Use caseWrite
Revise a draft with your own AI agent
Your own draft, revised inside the AI app you already pay for — and every change arrives as tracked changes on your text, waiting for your yes or no.
- You start with
- A draft you wrote, already in the folder MindMax has open, and Claude Code or Codex running next to it.
- Time
- About an hour, and most of it is reading the changes rather than writing them.
- Your part
- What the draft should become, which groups of edits you keep, and the two sentences you rewrite yourself because it missed the point.
What you'll end up with
- The same file, tightened — not a new copy you have to diff by eye.
- A panel of grouped changes, each with a reason, each accepted or rejected on its own.
- Your own comments answered in place, and the ones it could not answer still waiting for you.
- A version history entry naming the app that made the changes, so a bad pass is one click away from being undone.
The agent does the typing. The judgement stays on your side of the screen.
From “tighten this” to a file you accept, step by step
Open the draft in MindMax
YouA .docx, a Markdown note, a chapter you have been putting off. The workspace you have open is the one a connected app can reach, so the file has to live there.
Connect the app you write in
YouSettings → External AI → Connect. MindMax writes that app’s own configuration, lists it as connected once it makes a real call, and spends none of your MindMax credits — the thinking is billed to your Claude or OpenAI account.
Say where it hurts, in the margin
YouLeave comments the way you would for a colleague: this section repeats itself, this paragraph buries the point, this claim needs a number. Comments are part of the library, so an outside agent can read them — including the ones already sitting on the draft.
Ask for the pass, in your own words
You ask · MindMax answersAskRevise the report I have open in MindMax. Tighten section 2, keep my voice, and work through the comments I left. Don’t touch the appendix.
Naming the file and the section is enough. The app can also read what you have selected and which document is in front, so “this paragraph” usually works too.
It proposes, it does not write
A change to a document you already wrote comes back as a proposal: grouped edits located in your current text, each group carrying the reason it exists and, where there is one, the comment it answers.
Review it like a colleague’s markup
YouDeletions are struck through and additions highlighted, on your own sentences rather than in a fresh copy. Accept the groups you agree with, reject the rest, and fix by hand the one place where it flattened your argument. Nothing is written until you do.
Keep the paper trail
You ask · MindMax answersAskLeave the parts you were unsure about as comments instead of changing them.
Those arrive as ordinary comments on the exact passage, attributed to the app, so MindMax’s own AI can answer them later — or you can, on the train.
Why it works here
The review gate is MindMax’s, not the app’s
A capable agent handed a document will happily write a new version and call it done. Here the change cannot land that way: it comes back as a proposal shown on your text, and the accept button is in the app that owns the file.
Edits wait for you →Comments are a channel both sides can read
Your margin notes live in the workspace, not in a chat thread that scrolls away. The agent reads them through the library, groups its edits against them, and can reply inside the thread instead of starting a new conversation about the same paragraph.
What the connection exposes →The file stays where your work lives
Accepted edits go into the same file: a .docx gets Word tracked changes, a Markdown note is edited in place, a PDF gets an editable .md copy beside it. History records which app made the change and when.
Word and PDF →Also used here: Guide: connect an AI app · All the tools it can call · How revising works inside MindMax · Revising one comment at a time
Where it gets hard
- You pay your app, not MindMax
No MindMax credits are spent because no MindMax model runs on this path. A whole-document pass on a long report is a real amount of tokens in your Claude or OpenAI account — cheaper than reading it five times, but not free.
- The granularity is groups, not characters
A section rewrite comes back as groups you accept or reject; a suggestion at the cursor is one decision. If you want to keep half of one group, you edit after accepting — there is no per-word toggle.
- It only knows what is in the workspace
A file outside the open folder is out of reach, and an absolute path is refused for writing. Scans that have never been recognised read as images, so it will comment on the picture rather than the words in it.
- Tracked changes are a Word file privilege
In a .docx you end up with real Word tracked changes. Markdown is edited in place, and a PDF is never edited — you get an editable .md copy next to it and decide what to do from there.
- Some hosts are unverified
Claude Code and Codex are tested end to end. Cursor, Windsurf and WorkBuddy are configured for you and confirmed on first use; a given ChatGPT desktop build may not surface local tools at all.
Then what
About an hour
Same review gate, but the comments come from your supervisor and the AI answering them is MindMax’s own.
About an hour
When the draft is not the problem — you need the shelf read first.
An afternoon
When the document is fine and the thinking behind it is not.
A revision you did not read is a revision you did not make.