lpm: Skill for Claude Code

App Review asks for a video of lpm Link on a physical device connected to the Mac (the earlier one: unlisted YouTube, linked from App Review Information → Notes). This skill makes it: the left half is the lpm desktop app, the right half is the real iPhone through iPhone Mirroring, and every tap and click is a real one made with computer use.

scripts/review.js runs the session (node scripts/review.js prints the commands):

• setup builds (first run, or --build after desktop changes) an isolated debug app named lpm Review (bundle cx.lpm.appreview, so computer use can see it and never touches the user's own lpm), wipes and seeds ~/.lpm-review with three demo projects under /Users/Shared/lpm-review (storefront with api :4000 and web :5173, payments-api, docs-site), turns remote control on (port 8792), names the Mac MacBook Pro (--name; no owner name, no dev marker), neutral shell prompt, and places its window to the left of the iPhone Mirroring window, as tall as the phone.
• record starts a ScreenCaptureKit recording of the rectangle around both windows, with every other app left out (notifications, the user's windows and dialogs never show) and the pointer drawn. It also logs every mouse press, release and move and the moment of each key press (never which key) to events.jsonl, through a listen-only event tap. It returns at once; the capture keeps running.
• mark "<caption>" "<subtitle>" starts the next numbered step caption.
• checkpoint before an uncertain action; cut if it went wrong drops everything since the checkpoint from the video. Redo the action after the cut.
• stop, then render --device "<model>" [--build <n>] [--version <v>] [--cut a-b,…] [--redact <regex>] [--no-camera] writes ~/Movies/lpm-app-review/<stamp>/lpm-link-<version>-app-review.mp4 and steps.txt (where each step starts). The version defaults to mobile/ios/project.yml. The render:
  • keeps each click's response (the screen changes until the next click, at most 12 s) with a short hold, and drops everything between: waits, a server's log, a clock (edit.js);
  • draws a ripple at every click;
  • eases in up to 1.8x on the part of the phone a tap changed (a menu, the message field) when the response stays on the phone, and back out when it reaches the Mac or the next click is on the Mac; the header then names only the device in view (camera.js);
  • reads the kept frames twice a second with Vision and blurs any line in someone's possessive form ("Someone's MacBook Pro"), any email address and any --redact pattern, for as long as it is on screen; the review Mac's own name stays readable (redact.js);
  • with no events.jsonl (an older take) it cuts on screen changes alone.
• teardown quits the review Mac (the phone hears it close), stops its servers and deletes its data and demo projects.

Before the take

1. The iPhone has the build under review installed (Xcode or TestFlight) and is near the Mac, locked, on the Mac's Wi-Fi. Ask the user for the device model and build number for the title card if they matter.
2. request_access(["iPhone Mirroring"]), open_application("iPhone Mirroring"), screenshot. "iPhone in Use" means the user has the phone: ask them to lock it. Bring lpm Link to its first screen ("Connect to your Mac"): ⌘2 (App Switcher), swipe its card up, ⌘3 (Spotlight), type lpm Link, open the top hit. If it lists saved Macs instead, ask the user before removing any.
3. node scripts/review.js setup, then request_access(["cx.lpm.appreview"]) and a screenshot: lpm Review on the left with the three projects, the phone on the right.

The take (record, then for each step mark first, then act). Call open_application for the app you are about to click before every batch: a screenshot hands focus to another granted app, and the click then lands in the wrong one.