Capture the image properly
Fill the frame with the text. Cropping later throws away the pixels you needed; getting close in the first place keeps them.
Hold the camera parallel to the page. Perspective distortion makes the far edge of the page smaller than the near edge, and letters near the far edge start being misread first.
Use diffuse light. A window on a bright but overcast day is close to ideal. Direct overhead light creates glare on glossy paper and hard shadows from your own hand.
Prepare the file
Save as PNG for screenshots and text-heavy images. JPEG compression smears the edges of characters, which is exactly the information recognition depends on.
Crop away anything that is not the document. Desk edges, keyboards and hands confuse layout analysis, which may treat them as figures or as text regions.
If the image is rotated, rotate it before uploading. Small skews are corrected automatically; a sideways page often is not.
Run the conversion and read the output critically
Upload the file, choose the expected language, and run the extraction. Copy the result into an editor before you start correcting.
Check three categories first: numbers, proper nouns and anything in an unusual font. These are where the language model has least ability to help, so they carry most of the remaining errors.
Common confusions to scan for are 0 and O, 1 and l and I, 5 and S, and rn read as m. A find-and-replace pass over these pairs takes a minute and catches most of the damage.
A realistic example
A screenshot of a terminal window at 1920 by 1080, monospaced font, dark background: this converts essentially perfectly, including punctuation, because contrast is high and the font is designed to be unambiguous.
A phone photo of a conference handout in a dim room, taken at a 30 degree angle: expect several errors per paragraph, mostly in the smaller footnote text. Retaking the photo in better light usually costs less time than correcting the output.