Claude's image rules are more specific than most people expect, and they interact. On claude.ai you can attach JPEG, PNG, GIF or WebP files up to 10MB per image, with a maximum of 8000x8000 pixels per image and up to 20 images per chat. Claude's own help center adds a floor rather than a ceiling: use images that are 1000x1000 pixels or larger, because small blurry images are what produce bad answers. Then there is the token bill, which is the part nobody checks. Claude sees an image as 28x28 pixel patches, so an image costs about ceil(width/28) x ceil(height/28) visual tokens — a 2000x1500 photo lands around 1,564 tokens on a standard-tier model, and a 4000x3000 export costs more than three times that for detail the model discards. The docs are blunt about the fix: resize images before uploading to control token cost. Skwish compresses and resizes images locally in your browser with a live before/after readout per file, so you can sit in the readable-but-not-wasteful band. Batch the set, export a ZIP. Nothing is uploaded to a server.