agentleFS
pedrojaques99pedrojaques99/visant-skills-public/site-reels/SKILL.md

site-reels

Skill81 starsChanged 4 days agoMITView on GitHub

Records the best scenes of a website or web app (WebGL hero, 3D, scroll) as 1080p MP4 for decks, uploads them to Figma Slides, and builds brand or product showreels in Remotion. Use on "record the site scenes", "site video for the deck", "showreel", "launch video", "gerar vídeo do site", "vídeo pro deck", "reel do produto", /site-reels.

  • Installs packages

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Run the scripts copied into the project: ESM resolves puppeteer-core from the script's folder, not from the cwd.

    line 45
  2. Distinctive:

    Visual identity comes from the deck, not the site: check official type and palette in the slides (sites drift from the brand book).

    line 149
  3. Distinctive:

    Check the first frame of the stretch: a shared element "flying" in shows up loose.

    line 204

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Site Reels: site → MP4 → deck

Requirements

Needed forWhatCheck
everythingWindows, macOS or Linux with a GPU and a visible desktop (Chrome runs headful)n/a
everythingNode.js 18+node -v
everythingffmpeg + ffprobe on PATH (or FFMPEG=<full path>)ffmpeg -version
everythingChrome, Chromium or Edge (auto-detected, see below)node -e "import('./scripts/find-chrome.mjs').then(m=>console.log(m.findChrome()))"
record.mjspuppeteer-core + puppeteer-screen-recorder (npm, in the project)npm ls puppeteer-core puppeteer-screen-recorder
snap.mjspuppeteer-core + timesnap (brings timeweb)npm ls timesnap
showreelRemotion projectnpx remotion --version
audits (optional)ImageMagickmagick -version
Figma uploadFigma MCP (use_figma, upload_assets) + skills figma-use, figma-use-slidesMCP tool list

Chrome detection (scripts/find-chrome.mjs, used by record.mjs and snap.mjs): CHROME env var if set; else installed Chrome/Edge (Windows: Program Files and LocalAppData; macOS: /Applications and ~/Applications; Linux: google-chrome, chromium, microsoft-edge on PATH); else the newest ms-playwright/chromium-<N> (or PLAYWRIGHT_BROWSERS_PATH); else it stops and tells you how to set CHROME or run npx playwright install chromium. On Windows it adds --use-angle=d3d11; on other OSes Chrome keeps its default GPU backend.

What it does

Records scripted scenes of a site running locally and delivers MP4s ready for Figma (1920×1080, 30fps, x264, < 100MB). Validated engine: puppeteer-core + puppeteer-screen-recorder (CDP screencast + ffmpeg). Never screen-record.

Run

mkdir <proj>-reels && cd <proj>-reels && npm init -y && npm i puppeteer-core puppeteer-screen-recorder
mkdir -p scripts && cp ~/.claude/skills/site-reels/scripts/{record,find-chrome}.mjs scripts/
cp ~/.claude/skills/site-reels/scripts/scenes.example.mjs scenes.mjs   # edit the scenes
# in the site repo: npx next build && npx next start -p 3100   (production build: smoother WebGL)
OUT="<output folder>/Videos" node scripts/record.mjs scenes.mjs [scene ...]

Run the scripts copied into the project: ESM resolves puppeteer-core from the script's folder, not from the cwd.

Env: BASE (default http://localhost:3100), OUT, CHROME, FFMPEG, W, H, DPR, MOBILE=1 (touch emulation), KEEP_UI=1 (keeps the nav), MIN_FPS (24). Phone 4:5: W=360 H=450 DPR=2 MOBILE=1 gives 720x900.

Each scene = { name, path, theme?, trim?, crf?, css?, prepare(page, ctx), act(page, ctx) }. prepare runs before recording (scroll to the section, wait for 3D to load); act is what goes into the video. ctx has wait, scrollTo(page, y|selector, ms) and scrollThrough(page, selector|null, ms, from, to) with easing. Intro that runs on load: page.reload() inside act and use trim to cut the loading.

Brand script that works: logo/hero intro (light and dark), hero media swap, 3D section with the camera entering on scroll, immersive route (walk), footer wordmark and one full page pass. 5–7 scenes of 12–55s.

Audit before delivering (do not trust the exit code)

  • Contact sheet per video: ffmpeg -i x.mp4 -vf "fps=6/<dur>,scale=480:-1,tile=3x2" -frames:v 1 sheet.png and look.
  • 1MB for 19s = blank scene. > 100MB = does not go into Figma (scene crf fixes it; 24 for a scene with grain + trails).
  • Top strip (crop=1920:120:0:0) at several times to catch a leaking nav.
  • LOW-FPS with a long pause at the start/end is a false alarm (the meter counts still frames); check the contact sheet.

