screen-recording

Script-driven product capture — iOS Simulator, website, or Figma prototype — with auto-zoom and cursor navigation derived from measured interaction times, never hand-guessed. The synthetic cursor comes from the compositor, so it works identically on every surface.

The user's script must specify, per scene: start state, the interaction to show, the payoff to read, and the data that must exist. If it doesn't, get those four fields per scene before capturing anything (the companion product-film skill ships a full fill-in template in its references/writing-the-script.md).

Read references/capture-protocol.md before rolling anything. It has the staging checklist, the per-take protocol, and the derivation rule.

COMPANION for editing/compositing: the product-film skill supplies the downstream tools (diffscan.swift/diffscan2.swift for measuring takes, composite2.swift for rendering zoom + cursor). This skill produces their inputs; capture itself works standalone.

The loop

1. Parse the user's script → shot list (start state, actions, end state, readable moments). Seed any missing product data first.
2. Pick the capture surface (protocol §0: simulator / website / Figma proto — each has its drive+record pair) AND the presentation path by form factor: portrait/phone content → mobile path (mockup, dot cursor); landscape/desktop content → desktop path (FILL=1 full-bleed, arrow cursor, capture must be canvas-aspect 16:9). Stage it: sim = locale+status-bar+privacy+burned dialogs; web = logged in, banners dismissed, viewport sized to the aspect; proto = presentation mode, Fit zoom.
3. Per scene: dry-run for coordinates → roll detached to a durable dir → drive with deliberate pacing → SIGINT → verify the end frame.
4. Measure with product-film's diffscan (mobile) / diffscan2 (desktop) → real interaction times.
5. python3 scripts/autoplan.py taps.json (add --desktop on the desktop path) → composite2 camera plan + cursor/ripple args (zoom punches arrive 0.3s before each tap; cues complete before the UI transition).

Non-negotiables

• No take before staging is complete and one-time dialogs are burned.
• Dry-run before every recorded flow; verify field focus before typing.
• Desktop/web film takes: sckrecord (ScreenCaptureKit) only — never screencapture -v — with the mandatory pre-roll health gate (protocol §0).
• One pointer contract per film (synthetic arrow OR recorded cursor), decided before roll 1 and honored in every take and recapture; same stage (window, region, scale, recorder) for every act.
• The recorded viewport must contain every payoff the script promises — verified on a dry-run still of the end state, before rolling.
• Record to durable storage; a take without a verified end frame isn't done.
• Trim only from measured times — tap latency makes plans fiction.
• A contaminated take (dialog, toast, wrong state) is discarded and re-rolled, never rescued in the edit.
• Pass the udid explicitly on every drive command when more than one simulator exists.