kingsahil/blinky ? reverse-engineered prompt

Build me an AI desktop tutor app called Blinky that can look at my screen, understand what app I am using, and guide me step by step in a simple visual way. I want it to feel like a helpful teacher sitting beside me, so it should read on screen text, highlight the right button or menu, and explain what to do next in plain language.

It should work offline first when possible, with local AI and screen reading, but if needed it can also use online AI for better answers. I want a mode where it just guides me, and another safer autopilot mode where it can actually click around and do the task for me after checking the screen. It should also support voice input and voice replies, and be able to open a browser or search the web when needed.

Make it a modern desktop app with a clean overlay style interface, and if you need to look up current docs online while building it, go ahead.