Stop saying “the third paragraph.” Write the instruction on the sentence
Every chat instruction about a document starts by describing where it applies — “the third paragraph”, “the bit about the trial” — and pasting the passage in to be sure. A comment skips that: it is already on the sentence it is about, it carries a short instruction and a kind, and a reply hands the work to the next round. The instruction lives with the text, not across a window from it.
Ask an assistant to fix a paragraph and you have two jobs. The first is saying what is wrong. The second is saying where. The pasting is tedious, but it is not the real cost. The real cost is that after it, the sentence exists in two places: the document, where the editing has to happen, and the chat, where the instruction lives. Everything after that is a reconciliation between the two.
What a comment has that a chat message doesn't
A comment carries three things, and none of them has to be typed twice.
A position. It is attached to specific words. Whatever the AI receives, it receives with the sentence it is about — not with your description of where that sentence sits.
A short instruction. One line is normal. "This repeats the paragraph above." "This needs the 2024 figure, not 2022." You would say this to a colleague without preamble, and you can say it here without composing a prompt.
A kind. Note, question, rewrite, doubtful, needs a source. When the words are thin, the kind does the work: mark a doubtful number, set it to doubtful, and the request is complete without a sentence — the AI checks it and comes back with what it found.
A highlight is not a request. Highlighting says this mattered; commenting says do something about it. That distinction is the whole reason we kept comments and highlights as separate objects rather than one annotation with an optional note.
Then there is the question of who moves next
Once instructions live in threads, a thread has a state: either it is waiting on the AI, or it is waiting on you. We derive that from the messages rather than storing it. If the last word came from you, the AI's turn; if the last word came from the AI, yours. Resolve the thread and nobody's turn.
That sounds like bookkeeping. It is what makes the thing usable with two assistants. You have MindMax's AI, and you have whichever app you connected to the same folder — Claude Code, Codex, Cursor. They read the same threads, because the threads live in the workspace and the workspace is what a MindMax MCP connection exposes. The app brings its model and its own bill; what it finds here is the thing a chat window never had — the position, the kind, and whose turn it is. Without a notion of whose turn it is, the second assistant to look would answer everything the first one already answered, and every pass would cost twice. With it, a pass ends where it should: with a reply in the thread and the turn back on your side.
It also removes the thing that makes most AI review tools annoying: a queue. There is no assignee, no "sent to AI" status, no task list that silently drifts out of date. The panel says four comments waiting for the AI because four threads were last touched by you. That is all it means.
The direction that turned out to matter more
We built that turn rule for rewriting, and then noticed the more interesting use: an agent that reviews instead of rewriting.
Ask for a review — "read this the way a referee would, comment, don't change anything" — and the notes arrive in the margin, each on its passage, each signed with the app that wrote it, each waiting for you. You disagree with one and reply; the turn goes back to the AI, which reads the thread rather than a summary of it, and continues from your answer. You agree with another and either fix it yourself or ask for the pass.
That is a review shaped like the ones people actually have: not a report handed over, then discarded, but a conversation that happens where the text is. The AI's opinion is written down in a place that survives the session, next to the paragraph it is about, and the next thing that reads it — a week later, in a different app — sees the whole exchange.
What it does not fix
A comment points; it does not think for you. "Make this better" gets a confident guess, and the confidence is not evidence. Naming the problem — repetition, a stale number, a claim with no source — gets a change you can judge in five seconds.
Read-only material stays read-only. On a PDF or an ebook a comment can be answered or checked, and the answer lands in the thread; the text does not change. That is a limit of the format, not of the channel.
And someone still has to decide. Every proposed change waits for an accept. Forty comments is forty decisions, and a pass you skim is a pass you did not really make. We think that is the right trade — the alternative is a document that improves itself while you are not looking — but it is a trade, and it belongs on the same page as the feature. The thinking is not free either: it is billed to whichever app does it. This path spends no MindMax credits.
The margin is not a smaller chat box. It is the only place where the instruction, the text and the record of what was done end up in the same spot — which is why the next round can start there too.
If you want the mechanics rather than the argument: working in the margin with an AI agent covers what a comment carries, who has the turn, and what each side is allowed to do. Revise in the margin, one comment at a time walks the same thing through as a workflow, and revise with your own AI agent is the pass that comes back as tracked changes. MindMax MCP is what lets an app you already use read those threads — the protocol itself in plain terms.