Dataset prep for vision APIs is a repetition problem. Claude wants 1,568 px on the long edge unless a request carries more than 20 images, in which case nothing above 3,000 px per side. OpenAI high detail wants 2,048 px and rejects anything that balloons past its patch budget. Gemini tiles by dimension and lets media_resolution cap the token spend. Whatever number you pick, the expensive part is applying it consistently to a few hundred files without a desktop app, a Python environment and an afternoon. Doing it by hand also hides the failure modes: a folder where half the images were already small enough and got upscaled, a couple of animated GIFs that turned into one frame, a handful of HEIC originals the browser quietly skipped. Skwish takes the whole folder at once, applies one output format and one size limit across the batch, shows a per-file before and after readout so the odd file out is visible, and hands back a single ZIP with consistent names. Everything runs in the browser, so a private screenshot set never leaves the machine.