cinematic-scroll

name: cinematic-scroll description: Build cinematic scroll-driven, 3D-tilt, parallax, and environment-morphing websites — pinned chapter reveals, hero parallax, depth-image figures, hover-tilt cards, background-morphing layouts, release/launch pages, product story pages, or editorial commerce microsites. From a single self-contained scroll section (Mode A) to a full Shopify-Editions-style Next.js release site with optional AI-generated visuals via the fal.ai remote API (Mode B — requires your own fal.ai key; makes outbound API calls to fal.ai). Includes an optional audit mode in which the agent analyzes a user-supplied URL using its own browser/fetch access and scores the scroll experience on 4 dimensions (Pacing, Performance, Accessibility, Emotional Arc). Works through an optional 5-phase pipeline (cinematic audit → motion storyboard → technical spec → build → polish) with taste guardrails, 12 proven scroll patterns, 11 visual systems, and a transform/opacity performance budget as built-in craft constraints. Advanced capabilities are user-initiated, not absent: WebXR/AR sessions are feature-gated and only start on an explicit user action (an Enter VR/AR button that appears solely when the device reports support), and 3D GLB generation lives in the Mode B Next.js template (templates/nextjs/scripts/generate-flagship-assets.mjs, run with your own fal.ai key) — see activation guidelines below. version: 2.6.1 author: Simone Leonelli license: MIT metadata: hermes: tags: [Animation, Frontend, Design, 3D, Motion, GSAP, Parallax, WebGL] related_skills: [] permissions:

• filesystem:read # read project files to audit and build cinematic layouts
• filesystem:write # create and modify HTML, CSS, TypeScript, and asset files
• network:fetch # (a) call fal.ai remote API to generate images/3D GLB assets — optional, user-initiated, requires FAL_KEY; (b) fetch a user-supplied URL in audit mode — optional, user-initiated; (c) generated pages load pinned third-party CDN assets at runtime in the browser (GSAP + three.js from cdn.jsdelivr.net; three.js Draco decoder + @google/model-viewer from unpkg.com; Google Fonts) — self-host to avoid; all disclosed in manifest.json → security.thirdPartyNetworkCalls
• shell:execute # run npm/node scripts (setup, generate, typecheck, build) and the local page-proof tool (Playwright) for headless page screenshots — all optional, user-initiated
• env # read FAL_KEY / FAL_IMAGE_MODEL for fal.ai generation and CHROME_PATH / PLAYWRIGHT_BROWSERS_PATH to locate a local browser (all optional, user-initiated)

Context isolation: this skill acts only on content, files, and credentials supplied

in the CURRENT session — it never accesses, recalls, or acts on data from previous

sessions or other users, and FAL_KEY/credentials must be given explicitly here,

never inferred from memory or prior context. Two deliberate, user-initiated network

exceptions are disclosed below (and in manifest.json → security.thirdPartyNetworkCalls):

(1) audit mode fetches the user-supplied URL via the agent's OWN browser/fetch — only

sites the user owns or is authorized to test; (2) generated pages load pinned

third-party CDN assets (GSAP, Google Fonts, three.js/Draco, @google/model-viewer) when

opened in a browser — self-host these if your deployment policy requires it.

Activation: invoke this skill when the user asks to BUILD or AUDIT a

cinematic/scroll/parallax/3D-tilt website, launch page, or editorial microsite,

OR asks to review or score an existing URL's scroll experience,

OR asks to LEARN FROM / study / distill patterns from an existing URL.

Do NOT activate for: generic landing pages, CRUD apps, forms-based workflows,

CMS integrations, dashboards, or any request that does not explicitly involve

scroll-driven motion or cinematic web design.

Motion-heavy defaults apply only within this skill's output. If the user requests

minimal animation or a static fallback, respect that preference without requiring

justification.

Cinematic Scroll

Reusable patterns + production templates for building cinematic, scroll-driven React pages: pinned chapters, multi-depth parallax, 3D mouse tilt, environment-morphing backgrounds, reduced-motion-safe degradation, and (optionally) a full Next.js release site with fal.ai-generated visuals.

