The session boundary is the problem
A chat is a session. Close the tab and the context is gone; open a new one and the model starts from nothing. So people invent workarounds — paste the whole draft back in, ask the model to “read this first”, keep a note with the current state. Every one of those is a copy, and copies drift.
A document is not a session. It has one current version, it lives somewhere stable, and anyone — you, a reviewer, an AI app — can be pointed at it tomorrow. That difference is the whole reason this page exists.
What “connect once” changes
With a document link, the AI can see exactly one document, for as long as that link lasts. That is the right tool for a single job. For long-form work you want the other shape: an account connection, authorized once, which lets the AI list the documents it can reach and open the one you name.
- “Continue the chapter I was writing on Tuesday” — it reads the current text.
- “What changed since the version I sent my supervisor” — it reads the history.
- “Find the paragraph where I promised to discuss the counter-argument” — it searches the document, not your memory.
Why it must not be able to rewrite the draft
The longer the document, the more expensive a silent rewrite is. A model that “improves” chapter one while you are working on chapter four has undone work you cannot easily recover — and the usual defence, reading the diff, is impractical at forty pages.
So the write path is narrow by construction: proposed changes arrive as suggestions attached to the text they would replace, and each one is accepted or rejected on its own. The refusal is enforced by the server, not by asking the model to be careful.
Where the reviewer fits
Long documents also need other people, and other people do not want accounts. Send a review link: the reader sees a frozen snapshot with your formatting intact, comments on the sentence they mean, and their notes come back beside that paragraph. Each round is a new snapshot, so a comment never drifts away from the text it was written about.
An AI reviewer can sit in the same thread. Paste the review link into an MCP client and it reads the same snapshot, leaves comments in the same place, and sees the focus you wrote first — “check the citations”, “is the argument in the right order”.
What this is not
- Not a project manager: no kanban, no word-count goals, no writing streaks.
- Not a citation manager: references are text and links, not a bibliography database.
- Not real-time co-editing: review is asynchronous, on purpose — it is a different problem at a different price.