Broll Capture

B-roll & Screenshot Capture

Three lanes, in order of effort. Real captures read as proof; recreations read as ads — always prefer capturing the actual site.

Every capture, whatever the lane, ends the same way: re-encode with dense keyframes (-g 30 -keyint_min 30) so seeks don't freeze in the renderer, then frame-extract and LOOK at every take before using it.

Lane A — Full-page screenshots (shot-cards, docs, blog "proof" beats)

A1. Playwright MCP (zero code — use it if the client has it connected)

If the Claude Code session has a Playwright (or browser) MCP available: navigate to the URL, wait for load, take a full-page screenshot, save it into assets/shots/. Good enough for most shot-cards. Take at the largest viewport available and capture more height than you need — crop later.

A2. Scripted (bundled): scripts/capture-shots.mjs

Puppeteer script — edit the TARGETS list (name, url, capture height) and run node scripts/capture-shots.mjs. What it encodes:

• Real desktop UA + 1440×900 viewport at deviceScaleFactor: 2 (crisp 2x shots).
• A widget killer injected twice (cookie/consent/chat/intercom/banner/dialog selectors hidden with display:none !important) — once after load, once after the scroll pass, because chat widgets re-mount.
• A lazy-load scroll pass (step to the bottom, settle, back to top) so images and deferred sections are actually painted before the shot.
• Clipped full-page screenshot (clip to the min of wanted height vs body height).

Review every PNG. Common junk that survives: keychain popups (from a Chrome-for-Testing profile — use system Chrome + a fresh userDataDir + ignoreDefaultArgs:["--enable-automation"]), region banners, half-loaded heroes.

Lane B — Scripted scroll recordings (clean site B-roll, 3–6s clips)

The branded-ad-edit skill's references/design.md §3 carries the full recipe: kiosk system-Chrome, scripted eased scrolling (never native jerky scroll), recorded or screencast, cut into 3–6s clips. Key gotchas:

• Scroll the smallest matching node — a text anchor can match an ancestor div whose scrollTop never moves (the scroll silently no-ops).
• Sweep junk selectors on an interval during the take.
• Frame-extract-review EVERY take.

Lane C — Screen Studio / real-browser recording (highest fidelity, or when automation is blocked)

Some sites (Cloudflare and friends) block kiosk/fresh-profile Chrome outright. The proven fallback: record the user's REAL logged-in browser.

1. Drive the real browser — a browser-extension MCP (e.g. claude-in-chrome) or the user by hand — through scripted, eased scrolls of each target page.
2. Sync markers: flash the page pure white for ~0.3s before each take (document.body.style.filter or an injected overlay). Markers make take boundaries findable in one pass.
3. Record the display with Screen Studio (screenstudio-cli if installed — check screenstudio --help) or any screen recorder. Screen Studio's polish (smooth cursor, auto-zoom) is why it's worth using when available.
4. Find the markers afterwards: invert the video and run black-frame detection — ffmpeg -i cap.mp4 -vf "negate,blackdetect=d=0.1:pix_th=0.10" -f null - — each detected "black" run is a white flash; cut takes between markers.
5. Crop off the chrome: browser toolbar and the Dock/menu bar must go — e.g. a 1920×1080 display capture cropping to 1920×816 at y=176 keeps only the page. Compute your crop from a test frame, verify by eye.
6. Re-encode dense keyframes; clip into 3–6s beats.

Traps learned the hard way

• CDP on :9222 may not be the user's browser. Screen Studio runs its own embedded Chromium that can occupy the debug port (UA ScreenStudio/…, Target.createTarget unsupported). If an automation script suddenly sees a logged-out web, check WHOSE browser it attached to.
• Inner-container scroll areas (docs sites) often defeat page-level scrolling — takes come out static. Target the container, or skip that page.
• Never enter credentials for the user; if a login wall appears, hand the browser to them and resume after.

Lane D — Public social posts and reaction GIFs without a login (2026-09-28)

• X posts render through the public embed page https://platform.twitter.com/embed/Tweet.html?dnt=true&id=<id>&lang=en&theme=light&hideThread=true (Puppeteer, 420 px viewport, device scale 2 so the text wraps large enough for a phone). Clip from the User-Name row to the first tweetText (the last one is a quoted post). Triage a post before capturing it with https://cdn.syndication.twimg.com/tweet-result?id=<id>&token=a: text, likes, the parent post, photos[] urls and video.variants[] (pick the highest bitrate; portrait 1080×1920 variants exist for phone-shot posts and drop straight in as takeovers).
• Reaction GIFs: GIPHY's published beta key is banned and Tenor's API is discontinued. giphy.com's own search calls api.giphy.com/v1/gifs/search with a key visible in its page requests; read it from a headless page (page.on("request")) and search from there (rating pg-13), then loop the mp4 to the beat with -stream_loop. Prefer brand-, artist- and creator-made loops over film and TV frames, and check the loop's last frame for a channel watermark before using it.
• Scroll recordings: when Playwright's browser is missing, Puppeteer's page.screencast() at a 540×960 viewport with device scale 2 gives a clean 30 fps portrait recording; re-encode with a dense GOP before the comp.

Where captures go

assets/shots/*.png for stills, media/broll/*.mp4 for clips. In compositions, stills usually display as a sharp card floating over a blurred blow-up of itself, or inside a browser-chrome frame (see branded-ad-edit references/design.md).

More traps (profile captures, 2026-09)

• A TypeScript runner can poison page.evaluate. tsx/esbuild injects a __name helper into any function it serialises; inside the page that helper does not exist and the evaluate throws __name is not defined. Pass page scripts as strings, not functions.
• Some platforms never render their grid for automation (TikTok: "Something went wrong", survives Refresh clicks and long waits). The profile header still renders — take the still, and carry the scrolling beat with another platform's grid.
• A screencast that "worked" can be static. Instagram's grid lives in an inner scroll container; page-level scrollTo no-ops and every frame is identical. Frame-tile every take before using it — the tile is how this was caught.
• Check the handle before blaming the tool. Two 404s were a handle mismatch (the creator's handle differs by a dot between platforms), not blocking.
• An authenticated dashboard capture can violate a client rule. An Ads Manager screenshot shows real account IDs and campaign names; treat it as unusable unless cropped to chrome.