This is v2.0 — built on a 5-phase pipeline that is adaptively gated (see "Match the gating to the ask" below). Every phase produces a reviewable artifact: the user approves each phase before the next when they want the process or the brief is ambiguous, and the agent runs straight through when handed a complete brief or asked for a result directly (still emitting every artifact). This replaces the v1.0 one-shot model with a process that consistently produces production-quality output.

Agent quickstart — route, act, verify

Read this section first; read the rest as the route demands. Three rules:

1 · Route the request. Match what the user asked for and go straight to work:

1b · Speak the design contract. Before emitting any CSS/TS, resolve every color, type size, spacing value, easing curve, and pin height through a token — never a literal. The readable map is design.md; the machine source is tokens/ (DTCG: core primitives, motion, semantic roles). Components/chapters reference semantic role vars only (--bg, --accent, --ease-reveal); a visual-system swap is one themes/*.theme.json. Verify the contract with npm run tokens:check. Full reference: references/design-tokens.md. Reusable, doctor-verified building blocks (HeroParallax, PinnedReveal, DepthFigure, TiltCard, MorphBackground, HorizontalGallery, ScrubVideo, KineticHeadline, MagneticCursor) live in references/component-grammar.md + components/ (Mode A html + Mode B tsx) — start from these instead of writing motion from scratch.

1c · Reuse what the skill has learned. Before building, scan the ## Learned additions pointer sections of references/scroll-patterns.md, references/visual-systems.md, references/film-archetypes.md, and taste-guardrails.md, and fetch any relevant references/learned/<type>/<slug>.md entry on demand (pointer-first, loaded only when relevant). These are distilled, original recipes the skill learned from authorized sites. To add to them, see learn-mode.md.

2 · Match the gating to the ask. The 5-phase pipeline below produces an artifact per phase. When the user wants the process (or the brief is genuinely ambiguous), gate each phase on their approval as written. When the user asked for a result ("build me…", one-shot, CI, or another agent invoked you), run the phases internally without pausing, still emit the artifacts (cinematic-audit.md, motion-storyboard.md, technical-spec.md, polish-report.md) as the audit trail, and replace human gates with the verify loop below. Never block an autonomous run waiting for approval the user can't give.

3 · Verify before you call it done — every time.

node tools/cinematic-doctor/cli.mjs <your-page>.html   # 0–100; exits non-zero < 80

Fix what it flags and re-run until it passes — the same gate CI enforces. The doctor scores taste, performance, a11y, mobile, tokens, and 3D; its findings reference the exact guardrail sections to read. For Mode B also run npm run typecheck and npm run build in the project.

The doctor grades the static contract; pair it with the runtime half:

npm run proof -- <url-or-file>          # tools/page-proof — headless run +
                                        # console errors + scroll screenshots

page-proof opens the page in headless Chromium, scrolls it, collects every console error / uncaught exception / failed request, and writes screenshots at each depth (.page-proof/proof.json + shots). Exit 1 means runtime errors — fix and re-run. Add --fps on DOM pages to measure scroll smoothness (avg fps

• dropped-frame share — the §1 transform/opacity budget, measured, not asserted). Needs playwright-core + any Chrome (--wait 8000 for WebGL).

Then LOOK at the shots — this step is not optional. You can read images: open every screenshot page-proof wrote and grade the frame like a director reviewing dailies, against taste-guardrails.md:

• Composition — is there a clear focal point at every depth, or dead/empty frames mid-scroll? (A dwell with nothing composed in view is a failed shot.)
• Hierarchy — does the type read in order (eyebrow → display → body)? Any title colliding with imagery or another overlay?
• Reveal state — are entrance animations finished or stuck half-way (clipped masks, 0-opacity text that never arrived)?
• Canvas truth — for WebGL: is the scene actually rendering, or is it a black/empty canvas behind healthy-looking DOM?
• Edges — stretched or wrongly-cropped media, horizontal overflow, elements pinned off-screen.

Anything you would screenshot-and-complain-about as a user, fix and re-prove. A build is done when the doctor passes, proof exits 0, AND the shots would survive an art director's review.

