It looks like a screenshot. It is not one.
Asking AI about a screenshot of a chart works, but it throws away the text under the picture, the page it came from and the document around it. Boxing a region inside the document keeps all three, so numbers are read from the text rather than guessed from pixels, and the answer links back to the page.
Everybody has the same workaround. You hit a figure you cannot parse, you screenshot it, you paste it into a chat window, you type "what is this saying". It works often enough to become a habit.
It also fails in a specific, repeatable way, and the failure is not the model being bad at pictures. It is that by the time the picture arrives, three things that were sitting right next to it have been thrown away.
What the crop drops
Open the figure in the actual document and look at what is there. The axis labels are text. The caption is text. The table under the chart is text, with the exact values in it. The page number is known. The paragraph that explains the figure is two inches above it.
Crop to a PNG and all of that is gone. What reaches the model is a grid of pixels in which the
number 0.37 and the number 0.87 differ by a few dozen of them.
So the model does what it must: it reads the picture. Most of the time it gets the number right. Sometimes it gets it nearly right, which is worse, because a number that is nearly right does not announce itself. And the answer that comes back has nowhere to point — you cannot click it, the conversation has no idea which document that was, and the follow-up question ("what does the paragraph above it say?") is unanswerable without another screenshot.
What we send instead
In MindMax you drag the box inside the reader, and four things travel with it.
The image, because some things genuinely are pictorial: which box is under which heading, what the arrows connect, how the layout groups things. This is what the picture is for.
The verbatim text underneath the box, harvested out of the document rather than recognised out of the pixels. This is the part that does not exist when you paste a crop into a chat window.
Where it came from — file, page, chapter, slide. The answer can therefore be clicked back to the spot, and the thumbnail in the conversation reopens the document at the right page.
The document itself, attached to the question, so the model can go and read the surrounding section rather than speculating about the caption.
The rule that does the work
Having both the image and the text is not automatically an improvement. The model still has to be told which one wins, because it will happily read a value off a chart while a table containing that exact value sits in its context.
So the instruction is explicit: where a region has a real text layer, numbers, formulas and table values follow the text, and the image is only for layout, shapes and spatial relationships.
The interesting part is the other branch. Scanned pages, diagrams with the labels drawn in, and figures in image-only PDFs have no text underneath. We could have papered over that — transcribe the region silently and present the result as if it were source text. We do the opposite: when there is nothing under the box, the model is required to say that a figure was read off the image and may be wrong.
That branch matters more than it looks, because it is the one case where the honest answer is less impressive than the dishonest one. A tool that always sounds equally confident has told you nothing about when to check.
There is a third case, and it is the one that caught us out during implementation. If you ask a follow-up question two turns later, the image is no longer in the conversation history — history is text. Our first fix was to keep a transcription of the region around instead. Which means that on turn three, the "text under the box" is not text from the document at all: it is something the model wrote about the picture on turn one. Feeding that back as the authority on numbers is exactly the failure the rule was written to prevent, dressed up as compliance with it. So the two are labelled differently all the way through, and a model-written description is explicitly marked as not quotable and not authoritative on any figure.
Why it is the same box everywhere
This is one action in every reader — PDFs, EPUB and Kindle books, PowerPoint slides, images, your own Markdown notes — and the region adapts to what it was drawn on: a page in a PDF, a chapter in a book, a slide number in a deck.
That uniformity is not tidiness for its own sake. The moment it exists in only some readers, you stop reaching for it, because a gesture you have to check the availability of first is not a gesture. It has to be reflexive to be worth anything.
Keeping the answer
The last piece is the least clever and possibly the most useful. An explanation of a figure is worth having next to your work, not buried in a conversation you will not scroll back through.
So a region and the answer it produced can be saved into your daily note in one step. When that happens the picture is copied into your own folder — not referenced out of the conversation's storage — because a note that goes to a broken image after you tidy up old chats is worse than no note at all.
What is not claimed
Handwriting and tiny axis labels are still where image reading goes wrong, and it goes wrong confidently. The box does not fix that; it narrows it, by giving the model the text whenever text exists and making it admit when there is none. And a figure that spans a page break is two regions, not one.
The claim here is narrow and checkable: what you point at arrives with its context intact, the numbers come from the document when the document has them, and the answer knows the way back.