Known pitfalls

  1. Hidden window freezes rAF → frozen frames. The launch already passes --disable-backgrounding-occluded-windows --disable-renderer-backgrounding --disable-background-timer-throttling --disable-features=CalculateNativeWinOcclusion. headless: false on purpose: headless falls back to SwiftShader and WebGL stutters.
  2. Hide-UI CSS disappears on reload and when React hydration clears <head> → the script re-attaches a <style id="reels-hide"> to the body every 100ms.
  3. One invalid selector kills the whole rule: \[ in a JS template literal becomes [ and .z-[60] invalidates header, nav, .z-[60]. One rule per risky selector.
  4. Do not hide [data-cursor]: sites tag the 3D canvas and carousels with it. Target the custom cursor element (fixed + mix-blend + pointer-events-none).
  5. Lazy-loaded 3D (IntersectionObserver + dynamic import): wait 5–7s after reaching the section. Reading the WebGL canvas comes back empty without preserveDrawingBuffer: use a fixed wait or a screenshot.
  6. Screencast delivers ~32 unique frames/s → record at 30fps (60 gives irregular repeated frames). Default trim 0.3s drops the first cached frame.
  7. Screencast stutters on a loaded machine (Photoshop open, parallel render): the MP4 says "30fps" with 9–18 unique frames/s. record.mjs measures (mpdecimate=hi=64:lo=24:frac=0.05) and re-records up to 2x below MIN_FPS (24). For WebGL/3D plates prefer snap.mjs.
  8. Cookie banner: accept it in prepare (generic click on "Accept/Aceitar/OK") before recording.
  9. A tool page without scrolling (generator, effect editor) records still: the scene has to interact (upload a file to input[type=file], click presets by button text). A custom slider does not respond to a synthetic input.

60fps locked plates: snap.mjs (deterministic)

scripts/snap.mjs injects timeweb (timesnap's virtual clock: rAF, performance.now, setTimeout and video), advances 1/60s per frame and screenshots. Independent of machine load. Scene: { name, path, theme, css, dur, wait (real ms to load 3D), pre (virtual ms before scrolling, 2000; **0 for an intro that runs on load**), start (virtual s skipped), scroll: { sel, from, to, at, len } | { bottom: [from, to] in screen heights }, frame(page, t) (e.g. mouse) }. Scrolling is a function of virtual time.

npm i puppeteer-core timesnap && mkdir -p scripts
cp ~/.claude/skills/site-reels/scripts/{snap,find-chrome,patch-timeweb}.mjs scripts/
npm pkg set scripts.postinstall="node scripts/patch-timeweb.mjs" && npm run postinstall
FPS=60 OUT=... node scripts/snap.mjs scenes-snap.mjs [scene]
  • Do not use timesnap() itself: with Next + R3F + Lenis it recorded a blank page and left frozen stretches on the WebGL hero (rAF does not run). The own loop (goTo + scrollTo + screenshot) worked everywhere: 60/60 unique frames.
  • timeweb patch is mandatory: it swaps video.play() for a version with no Promise and no playing event → video.play().catch() throws and the page goes blank.
  • preserveDrawingBuffer: true forced on getContext (turn off with preserve: false).
  • Frames go to the OS temp dir, never to a synced drive (thousands of jpgs syncing).
  • No GPU work in parallel (Remotion render, Photoshop): WebGL loses its context and records seconds of background only. The script warns FLASH (frame flashing by luminance); a long blank stretch only shows on the contact sheet. Look.
  • Real smoothness audit: magick compare -metric MAE between neighbor frames at full resolution. 0 = repeated frame.
  • Delivery: scripts/deliver.example.sh (copy and rename the plates; V= required) builds reveals, loops (loop.sh) and clean 3D, raising crf by 3 until < 95MB.

Upload to Figma Slides (hidden, without disturbing the team)

The MP4 does not go in through MCP: use_figma has no network and the Plugin API does not create video from bytes; upload_assets only accepts images (≤ 10MB). So:

  1. Cover of each video: ffmpeg -ss <t> -i x.mp4 -frames:v 1 -q:v 2 _capas/x.jpg.
  2. use_figma (skills figma-use + figma-use-slides): createSlideRow(idx) before the final slide, row.name = 'WIP · Motion (hidden)', one createSlide(idx, i) per video with isSkippedSlide = true, black background, a 1920×1080 rectangle named VIDEO · x.mp4 and the instruction in speakerNotes.
  3. upload_assets with the rectangles' nodeIds → multipart POST of the covers.
  4. The user drags the MP4 onto the rectangle (replaces the cover with the video). Say this clearly; never promise an inserted video.

getSlideGrid() omits skipped slides: check the structure via SLIDE_GRID.children[i].children.

Brand showreel (Remotion): plates + designed scenes

For a branding showreel (logo, colors, type, layouts, manifesto, mockups), record only clean plates from the site (no text: per-scene CSS hides title/nav) and compose the rest in Remotion (frame-by-frame render, no jitter). Reusable building blocks worth having in src/fx.tsx: Mist, MistWords, Grain, KenBurns, Plate, a nevoa() (mist) transition, with brand data in brand.ts.

