Parallax Landing Page

Turn a short video into a one-screen cinematic landing page where the user's scroll gesture scrubs through the video frame-by-frame while five dramatic text scenes crossfade in and out. The document never actually scrolls — wheel/touch/keyboard input is intercepted in JS and converted into virtual frame progress.

This skill encodes the patterns proven in the bundled reference at ../../examples/parasites/ (the github / lion / hope landings — three working pages built with this exact skill). Treat that folder as the ground truth if anything here is ambiguous.

Where this skill sits in the YUV.AI pyramid

parallax-landing-page is in the middle tier of the YUV.AI skills pyramid, alongside yuv-design-system, yuv-decks, yuv-viral-video, and video-to-landing-page. The top-tier orchestrator yuv-pilot routes here whenever a 5–15s video clip needs to become a single-hero scroll-scrub landing.

This skill brings its own visual language (Anton + Caveat + Inter, gold/amber/accent/cream accent palette) — it does NOT call into yuv-design-system. The two systems coexist: yuv-design-system owns the brand palette across web/app/deck surfaces, parallax-landing-page owns the cinematic scroll-scrub aesthetic specifically. For YUV.AI parallax landings, you can OPTIONALLY add a phoenix watermark + the canonical Linktree URL in the footer area as a brand throughline — see yuv-design-system/assets/ for the watermark file. Do NOT swap the Anton/Caveat type stack for the design system's Neon palette — that would lose the signature look.

What you're building

Each landing is a single locked viewport with:

A canvas filling the screen, drawing one frame at a time (cover-fit, DPR ≤ 2)
5 absolutely-positioned text scenes (kicker + uppercase Anton headline + Caveat handwritten line), each with its own scroll-progress window
A fixed nav with PARALLAX // showcase logo and links to sibling landings
A loader that preloads every frame before letting the user interact
A vignette + film grain overlay
A bottom scroll-hint that fades out once the user starts scrolling
A bottom-right frame counter (e.g. 073 / 145)
A final scene that is the CTA section — no separate .end section below

Total weight at q:v 2: roughly 120–220 KB per frame × frame count. A 6-second 24fps clip = 145 frames ≈ 25 MB. A 10-second clip = 241 frames ≈ 53 MB.

Save location

Default: ~/Documents/yuv-projects/landings/<slug>/ — always save standalone landings here so you can find them again. Override only if the user explicitly picks a different location OR if Mode B (integrate into the existing examples/parasites/ showcase) is selected.

mkdir -p ~/Documents/yuv-projects/landings
cd ~/Documents/yuv-projects/landings
# The skill creates <slug>/ here as it extracts frames and builds the HTML.

Final path (Mode A — standalone): ~/Documents/yuv-projects/landings/<slug>/. Final path (Mode B — integrate): the existing showcase directory the user chose.

Tell the user the final path at the end of every build.

Workflow

There are four phases. Don't skip phases — each one feeds the next.

Phase 1 — Discover

Confirm with the user:

Video path — where the source MP4 lives.
Slug — short folder/file name (e.g. hope, marcus, desk). One word, lowercase, no spaces. This becomes both the frame folder name (<slug>/) and the HTML filename (<slug>.html).
Theme / topic — what is this landing about? A product launch? A personal moment? A wildlife encounter? You need this to write the copy.
Project mode — standalone (brand new folder) or integrate into the existing parasites showcase at examples/parasites/ (top of this repo). If the user says "add it to the showcase", they mean Mode B below.
Accent color — for the script/CTA highlight: gold (#f5b042), amber (#ff8a3d), accent (pink #ff4d6d), cream (#f5e9d4), or something new (add it to :root in style.css and write a new .script.<name> rule).

If the user hasn't specified some of these, propose defaults and move on — don't stall on questions they can answer once they see the result.

Phase 2 — Extract frames

Run the bundled extraction script. It probes the video with ffprobe and extracts every frame at native resolution with -q:v 2 (near-lossless JPEG), naming them frame-001.jpg through frame-NNN.jpg (3-digit zero-padded).

python <skill>/scripts/extract_frames.py <video_path> <output_folder>

Example:

python C:/Users/User/.claude/skills/parallax-landing-page/scripts/extract_frames.py \
    C:/Users/User/Documents/parasites/hope/hope.mp4 \
    C:/Users/User/Documents/parasites/hope

The script prints a JSON metadata block to stdout. Capture it — you need frame_count and suggested_scroll_budget for the HTML.

Why -q:v 2: ezgif and similar tools re-compress aggressively (~55 KB/frame). Native ffmpeg at q:v 2 gives ~120–220 KB/frame at the same dimensions — visibly sharper. Do not downscale; never upscale (it inflates files without adding detail).

Why the scroll-budget formula (≈26 px per frame): keeps the tactile "pixels of scroll per frame advanced" constant across clip lengths. 145 frames → ~4200, 241 frames → ~6300, 90 frames → ~2500 (floor). The script computes it for you and clamps to [2500, 8000].

Phase 3 — Author the page

Read the template at <skill>/assets/landing-page-template.html. It is a complete working file with {{PLACEHOLDER}} markers and  comments next to creative bits. Write a new file at <target>/<slug>.html with everything substituted.

