Read
Ask AI about screenshots, charts and photographed pages
A picture is a document too. The chart someone screenshotted into a chat, the page of a book you photographed because you could not borrow it, the diagram that explains the whole architecture. These are the files most tools treat as decoration and then cannot find again.
Images arrive in knowledge work constantly and then vanish into the folder, because nothing about them is searchable and opening one tells you only what you already remembered. The fix is not a better gallery. It is that the picture should be readable — once, cheaply, and in a way that every other part of the app can then use — and that what is locked in the pixels can leave as text you can edit.
Read once, used everywhere
A description is stored against the picture itself, not against the place you happened to be looking at it from.
- Ask about an image in the viewer and drop the same image on a canvas, and it is the same reading — you do not pay twice for the same picture.
- The description goes into your knowledge base, so the image turns up in later answers about the subject instead of being invisible to everything.
- It is labelled as a description of a picture. An answer can say what the diagram shows; it cannot present that as a sentence quoted from a page.
- Reading happens when you ask for it, or the first time an operation genuinely needs the pixels — not automatically on every file you scroll past.
Ask about part of the picture
Most useful questions about an image are about one corner of it.
- Drag a box around a region and ask — one panel of a multi-part figure, one column of a screenshotted table, the legend you cannot read.
- The region travels with the question, so the answer is about what you pointed at rather than about the picture in general.
- On a canvas the same region becomes an item of its own, which can be expanded, compared with another image, or argued with.
- Six of the canvas skills insist on the original pixels rather than a description — reading a chart back into numbers, taking a layout apart, criticising a design.
A picture becomes a note you can edit
A table screenshot, a chart, a formula, a whiteboard, a page you photographed — whatever is locked in the pixels can leave as ordinary Markdown, saved next to the picture with the picture still in it.
Markdown table
Every row and column in the original order. Numbers, units and symbols kept as printed. Uncertain cells are flagged, not filled in.
Data + reading
The plotted values come back as a table (axis estimates marked with ≈), then a short reading: what is measured, the trend, the extremes, the caveats.
LaTeX
Standalone equations in their own blocks, equation numbers kept, the conditions around them as plain text between the maths.
Faithful outline
Title and nested bullets, key sentences verbatim, tables still tables. Page numbers and running heads dropped — page furniture is not the argument.
Structured notes
Sections and lists recovered, arrows spelled out as relations — A → B: the reason. The author’s wording kept; nothing added that was not there.
Or box just the part you need — one table in a screenshot, one panel of a figure — and convert that alone; the same box and the same menu work inside a PDF, a book or a deck, where the note lands beside the document instead. You are not guessing at a menu: the toolbar decides which of these it is. The result is an ordinary .md file beside the image — searchable, linkable, editable — and anything unreadable is marked rather than invented.
Living with a folder full of pictures
A contact sheet for the folder
Every image in a folder as one sheet you can scan, instead of opening thirty files to find the one you meant.
Page through and zoom
Arrow through the folder without going back to the file tree, zoom into the part that matters, and box a region at any zoom level.
Names that say what they are
IMG_4821.png can be renamed from what is actually in it, which is the difference between a folder you can search and a folder you scroll.
Figures inside documents too
A chart in a PDF or a slide can be boxed and asked about the same way, without extracting it into a separate file first.
Stored beside the work
Images you bring onto a canvas live in a folder next to it and follow it if you rename or move it. Nothing is copied into a database.
Filter by what is in the picture
Once a picture has been read, the contact sheet can be filtered by its contents — the chart about Q3 revenue, not the file called IMG_4821.
Where it gets hard
- A description is not a transcript
What the model writes about a picture is an interpretation, and it is stored as one. For a photographed page of text where you need the words exactly, recognition — not description — is the right tool, and it is a separate, auditable step.
- Conversion is transcription with rules, and it can still be wrong
It only writes what is visible. Unreadable marks come back as [illegible] and anything inferred is labelled as inferred — but scribbled notes, tiny axis labels and crowded tables are where image reading goes wrong, and it goes wrong confidently. Check the result against the original before building on it.
- Reading a picture costs credits
Descriptions are cached against the image, so the cost is once per picture rather than once per question — but pointing at a folder of four hundred photographs is a real bill. Conversion is a separate run from reading.
- Some formats view but do not read
SVG, AVIF and ICO can be opened and looked at, but are not sent to the model. HEIC and TIFF are converted for display on macOS only.
- It is not a photo library
There is no face grouping, no albums, no editing. This is for pictures that carry an argument, not for the two thousand photographs on your phone.
Common questions
Can I ask questions about an image?
Yes. Open it and ask, or drag a box around one part of the picture — a single panel of a figure, one row of a table, one region of a screenshot — and ask about just that. The answer is about the pixels, not about the file name.
Are images included when I search my folder?
Once an image has been read, yes. What enters the knowledge base is the model’s description of the picture, and it is marked as a description rather than presented as text printed on the page — so a later answer cannot quote it as if it were a verbatim source.
Does it read the same image twice?
No. A description is stored against the image itself, so the viewer and the canvas share one reading. Opening the picture again, or using it in a different place, does not pay for it a second time.
Can I turn a screenshot into a note?
Yes. A picture can become an ordinary Markdown note saved next to it in your folder — a table screenshot becomes a real Markdown table, a chart becomes its data plus a short reading, a formula becomes LaTeX, a whiteboard becomes structured notes. The note then behaves like any other note: searchable, linkable, editable.
Can it pull a table out of a picture?
Yes. Choose table conversion and every row and column comes across in the original order, with numbers, units and symbols kept as printed. Or box just the table inside a larger screenshot and convert that region alone.
Which image formats work?
PNG, JPEG, WebP, GIF and BMP are both viewable and readable by the model. HEIC and TIFF are converted for display on macOS. SVG, AVIF and ICO can be viewed but not read.
See also
When what you need is the exact words on a photographed page rather than a description of it.
Where images become cards, regions become items, and a whiteboard photo becomes a map you can keep expanding.
Boxing a figure inside a document and asking about just that region.
What a boxed region carries besides the pixels, and why the numbers come from the text under it.
The other format that is half picture and half argument.
How a picture that has been read turns up in an answer weeks later.
What happens to the note once the picture has become one.