Verification map — four surfaces, each answers a different question:

The contract itself: npm run tokens:check · themes:check · links:check · evals:run, or all gates via npm test.

The aesthetic is the user's — the motion is yours

This skill supplies the motion grammar, never a fixed look. The pinned chapters, parallax, tilt, title choreography, and morphing backgrounds are the constant; the visual world — palette, typography, imagery, mood — comes entirely from the user's brief. Derive the aesthetic from what they ask for (brand, references, palette, vibe, or a visual system from references/film-archetypes.md). If they haven't said, ask or offer 2–3 distinct directions — never default to any one style. The same machinery must produce a brutalist black-on-white drop, a quiet-luxury launch, a neon Gen-Z page, a sci-fi noir reveal, an organic wellness story, or a Renaissance editorial. None is "the" style. The five public examples (examples/renaissance, examples/studio, examples/noir, examples/luxe, examples/pop) are different worlds from the same engine — proof the look is a variable, not a default.

Philosophy

1. Taste is non-negotiable

The difference between slop and craft is anti-convergence. This skill ships with taste-guardrails.md — 11 banned patterns, a cinematic vocabulary, pacing rules, and anti-convergence principles. These are the skill's default craft constraints (anti-slop quality, not a forced aesthetic or locale): the user's explicit preferences always win — palette, tone, intensity, language, or a minimal/static fallback are theirs to set. Absent such direction, an agent that skips these guardrails produces tasteless output regardless of prompt quality, so every generated file is checked against the banned-patterns list before delivery.

2. Process over prompt

A great prompt is not enough. The 5-phase gated pipeline ensures that auditing, planning, specifying, building, and polishing happen as discrete, reviewable steps. The user sees a cinematic-audit.md before any code is written. They approve a motion-storyboard.md before any animation is implemented. Process de-risks the output.

3. Film grammar over web patterns

Scroll is not "web design." It is digital cinematography. The cinematic vocabulary in taste-guardrails.md (Section 2) maps 12 film techniques to scroll equivalents — dolly zooms, whip pans, rack focus, tracking shots, crane shots. Every scroll behavior names the film technique it implements. This is how we produce cinema, not PowerPoint transitions.

4. Measurable quality

Every output has reviewable artifacts. Every phase has a decision gate. Every build is checked against performance-budget.md (Section 6, 11-point pre-launch checklist). Quality is not a feeling — it is a checklist.

The 5-Phase Pipeline

Each phase produces a reviewable .md artifact. Gating is adaptive (see the quickstart's "Match the gating to the ask"): when the user wants the process or the brief is ambiguous, the user reviews and approves each phase before the next; when the user asked for a direct result (a complete brief, one-shot, CI, or another agent invoked this skill), the agent runs the phases internally without pausing and still emits every artifact as the audit trail. The agent never silently drops a phase's artifact.

Phase 1: Cinematic Audit

Purpose: Analyze the brand/content, define the emotional arc, select the visual system, and establish the motion personality.

Agent instructions

1. Ask the user about their brand's motion personality if not provided:

   • "What emotion should the first 3 seconds produce?"
   • "Is your brand closer to a Symmetric Monument (meticulous, formal) or a Warm Scrapbook (intimate, playful)?"
   • "Who is scrolling this — a curious visitor or a decision-maker?"

2. Select a visual system from references/film-archetypes.md. Read the archetypes file (Section 1-7) and match the brief to ONE primary visual system. Document the choice in the audit with rationale. Never mix more than 2 visual systems; if hybridity is needed, choose one primary and one accent.

3. Define the emotional arc across the full scroll journey:

   • Opening emotion (what the user feels at scroll position 0)
   • Mid-journey turning point (where the narrative shifts)
   • Closing emotion (what the user carries away)
   • Pacing rhythm: glacial / medium / energetic / variable

4. Document:

   • Brand motion personality (3-5 adjectives)
   • Emotional arc definition (opening → midpoint → closing)
   • Audience analysis (device split, technical sophistication, attention span)
   • Device context (primary viewport, performance tier expectation)
   • Accessibility requirements (reduced-motion needs, WCAG target)
   • Visual system selection (primary + optional accent, with rationale)
   • Color temperature progression across chapters (warm → cool → neutral)
   • Typography strategy (display font + body font, from archetype)

Output: cinematic-audit.md

→ Full template: references/artifact-templates.md. Copy the cinematic-audit.md section and fill every field.