  • Visual identity comes from the deck, not the site: check official type and palette in the slides (sites drift from the brand book).
  • Scenes that work: opening phrase in mist, recorded reveal, logo construction (evolvePath + guides measured on the logo raster), manifesto phrase by phrase, pillars with variable wght axis, palette in bands, typography, layered graphics, 3D with title in mix-blend-difference, scrolling layout columns, 1/2/3-up mockup montage, product synced to the name, closing on the wordmark.
  • Review by stills (npx remotion still ... --frame=N + montage) before rendering.
  • Render with --concurrency=3: at 8, with Photoshop open, a 32GB machine ran out of memory.
  • Loops: scripts/loop.sh in out [start] [dur] [crossfade]. Check the seam with magick compare -metric RMSE between first and last frame (< 0.5% = invisible).
  • Bad plate is a discard: a solid of revolution spinning shows no motion; a camera framed for a portrait canvas looks small in 16:9.
  • Frantic reel (approved rhythm): 25–35s, 60fps, Series with hard cuts on a 120 BPM grid (30 frames), image bursts every 3–6 frames, giant logo crop over a grid, flashing font weights, palette blocks rearranging, portrait in a circular mask, graphic burst, kinetic phrase, one-beat mockups, closing logo held ~2.5s. Zero labels: text only as graphics. A slow 1:40 showreel was rejected.
  • Soft/smoky reel: record the hero smoke with no logo or photo, black on pure white (scene with storage = params in localStorage and block: ['assets/logo.png']; the blank PNG must be transparent; do not block the video: a broken video freezes timeweb). Use the plate as a mix-blend-mode: multiply overlay: opens at the start of the scene, an inverted version (ffmpeg -vf reverse) closes it; the cut lands on black. Transition at ~1.1s (rate 2), or half the scene goes black.
  • Figma Slides layouts assembling: clone the slide (skipped), duplicate each card 3x with only background / logo / text visible, download each node with download_assets (scale 2) and delete the clones. Slides does not export transparency: export logo/text with the slide black and then white and extract alpha by difference (ImageMagick). Never return base64 through use_figma (blows the context).

Product showreel (real app filmed)

Screens are filmed for real (Playwright + CDP screencast at 2x, steps goto/move/click/type/scroll/wait, pointer trail saved as <id>.cursor.json) and the cursor is redrawn in Remotion on top.

Rejected, do not repeat:

  • Gluing a designed reel to the real screen (illustrated opening + filmed middle + drawn CTA): two languages, the character vanishes and returns, the black card between acts stops the rhythm. One language per video.
  • Small, sped-up product screen in a 16:9 window: feed and ranking become an unreadable wall.
  • Zoom that crops the screen and text stamped over the video (giant verb, 01/07 counter, gradient): intrusive.
  • Empty waiting on screen ("waiting…"): cut straight to the action.
  • Vague copy: it has to be obvious and in the audience's plain spoken language.

The form that works: per action, text alone first, centered and short (2 beats), then the whole screen, uncropped, in a centered window at the capture aspect, on a clean background. Discreet text (weight 700, per-word entry with light blur), no counter, no gradient. Cuts on the track grid (14 frames = 128.57 BPM). 40s max.

Technique that paid off:

  • Speeding a capture: <OffthreadVideo trimBefore={from} playbackRate={rate}> and recompute click frames / rate (sound and cursor ring).
  • Starting another composition mid-way: <Sequence from={X}><Sequence from={-start}><Comp/></Sequence></Sequence>. Check the first frame of the stretch: a shared element "flying" in shows up loose. Skip a few frames and re-place the impact sound that fell before.
  • Language per composition with a module-level timeline: a globalThis flag set by its own entry file imported before the Root.
  • Continuous track across cuts: trimBefore on the bed so it continues from where the previous stretch stopped.
  • Loudness: render, measure and normalize with 2-pass loudnorm (-14 LUFS / -1 dBTP), copying the video. Measure only at the end: Remotion's raw mix goes over 0 dBFS.
  • Audit: timestamped contact sheet (fps=2,drawtext=%{pts\:hms},tile=8x6) of every render, and a dense sheet (fps=6) at every stretch junction.
  • A project without git, edited by another session at the same time: new cut in a new file, *.bak-* backup before editing a shared file, and tell the user.

Portfolio cards and studio reel

  • record.mjs with KEEP_UI=1 (the card shows the site as it is) and the phone preset above.
  • Scenes generated from an inventory file (only public items with a live URL); no live URL = Ken Burns of the screenshot.
  • Card package: crossfade loop from end to start, 2-pass bitrate to a KB ceiling, VP9 WebM alongside, WebP poster of frame 1.
  • Logo with alpha for the web: --codec=vp9 --pixel-format=yuva420p (Safari does not play it) + ProRes 4444 for editing; Config.setVideoImageFormat("png") keeps the alpha.
  • Cut-out/stop-motion look in Remotion: time quantized on twos (24 fps, pose every 2 frames), SVG filter per pose (feTurbulence + feDisplacementMap on the edge, feMorphology for a cream border, hard offset shadow) with 3 seeds cycling ("boil"), paper texture in multiply. Mascot parts read by id from the designer's SVG (DOMParser), so a new branding version drops in without code.
  • CC0 music: Free Music Archive (CC0 albums). Tempo and fullest stretch with librosa.beat.beat_track + windowed RMS (Python, optional).

Origin stories, reference projects and machine-specific tips: references/cases.md.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.