MankhongGarden

Windows Computer Use Notes — AI skill for Claude Code

AI community

Field notes for AI agents driving Windows apps on a shared desktop: act on minimized windows invisibly, stop typed text leaking, non-Latin keyboard layouts, Thai OCR.

How to install Windows Computer Use Notes

This entry records only its repository, not the path inside it, so there is no exact command to give. Open MankhongGarden/windows-computer-use-notes and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Windows Computer Use Notes does

Field notes for AI agents driving Windows apps on a shared desktop: act on minimized windows invisibly, stop typed text leaking, non-Latin keyboard layouts, Thai OCR.

Alternatives in AI

Guizang Ppt Skill — AI-agent Skill for generating polished HTML slide decks: editorial magazine and Swiss layouts, image prompts 24.9k ★
Azure Vision Java — Image captioning, OCR, and object detection 1.8k ★
Agentboard — Web GUI for tmux optimized for AI agent TUIs, with support for iOS safari and mac w/ keyboard shortcuts 410 ★

README

Windows computer-use for AI agents: four problems the popular tools don't handle yet

**Field notes · partly day one.** Section 1 (minimized windows), 2 (typing leaks), 3 (keyboard layouts) and 4 (Thai OCR) were built and tested on **2026-10-30**. The rest of the tool has been in daily use since 2026-10-07. Tested on one machine: Windows 11 build 26300, Python 3 + `ctypes`, 125 % DPI, apps from classic Win32 to UWP and Chromium. Expect rough edges; this README will be updated after two more weeks of real use.

I run several AI coding agents (Claude Code) on one Windows laptop, and they need to click and type in desktop apps **while I keep working on the same screen**. I built a small private MCP server for that. Before adding more to it I compared it with the two open-source tools most people use:

	trycua/cua (cua-driver, Windows backend)	CursorTouch/Windows-MCP
Minimized target window	Refuses: `cannot capture minimized window` (`platform-windows/src/capture.rs`, `wgc.rs`)	Restores it with `SW_RESTORE` and switches to it (`desktop/service.py`) — the user's screen changes

Both are good projects with far more scope than mine (Cua runs whole sandboxed VMs on three OSes). These notes are only about four details that matter when **an agent and a human share one desktop**. The code here is short illustrative `ctypes`; the tool itself stays private.

1. Act on a minimized window without the user seeing it

Agents are usually asked to work in an app the user has minimized. Restoring it steals the screen; refusing it means the user has to babysit. UWP apps unload their UI while minimized (their UIA tree is just the frame; File Explorer also needed the trick below in my tests), and a minimized window has no rendered content to capture.

What works: make the window fully transparent, restore it **without activation**, do the work, minimize it again.