Phase 1.5: Asset Direction — the world before the layout

Purpose: Decide the physical world the brand lives in and where every visual comes from — so the build can't drift into a generic dark-landing-page. This is the module that makes wow reproducible instead of occasional. Required for any release · launch · immersive · premium · flagship · "wow" brief; skippable only for a single utilitarian section.

Read references/asset-direction.md (the five decisions + the coherence rule) and references/wow-gate.md (the gate). The storyboard (Phase 2) then makes every chapter carry ≥1 motif, and the polish phase (Phase 5) verifies the signature moment survived in the page-proof frames — paired with cinematic-doctor, which scores the executed contract. Doctor = "not slop"; Wow Gate = "actually memorable."

→ Artifact template: references/artifact-templates.md, the art-direction.md section.

Phase 2: Motion Storyboard

Purpose: Plan the scroll sequence — chapters, patterns, transitions, depth layers, timing, and mobile degradation.

Agent instructions

1. Design a chapter breakdown of 5-8 chapters. Each chapter is one pinned section with a distinct visual world. The total scroll distance should be 1500-3000vh for the full experience.

2. Select ONE pattern from references/scroll-patterns.md per chapter. The 12 available patterns (Section 1-12) are:

   • Pinned Hero, Scrubbed Timeline, Velocity-Reactive, Sticky Narrative, Chaptered Release, Parallax Gallery, 3D Product Orbit, Editorial Longread, Data Story, Landing Sequence, Portfolio Reveal, Archive Explorer. Document the pattern choice and rationale for each chapter.

3. Ensure no adjacent chapters use the same pattern or transition type. This is a hard rule from taste-guardrails.md Section 4.4. Alternate: fade → slide → scale → rotate → crossfade → wipe.

4. Configure depth layers per chapter following the selected pattern's depth configuration. Reference taste-guardrails.md Section 4.3: never repeat a depth multiplier between adjacent chapters. Maximum 7 layers per chapter (taste-guardrails.md Section 1.7).

5. Verify all pinned sections respect the 150-400vh rule from taste-guardrails.md Section 3.2 and 3.3. No pin shorter than 150vh, no pin longer than 400vh.

6. Ensure breathing room between chapters: minimum 80vh of free-scroll space between pinned chapters (taste-guardrails.md Section 3.4).

7. Specify the title reveal style per chapter, rotating through the vocabulary in taste-guardrails.md Section 4.5. Never use the same treatment twice in a row.

8. Document the mobile degradation plan per chapter using the tier system from performance-budget.md Section 3.

Output: motion-storyboard.md

→ Full template: references/artifact-templates.md. Copy the motion-storyboard.md section and fill every field.

Phase 3: Technical Spec

Purpose: Output the Lenis/GSAP/ScrollTrigger implementation plan with exact configs, performance budget allocation, and asset requirements.

Agent instructions

1. Select packages:

   • Smooth scroll: Lenis (lenis npm package — NOT @studio-freight/lenis) OR GSAP ScrollSmoother (preferred when GSAP is already in the build)
   • Animation: GSAP + ScrollTrigger + SplitText (all now free)
   • Motion primitives: choreo-3d for pinning orchestration
   • Framework: React 19 + Next.js App Router (Mode B) or vanilla (Mode A)

2. Specify exact GSAP ScrollTrigger configs for every pinned chapter:

   • scrub value (0.3-0.8 range per performance-budget.md Section 7)
   • start and end positions
   • pin configuration
   • snap behavior if applicable
   • Easing functions per role (hero entrance, exit, micro-interaction, chapter transition — from taste-guardrails.md Section 4.1)

3. Allocate performance budget from performance-budget.md:

   • Layer count per viewport (max 10 desktop, 4 mobile — Section 2)
   • will-change strategy (Section 2)
   • Image budget per chapter (Section 5)
   • Font budget (Section 5)
   • JS budget (Section 5)

4. Flag any performance risks: If the storyboard requests more than 7 layers per chapter, more than 3 simultaneous motion types in a 50vh window, or pins approaching the 400vh limit, flag it here with mitigation.

