Reading and reviewing with an external AI agent
A connected AI app can read a whole book, look at the figures inside a PDF, and review a paper the way a referee would — using your own library, with its own model, and leaving its notes where you can answer them. Two things make this more than pasting the file into a chat window: a long document arrives as a map the app can navigate, and pages can arrive as images so charts, formulas and scans are not silently invisible.
Before you start
- An app connected to this workspace. Reading works at Read only too; writing notes back needs Collaborate.
- The book, paper or scan in a folder MindMax has opened.
- One sentence naming what you want back — see handing work to an agent.
Long books: map, drill down, read back
A short document comes back whole. A long one comes back as its map: two levels of structure, a note per chapter, a section id for each, and a list of tables, figures and references with their pages. A 550-page book is about 12k characters of map — enough for the app to decide which chapters to read without pulling every page into its context. It then reads those sections whole, one at a time.
Read the book I have open in MindMax as a referee would. Start from the map, read the two chapters where the method is defined, and leave the problems as comments on the passages. Don't change anything.
Searching your folder is the narrow path for one exact
sentence; the map is the path for "what does this book actually argue". The same three calls power
long jobs with an agent, where doc_map_notes writes
chapter notes back into that map and reading_guide runs a full annotation pass.
Let it see the page, not just the text layer
This is the part an external agent cannot do on its own. What it receives from a PDF is normally text; a chart, a formula set as an image, a two-column layout or a scanned page is simply absent. Ask MindMax to render the pages and hand them over as images:
Treat pages 12–15 of the paper I have open in MindMax as images and tell me the trend in Figure 3. Don't guess it from the body text.
read_document {path, pages: "12-15", as: "image"} returns up to four pages per call, and the
result carries the parameters for the next batch. A single page is "12". To read small print or one
chart closely, the app adds region {x, y, width, height} as fractions from 0 to 1 — the lower half
of a page is {x: 0, y: 0.5, width: 1, height: 0.5} — and MindMax renders that part larger. Image
files come back as the image itself. A .pptx cannot be rendered as a page image here; the app can
read its text by slide, or you can look at the slide in MindMax.
If a scan has no text layer, looking at the page is the first step, not a dead end. To give MindMax
searchable text, the app can run the ocr pipeline: its own model transcribes the pages page by
page, and MindMax checks and saves the result with a per-page quality record. That costs roughly one
pass over the material, so ask it to tell you before transcribing a whole book. See
recognising scanned PDFs for what the quality report means.
Four lenses, plus three ready-made jobs
reading_guide runs a full annotation pass with a chosen lens:
| Lens | The question it answers |
|---|---|
| Guide | What is this saying, chapter by chapter, and where should I slow down? |
| Review | What is wrong with the argument, as a referee would put it? |
| Critical | Which assumptions, gaps and leaps does the evidence not support? |
| Study | What should I be able to answer afterwards, and what should I test myself on? |
The prompts a host can show in its / menu cover the same
jobs without a long sentence: prepare-book writes a chapter map and then a prioritised reading
guide, review-paper reads a paper like a referee and leaves comments, critical-read marks the
assumptions, and study-guide turns a document into questions you answer yourself. For a worked
seven-step path, see
review a paper with your agent and
have your agent annotate a book you are reading.
Comments to answer, or notes to read
Both land in the margin, but they are not the same thing:
- A few opinions you need to respond to — "this number is wrong", "this claim needs a source" —
go in as comments with
add_comments. They enter the comments panel, are signed with the app's name, and wait for you. You reply, and the turn goes back to the app. - A whole document marked up for reading — highlights, lens notes, "stop here" marks — belongs to the reading layer. Those notes appear when you open the passage, can be ignored, and do not pile up in the waiting count.
The rule is the number of decisions you are being asked to make. Two or three challenges are a conversation; two hundred highlighted lines are a reading aid. See reading and annotating and AI annotation lenses for the reading side, and working in the margin with an AI agent for the conversation side.
Read-only material stays read-only. PDFs, ebooks and slides take comments and replies, not edits. If a referee pass needs to become a changed draft, the app writes a new note or proposes changes to an editable copy.
Common problems
The answer describes a figure that is not there. The app guessed from the surrounding text. Ask for the page as an image and name the figure.
Only four pages came back. That is the per-call limit; the result tells the app which pages to request next.
The app says a scan has no text. Correct — look at the pages first, then decide whether the
ocr pipeline is worth the tokens.
Every highlighted line became a waiting comment. Ask for a reading-lens pass rather than a comment pass, or narrow it to the sections you actually want to challenge.
Next
- Canvas requests with an agent — when the question is not in the text but in what you are trying to decide.
- Understanding a paper outside your field — the same material read inside MindMax.