Copy the text out of a screenshot

Yes, and it is the fastest route. Take your screenshot to the clipboard as usual, click this page and press Ctrl+V or Cmd+V. The image goes straight into the tool without ever being written to disk, which also means there is no stray screenshot file left in your downloads folder afterwards.

Because the text was drawn by a computer rather than captured by a lens. There is no motion blur, no perspective distortion, no uneven lighting and no paper texture, and the characters sit on clean pixel boundaries. Recognition quality on screenshots is usually far better than on a photograph of the same words.

Usually well, because both tend to use clear monospaced fonts at high contrast. Punctuation is the weak point: brackets, quotes and underscores are the characters most often confused, so check any command or code path before you run it. Long unbroken strings such as hashes are the least reliable.

Yes. Light text on a dark background is handled, though results are often slightly better from a light-mode screenshot because the engine was trained mostly on dark-on-light documents. If a dark screenshot reads poorly, retaking it in light mode is usually the quickest fix.

It can, and it shows you its working before you use it. When the words on the page line up into columns the tool offers a grid, displays it, and states what share of the lines actually fitted - because this is worked out from where the words sit, not read from the image as a table. A neat spreadsheet with values in the wrong cells looks exactly as convincing as a correct one, and the screenshots people bring here are usually of numbers, so the grid is something to check against your original rather than to take on trust.

They survive if you pick the right language, and are destroyed if you do not - so this is the one setting worth getting right. The English model's output alphabet holds 112 characters and exactly one is an accented letter, e-acute; it has no way to write a-acute, n-tilde, c-cedilla or u-umlaut at all, so a screenshot of French, Spanish, Portuguese, German or Italian read with it comes back with each of those replaced by whichever plain character looked closest. The confidence score is not a reliable warning: measured against the shipped model, a page of accented text scored 83 where its unaccented equivalent scored 91 - a drop, but not an alarming one, because the engine is confident about the wrong character it chose. Choosing the matching language loads a model whose alphabet does contain those letters, which is why the selector is at the top of the page rather than hidden in an advanced section.