Open source MCP for Cursor: demonstrate a real Windows or macOS process, save a standard Skill (intent + inputs), then replay in a new chat with tools you already have — not a mouse-coordinate macro.
Flow: /learn → Start Learning (on-screen HUD) → do the work → Stop (Ctrl+Shift+L / Cmd+Shift+L) → Save Skill to ~/.cursor/skills/
Secrets are redacted. Learn Mode does not execute the recording.
This is a really interesting way to think about Skills. For non-technical users (like me), demonstrating a workflow feels much more natural than trying to write detailed instructions for an agent. Curious how well the generated Skill handles situations where the workflow changes slightly the next time it runs 
That’s the right question — and why Learn Mode is not a click recording.
The saved Skill stores intent: which apps, named inputs (email, issue title, file path, …), and a step outline. It does not store mouse coordinates. Replay is a new Cursor Agent chat that reads SKILL.md and uses the desktop / browser / terminal tools already in the session. If a button moved, a window resized, or a label is slightly different, it should still find the control by name/role instead of replaying last pixel position.
It gets weaker when the process itself changes: a new required field, an extra confirmation dialog, a redesigned page. Then the original Skill can miss that step. That’s what Teach Another Example is for — a second demonstration with different values, then the shared structure is merged.
If the UI is ambiguous, the agent is supposed to stop and ask rather than guess a dangerous click.