then let it draft into the file.

Start in the chat: who is in the scene, what has to happen, what the reader already knows. When the shape is right, switch to Edit and it writes into the chapter file while you watch the editor update. Then you rewrite it. That is the point of the tint: as you replace its sentences with yours, the colour drains out of the page, and what is left coloured is what you have not yet made your own.

A new book can be created with a first chapter already drafted, so you are editing rather than staring at an empty file.

The work's style, voice and topics are given to the model on every turn, so it is not guessing at register.

A turn you dislike is undone whole. There is no half-applied rewrite to unpick.

What this is not

It is not a button that produces a novel. A model drafting into your chapter writes competent, characterless prose, and Byline's honest position is that this is fine as a first pass and not fine as the finished text. The tool exists to make the difference visible: if the page is still coloured at the end, the book is still the model's.

A paper

Use it for the reading. Write the argument yourself.

On a paper the risk is not style, it is a confident sentence with nothing behind it. So the model does the reading and you do the claiming. Ask mode answers questions about the field without touching the manuscript. Ideas mode returns an outline with its sources named. Papers you import land in the work's library with a citation key you can cite directly, and the model is told never to invent a citation.

Import from arXiv: the paper is saved with the work, keyed, and out of the compiled text.

Reference lookup on any selection, so a claim can be checked where you wrote it.

Text that came from elsewhere carries a declaration of where it came from.

A work can be set to research-only, and then the model cannot write into it at all.

Research-only mode

Set per work. Every writing tool is refused, in the app and over MCP, and the model is told to describe the change it would make instead of making it. Journals that require the text be the author's own are the reason this exists: the guarantee is enforced by the tool, not promised by the writer.

A policy

Describe it once, then edit what comes back.

Policy is the case where a generated first draft is worth the most, because the structure is known and the blank page is the whole problem. Create the work, say what the policy covers and which frameworks it answers to, and Byline generates a complete draft from the template: document control, scope, roles, numbered statements, exceptions, review and maintenance, definitions. You then co-edit it. Arguing with a wrong draft is faster than writing a right one from nothing, and every statement you rewrite becomes yours in the record.

Statements follow the numbered shall, should and may style the template sets.

Control claims are mapped from a shared catalogue, so the appendix is generated rather than retyped.

Owner, approver and effective date are left as marked placeholders. It will not invent your organisation.

Why a blank stays blank

A model asked for a complete policy will fill every field convincingly, including the ones only you can answer. The draft prompt names those fields and the compiled document prints them as not set. A policy issued with a plausible wrong approver is worse than one issued with an obvious gap.

True in all three

Snapshotyour uncommitted work is saved before any turn

Per turncommitted with the model that produced it, undoable whole

Per wordownership replayed from the history and tinted live

Per worktokens and cost accounted as you go

Hosted or local models. A model reaching the manuscript from another tool over MCP is committed and disclosed exactly as one inside the app is.

AttributionWhat the tint is reading, and what the disclosure says.Policy and standardsControl catalogues, coverage, review dates, issued bundles.Talk to usTell us what you are writing and we will say whether this fits.