stop screenshotting text for ai. do this instead.

elisabeth hitz · july 14, 2026 · 4 min read

when you screenshot a message, a table, or a paragraph and paste the image into ai, you force the model to read a picture of words instead of the words themselves. that swap is invisible to you, but it quietly costs you accuracy and tokens. the fix is boring: copy the actual text, or if the text only exists as an image, convert it first. paste words when you have words.

almost everyone does the screenshot thing. it feels fast. you see the text on screen, you grab it, you drop it in. but the model does not see what you see.

what actually happens when you paste a screenshot?

a screenshot is not text. it is a grid of pixels. when you hand the model a picture of words, it has to look at that grid and infer which shapes are which letters before it can even start on your question. that step is the same job your phone does when it reads a photo of a receipt, and it is never perfect.

most of the time the model gets close. but close is where the trouble lives. a lowercase l reads as a 1. a comma reads as a period. a zero reads as an O. in ordinary prose you would never notice, because context papers over the gap. in a table of numbers or a price or an account id, one misread character changes the answer and you have no way to catch it.

why does it cost more too?

there is a second, quieter cost. an image gets processed as a block of visual tokens, and a screenshot of a full page can cost noticeably more than the same words pasted as plain text. you are paying extra to get a less reliable read. that is the worst trade in the whole workflow. if you want the intuition behind the pricing, i broke it down in why images cost way more than text in ai.

none of this shows up on your screen. the output looks complete and confident, which is exactly why the habit survives. the model does not tell you it squinted at a blurry 3.

what to do instead

• select and copy the real text. if you can highlight it, copy it. paste the words, not a picture of the words.
• if it only exists as an image or a scan, convert it first. run it through ocr so the pixels become real characters, then paste that. more on this in ocr for ai, explained simply.
• for anything with numbers, always use text. tables, invoices, ids, prices. this is where a misread hurts most and hides best.
• then feed clean text, not a wall. once you have the words, structure them. the pillar on markdown vs pdf for claude covers why clean plain text reads best.

when is a screenshot actually the right call?

screenshots are not the enemy. they are the wrong tool for text and the right tool for everything that is genuinely visual. if the picture itself is the point, send the picture.

send a screenshot when you want a chart or a graph described, because the shape and the trend are the information. send one when you are showing a bug in a user interface, because the layout is what you are asking about. send one of a photo, a design, a diagram, a handwriting sample. in all of those the image carries meaning that no amount of copied text would capture. the rule is simple: if the words are the point, send words. if the visual is the point, send the visual. and if you are pasting a picture of text mostly because you are getting vague answers from your ai, clean input is usually the real fix.

faq

does claude read screenshots of text at all?

yes, it usually can. the problem is not whether it reads them, it is that reading pixels is less reliable than reading real text and costs more. it works until the one character that matters gets misread.

what if the text is inside an image i cannot select?

run it through ocr to turn the picture into real characters, then paste that text. the model can often transcribe an image for you directly, but verify the numbers before you trust them.

is it ever fine to just paste the screenshot?

yes, when the visual itself is what you are asking about. a chart to describe, a ui bug to diagnose, a photo or design to react to. for plain text, copy the words instead.