Substitution checklist (every placeholder must be replaced):

Placeholder	What to put	Example
`{{PAGE_TITLE}}`	HTML `<title>`	`Hope — The Fastest Heart in the Negev`
`{{PAGE_DESCRIPTION}}`	meta description, one sentence	`A moment with Hope, the cheetah of Midbarium.`
`{{SLUG}}`	folder name (no path, no slash)	`hope`
`{{FRAME_COUNT}}`	integer frame count	`241`
`{{FRAME_COUNTER_TEXT}}`	`001 / NNN` with matching width	`001 / 241`
`{{SCROLL_BUDGET}}`	integer from the extractor JSON	`6300`
`{{PREVIEW_FRAME_MID}}`	zero-padded middle frame	`121`
`{{PREVIEW_FRAME_END}}`	zero-padded last frame	`241`
`{{LOADER_TITLE}}`	one word, uppercase	`HOPE`
`{{LOADER_SCRIPT}}`	handwritten line ending in `…`	`summoning the runner…`
`{{NAV_LINKS_HTML}}`	full `<a>` tag block for nav	see below
`{{SCENE_N_*}}`	kicker, headline, handwritten line for each of 5 scenes	see copy guide
`{{SCROLL_HINT_TEXT}}`	2–4 word uppercase prompt	`SCROLL TO RUN`
`{{CTA_BUTTONS_HTML}}`	2–4 `<a class="cta">` tags	see below
`{{ACCENT_CLASS}}`	accent color suffix for `.script.X` and `.cta.X`	`amber`

Read references/copy-guide.md for the cinematic copy patterns (kicker phrasing, headline length, handwritten line tone). The copy is where this skill earns its keep — generic copy will sink an otherwise beautiful animation.

Phase 4 — Wire up assets

If the target folder already has style.css and parallax.js (showcase integration), do nothing — they're shared. Otherwise copy both from <skill>/assets/ into the target folder:

cp <skill>/assets/style.css <target>/style.css
cp <skill>/assets/parallax.js <target>/parallax.js

If integrating into an existing showcase, also do the steps in references/showcase-integration.md — add a card to index.html, add a link to every existing landing's navbar, and add the chain CTA on the previous landing's final scene.

Tell the user how to view:

Open <target>/<slug>.html via python -m http.server 8000 from the parent folder (not file:// — Chromium can taint the canvas with local image origins).

File contract

<target>/
├── <slug>/                  # frame folder
│   ├── <slug>.mp4           # the source (preserved)
│   ├── frame-001.jpg
│   ├── frame-002.jpg
│   └── … frame-NNN.jpg
├── <slug>.html              # the landing page
├── style.css                # shared (copy if standalone)
├── parallax.js              # shared (copy if standalone)
└── index.html               # the hub (only if standalone or integration)

The frame folder name and the HTML filename MUST share the same slug — the folder option in the page's new ParallaxPage({folder: '<slug>', ...}) call is the literal folder path.

The ParallaxPage API (already in parallax.js, do not redefine)

new ParallaxPage({
  folder:       '<slug>',     // folder containing frame-001.jpg, etc.
  frameCount:   145,          // integer
  prefix:       'frame-',     // always 'frame-' for this skill
  scrollBudget: 4200,         // pixels of accumulated input for full pass
  // optional: padWidth (default 3), ext (default 'jpg')
})

The class auto-attaches body.scrub-page (which locks scroll via CSS), preloads all frames, runs an RAF-driven lerp (0.22 factor) toward the scroll-derived target frame, and crossfades scenes whose data-start/data-end ranges include the current progress.

Hard constraints (the non-negotiables)

These are baked into the reference and break the experience if violated:

5 scenes, fixed progress windows: 0.00–0.20, 0.22–0.40, 0.42–0.60, 0.62–0.80, 0.82–1.00. The 0.02 gap between scenes is intentional — it's a moment of "just the image, no text" between beats. The hero (scene 1) and final (scene 5) use data-fade="0.30" and data-fade="0.18" respectively so they hold longer at the edges; middle scenes use the default 0.25.
No .end section beneath the scrub. The final scene contains the CTAs. The whole page is one locked viewport.
Anton + Caveat + Inter — uppercase Anton for everything bold, Caveat for handwritten accents (gold/amber/cream), Inter for any body lead text. Never substitute, never add a fourth font.
Off-white #f5f1ea, never #fff — already in CSS as var(--text).
Frame names are frame-NNN.jpg zero-padded to the width of the total (3 digits is the standard since clips < 1000 frames; the JS auto-computes counter width from the total).
The slug is the folder name AND the HTML filename basename — hope/, hope.html. No exceptions; the integration steps assume this.

Where to look when you get stuck

The reference showcase bundled in this repo: examples/parasites/ — github.html, lion.html, hope.html, style.css, parallax.js, index.html. If your output diverges visibly from those, you've drifted; reread.
references/copy-guide.md — how to write the 5 scenes so they actually feel cinematic.
references/showcase-integration.md — exact edits needed when adding to an existing hub.
assets/landing-page-template.html — the canonical structure with placeholders.
assets/hub-template.html — only used when creating a brand-new showcase (3 cards from scratch).

Two real example invocations

Standalone: user has a single clip desk.mp4 and wants a landing page.

Ask for slug + theme. Run extractor on desk.mp4 → ./desk/frame-NNN.jpg.
Copy style.css + parallax.js to project root.
Author desk.html from template. Nav has just one link: itself.
Done. Tell user to serve with python -m http.server.

Showcase add: user adds hope.mp4 to the existing parasites folder.

Slug = hope. Extract → hope/frame-NNN.jpg. (Reuse existing style.css + parallax.js.)
Author hope.html from template.
Run showcase integration: add a third card to index.html, append a Hope link to the navbar in github.html and lion.html, chain the previous landing's final scene with a Next: Hope → CTA.
If introducing a new accent color, add it to :root and add .script.X + .cta.X rules in style.css.

hoodini/ai-agents-skills

What is ai-agents-skills?

Ask in your favorite AI

Documentation