windows-use
Skill2 starsChanged 4 months agoMITView on GitHub
Control approved Windows desktop applications through CodeFlowMu Windows Use tools. Use when a Cursor Agent must inspect, click, type, scroll, invoke UI Automation controls, or capture a window in a native Windows app; do not use for browser-only work when a browser-specific capability is available.
In this repository42 other files it gives its agents.
- architect-boundary-reviewSkill
- architect-decision-recordSkill
- architect-system-designSkill
- browser-playwright-checkSkill
- browser-useSkill
- code-search-navigationSkill
- controlled-emergence-observerSkill
- dev-code-locationSkill
- dev-small-scope-patchSkill
- dev-test-and-explainSkill
- eval-observation-writingSkill
- eval-promotion-adviceSkill
- eval-risk-gap-analysisSkill
- ops-log-diagnosisSkill
- ops-runtime-healthSkill
- ops-stuck-workflowSkill
- pm-acceptance-criteriaSkill
- pm-architecture-reviewSkill
- pm-delivery-planSkill
- pm-evolve-skillsSkill
- pm-long-horizon-planningSkill
- pm-priority-triageSkill
- pm-product-design-briefSkill
- pm-product-requirementsSkill
- pm-scope-controlSkill
- pm-solve-problemsSkill
- pm-tech-scopeSkill
- qa-regression-checkSkill
- qa-reproduce-issueSkill
- qa-verify-fixSkill
- run-local-command-checkSkill
- test-selectionSkill
- tm-delivery-governanceSkill
- tm-review-routingSkill
- tm-risk-dependencySkill
- ui-information-architectureSkill
- ui-requirementsSkill
- ui-usability-acceptanceSkill
- ui-visual-consistencySkill
- web-extractSkill
- web-researchSkill
- web-searchSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
17 rules. About 0 are common, 6 are distinctive, and 11 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
11
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 12
Call windows.capabilities when host readiness is unknown.
- Very specific: line 13
Call windows.list_targets before handling login.
- Very specific: line 17
Call windows.inspect_ui when labels or standard controls can identify the target.
- Very specific: line 18
Only when an operation truly needs foreground input, call windows.activate.
- Very specific: line 19
Use windows.screenshot and window-relative coordinates only when UI Automation is insufficient.
- Very specific: line 21
Call windows.cancel when the user asks to stop.
- Very specific: line 26
Read approved targets and login characteristics: windows.list_targets
- Very specific: line 28
Wait for a launched native window: windows.wait_for_app
- Very specific: line 30
Verify/pause/resume the current session: windows.status, windows.cancel, windows.resume
- Very specific: line 32
Read control structure: windows.inspect_ui
- Very specific: line 37
Use literal text with windows.type_text.
- Distinctive: line 41
Never attempt to authorize an application through the tools.
- Distinctive: line 42
Never automate terminals, CodeFlowMu, ChatGPT/Codex, authentication dialogs, password managers, Windows security tools, the Run dialog, or …
- Distinctive: line 43
Do not bypass CAPTCHA, MFA, paywalls, safety warnings, or access controls.
- Distinctive: line 44
Ask for action-time confirmation before deleting data, sending messages, submitting forms, uploading files, changing sharing/access, …
- Distinctive: line 45
Stop if the desktop is locked, the target window cannot be activated, observation fails, or the window no longer belongs to the approved …
- Distinctive: line 46
Prefer a browser-specific capability for Chrome or web applications so logged-in browser state and page semantics remain available.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Windows Use
Control only applications the user approved in CodeFlowMu Panel. Treat application approval as permission to access that app, not blanket permission for risky side effects.
Workflow
- Call
windows.capabilitieswhen host readiness is unknown. - Call
windows.list_targetsbefore handling login. FollowloginMethod,verificationChannel,loginInstruction, andrequiresUser; never infer an authentication flow. A saved password is reported only as a boolean and is never returned. - For
qr_codeor verification-code methods, stop at the login step and wait for the user. If the method isunspecified, ask the user instead of guessing. - To open an approved target, first reuse any window returned by discovery. Call
windows.launch_targetonly when the target is not already running. Native launch is idempotent and may returnalready_running=true; never relaunch a minimized, hidden, or initializing application. Web launch only opens the URL—route all page DOM work to a browser-specific capability. - For a native target, call
windows.wait_for_appinstead of using Shell sleep, then select the returnedapp_idandwindow_id. Never invent or reuse a stale window identifier. - Call
windows.inspect_uiwhen labels or standard controls can identify the target. Preferwindows.invoke_ui: UI Automation InvokePattern works without taking foreground focus. - Only when an operation truly needs foreground input, call
windows.activate. Never use PowerShell, Alt+Tab, or a terminal to force focus. If activation is blocked,windows.clickcan deliver a bounded window-message fallback, while text/keyboard input must pause for one user click. - Use
windows.screenshotand window-relative coordinates only when UI Automation is insufficient. - Batch a small set of related actions, then inspect again to verify the result. Stop after an error or if the user takes over the foreground window.
- Call
windows.cancelwhen the user asks to stop. Claim “paused” only after it returnspaused=true. This is an MCP-session pause, not project-level disablement. Callwindows.statusto verify; callwindows.resumeonly after explicit user instruction.
Tool selection
- Discover applications:
windows.list_apps - Read approved targets and login characteristics:
windows.list_targets - Open an approved native or Web target:
windows.launch_target - Wait for a launched native window:
windows.wait_for_app - Restore and activate a window:
windows.activate - Verify/pause/resume the current session:
windows.status,windows.cancel,windows.resume - Observe pixels:
windows.screenshot - Read control structure:
windows.inspect_ui - Invoke a standard control:
windows.invoke_ui - Point interaction:
windows.click,windows.scroll - Text and keys:
windows.type_text,windows.keypress
Use literal text with windows.type_text; use windows.keypress for Enter, Tab, Escape, arrows, and chords. Focus the observed editable surface before typing.
Safety
- Never attempt to authorize an application through the tools. Ask the user to approve it in Panel.
- Never automate terminals, CodeFlowMu, ChatGPT/Codex, authentication dialogs, password managers, Windows security tools, the Run dialog, or permission/security settings.
- Do not bypass CAPTCHA, MFA, paywalls, safety warnings, or access controls.
- Ask for action-time confirmation before deleting data, sending messages, submitting forms, uploading files, changing sharing/access, installing software, making purchases, or transmitting sensitive data unless the user explicitly pre-approved that exact action and destination.
- Stop if the desktop is locked, the target window cannot be activated, observation fails, or the window no longer belongs to the approved application.
- Prefer a browser-specific capability for Chrome or web applications so logged-in browser state and page semantics remain available.
Windows Use actions bring the target app to the foreground. Keep operations bounded and allow immediate user interruption.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

