ThioJoe/Thio-Universal-Agent — reverse-engineered prompt

Reverse engineered prompt

GitHub

Build me a Windows desktop AI assistant that can look at my screen and help me use any app the same way a person would. I want it to work visually, by understanding screenshots and then either telling me where to click and what to type, or optionally moving the mouse and typing on its own.

Make it a single portable exe with no installer. The safe default should be human control only, where it draws clear markers on the screen, gives step by step instructions, and lets me copy suggested text. Add a setting to switch to autonomous mode, plus global hotkeys to pause, resume, or stop it immediately.

It should support common AI providers like Gemini, OpenAI compatible APIs, Claude, and local models where possible, using my own API keys. Include settings for max steps, context handling, import and export config, and a way to redirect the agent mid task with new instructions.

Keep the app focused on Windows, simple to run, and useful for troubleshooting, finding things on screen, or doing one off computer tasks.

Want more depth? Deep Reverse