The day I shipped this started with me yelling at Claude Code for the fifth time. I'd pasted a screenshot of a misaligned form. I'd typed "fix this." Claude moved the wrong input. I retyped. Claude moved a different wrong input. I gave up and fixed it manually. The reason it kept guessing: it was reading raw pixels. It had no way to know which rectangle was the input I meant, so it picked one that looked plausible. SlimSnap converts the screenshot into a spec the AI can parse element by element. Each element has coordinates, OCR text, color values, and (if you drew an arrow on it) a target reference saying "this one." It also happens to be ~700 tokens versus the 1,568 raw screenshots cost on Sonnet (up to 4,784 on Opus 4.7+). That part is just bonus. Open: the JSON schema (MIT, github.com/bickov/slimsnap-schema) and a Claude Code skill that auto-loads your latest capture (MIT, github.com/bickov/slimsnap-skill). The Mac app is closed but free. Other tools (Cursor, Lovable, bolt.new, Replit, ChatGPT Vision): the spec works, but you paste the JSON into chat yourself. Cleaner than raw images. Not as smooth as the Claude Code auto-loader. Someone with time on their hands could write the equivalent skill for any of them. A real question: which AI tool do you reach for most when you need to point at something specific on screen? Tells me where to build the next auto-loader.

SlimSnap

Your AI doesn't know which button you mean

Your AI doesn't know which button you mean