Your AI never sees the screenshot you paste
by•
Captured a page at 3066 x 28800, pasted it into Claude, and the model received 276 x 2600. Same shape, 11x smaller, about 99% of the pixels gone. Text that was 28px tall arrives under 3px.
That's the whole reason long screenshots get useless answers. The model isn't being lazy, it's reading a gray smear.
SlimSnap handles it in two parts. The capture gets cut into frames small enough to survive the resize, and the text goes over separately as JSON with coordinates, which has no height limit at all. A 28,000px page arrives intact.
Still building the table of what other tools do to a tall image. If you've measured Figma, ChatGPT or Cursor, I want the numbers.
45 views
Replies
Ugh—that's strange, especially since it takes more tokens to process an image. I've moved on to pasting only text and logs, even for a full site. SlimSnap is useful when I'm testing HTML outputs to see what works and what doesn't.
SlimSnap
@roopesh_donde Text only works until the bug is visual. No log is going to mention that two elements overlap.
Separate thing: that HTML-output line is better than anything on my site right now.
Can I quote you? Name and link.
@bickov Yes, go ahead.
this explains a lot of “bad AI output” that people blame on the model.
sometimes the problem is not reasoning, it’s input loss. if the context reaching the model is already compressed, cropped, or unreadable, the answer will only be as good as the damaged signal it received.
SlimSnap
@nipuntaneja And it's silent. No error, no warning, no "I received a smear." The model answers with the same confidence either way, so you go and debug your prompt when the damage happened before the model ever ran.
The part that worries me: every tool is now being asked to auto-resize oversized images so the request stops failing. That trades a loud error for a quiet one.