Screenshots are the most common thing people feed an AI chat and the worst case for blind compression. A Retina screenshot of a full browser window is 2880x1800 and 3-6MB as PNG, most of it flat UI colour and crisp small text. Compress it the way you would compress a photo and the interface text turns into grey smears, which is precisely the content you wanted the model to read. Screenshots need dimension control more than quality control: downscale to a size where the smallest text you care about is still several pixels tall, then compress with a setting that keeps sharp edges. Claude's help center recommends images of at least 1000x1000 and its vision docs suggest keeping dimensions under 2000px when sending many images at once; Gemini's tiling means a screenshot that lands in one 768px tile costs a flat 258 tokens while a 4K capture costs six or more tiles for the same layout. And every one of these platforms has a per-image byte cap that a full-page capture can quietly exceed. Skwish compresses and resizes images locally in your browser with a live before/after readout per file, so you can check both dimensions and bytes on the same screen. Batch a run of screenshots, export a ZIP. Nothing is uploaded to a server.