Windows Automation

Context

Use this skill with a configured Windows MCP Server in any compatible host. Let Windows expose control names, state, and IDs instead of guessing from screenshots. The skill itself does not install or configure the server.

Preferred workflow

1. Use window_management to find or activate the target window.
2. Discover with ui_find or ui_snapshot; pass the returned IDs to ui_read, ui_click, and ui_type.
3. Use ui_read_table to extract a grid, table, or details-view list into structured rows + headers in one call instead of scraping cells with repeated ui_read.
4. Use file_save for supported Save / Save As dialogs and file_open for Open flows. After failure or confirmation_required, inspect the existing dialog and choose the next action explicitly. These helpers do not approve overwrites, dismiss errors, or retry dispatched actions.
5. Use clipboard (get/set/clear) for fast bulk text IO — pair it with copy/paste hotkeys.
6. Use ui_batch to run a multi-step sequence in one call. Keep reusable steps in project files.
7. Only fall back to screenshot_control, mouse_control, or keyboard_control when the UI Automation tree is missing or the target is a custom canvas.

Patterns

Semantic-first automation

• Discover using names, control types, and automation IDs; act using the returned opaque elementId.
• Click, double-click, type, select, table reads, and state waits require IDs. Selection values still identify option text, while the ID identifies the containing control.
• Selectors are only for discovery and appear/disappear waits. Removed action selectors are errors.
• Read an element with its ID; omit the ID only for an explicit whole-window read. Element reads never widen to window OCR.
• Stale IDs require rediscovery. IDs belong to one owner and never transfer between MCP instances or the CLI daemon.
• Batch $prev requires one unambiguous immediately preceding result. Reusable batches should discover fresh controls on every run.
• Each batch step requires its matching enabled tool (key: keyboard_control; mouse/polyline: mouse_control). Forbidden steps reject the whole batch before it starts, even with stopOnError=false. withSnapshot=true requires enabled ui_snapshot before any action. Never route around disabled tools.
• Re-check the UI tree after dialogs, page changes, or tab switches.
• Treat screenshots as discovery or fallback tools, not the primary control surface.

Token-efficient observations

• For repeated checks, use ui_snapshot(mode='auto') on the first and later views. A default full view does not establish the remembered comparison.
• Use mode='reset' to replace an older comparison. Scope to a known subtree when only that part of the window matters.
• Use withSnapshot=true and snapshotMode='auto' on supported actions to combine action and observation. A full response is expected when a diff is unsafe or too large.
• Prefer a table read or article read when that is the information needed; do not repeatedly request whole-window trees or images.

Outcome verification

• Click success reports dispatch, not a verified application-specific outcome.
• Read the result or use a bounded wait for the expected state.
• Do not repeat a click just because its target disappeared, became disabled, or changed its label. It may already have completed.