Product Video Capture

Produce reusable clean masters first; derive presentation from them later. Preserve the raw master and camera config so every camera choice is reversible.

Prerequisites And Repo Discovery

• Require Linux Xvfb, Chrome for Testing, Playwright/CDP, FFmpeg/FFprobe, a Kandev checkout with E2E fixtures, and the landing repository.
• Resolve the Kandev root with git rev-parse --show-toplevel, then verify it contains scripts/dev-isolated and apps/web/e2e/.
• Resolve the landing root from KANDEV_LANDING_REPO when set and verify it contains both scripts/product-loop-camera.mjs and scripts/product-loop-encoder.mjs. If the variable is unset or its path fails verification, search the available workspace and sibling checkouts for both marker files; do not select a directory by name alone.
• If discovery finds zero or multiple landing candidates, ask the user to identify the checkout. Record the resolved roots as KANDEV_REPO and LANDING_REPO, and use those variables for every command and copy operation.

Choose Deliverable

Pipeline

1. Always invoke /product-demo-seeding before this skill. For alternate framing, it validates provenance and isolation instead of manufacturing new visible state. An approved raw is reusable only when its source SHA still equals freshly fetched origin/main; otherwise recapture. Then resolve and verify KANDEV_REPO and LANDING_REPO as described above. Do not assume task-specific absolute paths.
2. Create a unique writable CAPTURE_ROOT, for example with mktemp -d "${TMPDIR:-/tmp}/kandev-product-capture.XXXXXX". Use it for every raw, proof, config, and staged delivery path; do not write candidates into production assets.
3. Review the seed handoff and rehearse the full native desktop/mobile story once. Reject stale UI, fixture accumulation, duplicate tasks, selector drift, or a theme that differs from the requested product surface before starting the recorder.
4. Reset to a fresh per-take database or deterministically clean all rehearsal-created state. Record one continuous take as an unzoomed high-resolution master per form factor. Use true physical pixels, not a padded Playwright video canvas.
5. Record semantic action timestamps, target bounds, target glyph bounds, and dense pointer/touch metadata beside the raw file. Include each intentional movement's start, intermediate samples, arrival, pointer glyph bounds, and visibility interval.
6. Stop recording before capturing the clean poster.
7. Inspect raw frames before post-production. Reject UI bugs, padding, double cursors, fixture text, dead waits, and unreadable states.
8. Build a smooth, center-biased post camera from semantic events. Ignore micro-jitter, but keep every intentional pointer/touch journey inside the tested safe frame. For long travel, widen, pan with the active pointer, then tighten after arrival; complete menus, provider panels, sheets, and dialogs have framing priority.
9. Encode WebM, MP4, and WebP through landing's tested camera/encoder scripts.
10. Review fixed-fraction frames and playback on desktop, native mobile, and reduced motion. For landing films, also review at the actual theater width; full-resolution frames alone do not prove readable marketing media.
11. Keep all candidates staged outside production. Promotion is a separate, explicit action after approval; a capture request alone does not authorize overwriting landing assets.

Read capture-pipeline.md before recording and camera-encoding.md before conforming media.

Non-Negotiable Capture Properties

• Fresh isolated data; no main instance, credentials, database, or production ports.
• Separate desktop and native-mobile scripts. Never crop desktop footage into a mobile deliverable.
• Exact profiles: desktop source 3840x2400 at 25 fps delivers 1920x1200 at 25 fps; native mobile source and delivery are 1290x2796 at 25 fps.
• Raw master is one continuous take at 1x with no internal cuts, speed ramps, or audio. It also has no body transform, camera, crop, or concat.
• Disable OS cursor in X11 capture; show one intentional high-contrast DOM cursor/touch treatment.
• Use real UI input and retain dense semantic cursor/touch samples, target bounds, independently measured target and pointer glyph bounds, and motion-inclusive visibility intervals for camera design and audit.
• Browser chrome absent; product fills the physical frame.
• Any responsive/product defect remains visible or blocks capture. Do not hide it with capture CSS.

Camera And Delivery

Use $LANDING_REPO/scripts/product-loop-camera.mjs and $LANDING_REPO/scripts/product-loop-encoder.mjs. Their tests define dimensions, frame rate, maximum zoom, smoothness, loop reset, codecs, and poster quality. For semantic landing desktop stories, use formFactor: "landing" with cameraProfile: "landing-editorial" and cap maximum zoom at 2.0x. The landing-editorial profile requires both focusTrack and pointerTrack; configs without either are invalid. Native mobile stays native-size and uses the tested mobile cap, never more than 2.0x. Change these contracts test-first when the user requests a genuinely different delivery format.

Do not remove waits with cuts or speed changes. Improve the source choreography or recapture. One trim at the beginning/end is acceptable; time skips are not.

The camera is center-biased, not cursor-naive: follow the active semantic target while retaining the context needed to understand the action. Use one smooth establishing tighten, then keep a stable working depth. Widen-pan-tighten for long journeys. Keep the full dialog or menu visible as a priority, even when that means less zoom. The visible cursor never leaves the frame, including its complete rendered glyph and safety margin.

Explicit rejection rules:

• Reject lazy global zoom that holds one crop regardless of the active target.
• Reject zoom breathing: repeated in/out depth changes around one subject or state.
• Reject any zoom or camera move away from the active cursor while it is traveling.
• Reject stale UI, stale builds, accumulated rehearsal data, or captures from a non-current source checkout.
• Reject wide shots that make product text unreadable at actual size.

Acceptance Gate

Follow qa-checklist.md. Do not ship until:

• raw and delivery dimensions are proven by decoded pixels, not only container metadata;
• cadence is constant 25 fps, timestamps have no gaps, the FFmpeg log proves zero duplicated and dropped frames, and no audio stream exists;
• 10/25/50/75/90% frames, an overview contact sheet, actual-size 100% frame proofs, and full playback pass visual review;
• camera motion is smooth, reaches intended depth, and uses the profile's tested loop frame: centered 1x for a wide loop, or identical first, settled penultimate, and final crops for an approved focused landing/docs loop that improves readability while preserving identifying context;
• text, menus, diffs, pointer, and touch targets remain inside frame; frame-by-frame pointer containment passes with a deliberate edge margin;
• staged WebM, MP4, WebP, responsive source selection, lazy loading, and reduced-motion behavior pass codec probes and real-browser playback;
• SHA-256 hashes cover raw masters, camera configs, semantic metadata, proofs, and every delivery candidate;
• provenance records Kandev and landing commit SHAs, commands, source/delivery profiles, isolated ports/display/browser profile, and deterministic seed identity;
• teardown proves all capture processes, ports, X displays, browser profiles, temporary specs, databases, and temp data are gone.

Report story, seed, form factors, raw/delivery dimensions, durations, codecs, camera profile, output sizes/hashes, visual audit, browser checks, teardown, and any unsupported or blocked surface.