5. Document asset requirements: Images, videos, fonts, with specifications for each (format, dimensions, generation prompts if using fal.ai).

6. Specify mobile degradation implementation per chapter, referencing performance-budget.md Section 3 (Mobile Degradation Matrix).

7. Select the 3D / shader stack tier (see block below) and declare the chosen tier in the technical spec. Default is Tier A (no 3D). Every step up must be justified by a named narrative need, not a vibe.

3D / Shader Stack Selection

3D is the most expensive thing you can put on a scroll page. The default answer is Tier A: no 3D. Earn each step up the ladder with a reason. Pick the lowest tier that satisfies the narrative — climbing a tier multiplies cost (bundle, GPU memory, battery, asset-production time, failure surface). Read references/3d-stack.md and references/webxr.md before specifying any 3D.

Does the story actually need real 3D depth / rotation / parallax-in-space?
│
├─ NO  ──────────────────────────────────────────►  TIER A  (GSAP / CSS only)
│       Fake depth with layered transforms + CSS 3D. 95% of cinematic
│       scroll pages live here. (scroll-patterns.md #1, #6, #7.)
│
└─ YES → Do you have (or can you commission) a real model?
         │
         ├─ YES, a discrete object/scene/figure  ──►  TIER B  (Three + GLB)
         │        A product, an environment, an avatar. Asset-driven.
         │
         └─ NO model — the visual IS the math  ─────►  TIER C  (Three + shaders)
                  Fields, flows, particles, generative surfaces. Zero assets.

         …and on top of B or C, only if a flat screen genuinely cannot deliver
         the moment (scale / presence / embodiment):
                                          ──────────►  TIER D  (+ WebXR)
                  Immersive VR / room-scale AR. (references/webxr.md.)

• Tier A — GSAP / CSS only. The depth is illusory and a screen is the final medium. The "3D Product Orbit" pattern is CSS rotateY on layered images, not a mesh. If you cannot name a specific thing the user does that requires a camera moving through real geometry, stay here.
• Tier B — GSAP + Three.js + GLB. A discrete hero artifact whose form carries the story and that the user orbits / inspects / configures, or a place the camera flies through. Requires the GLB (or a committed path to it) plus the ASSETS-3D.md hand-off and a manifest.
• Tier C — GSAP + Three.js + procedural shaders. The visual is the computation — a field, flow, surface, or particle system with no object to model. Zero external assets; this is the procedural backbone (and the fallback every Tier-B chapter degrades to).
• Tier D — any of B/C + WebXR. Flatness is the actual limitation (scale, presence, embodiment). XR is a session the user explicitly enters; never the default render path. The 2D page must be complete on its own.

Pin every renderer version exactly — no latest, no floating majors. Three.js makes breaking changes between minors; an un-pinned 3D stack is a future blank chapter. Pin vanilla Three to an exact patch (three@0.160.0) via a versioned CDN import map; pin Three exactly in package builds (caret only on the R3F wrappers, lockfile freezes the tree). Full pin table and import map in references/3d-stack.md Section 2.

Declare the stack tier in the technical spec (the table in the output template). State the chosen tier, the named narrative need that justifies it, and the pinned versions. If two tiers both satisfy the story, ship the lower one — "could be 3D" is not "should be 3D." Full decision criteria, performance caps, fallback rules, and the scroll-camera pattern: references/3d-stack.md. WebXR session setup, comfort/safety, and AR quick-look: references/webxr.md.

Output: technical-spec.md

→ Full template: references/artifact-templates.md. Copy the technical-spec.md section and fill every field.

Phase 4: Build

Purpose: Generate the code.

Agent instructions

1. Apply ALL taste guardrails as hard constraints. Before delivering, check every output against the banned patterns list in taste-guardrails.md Section 1. Violating these rules is a bug, not a style choice.

2. Ensure reduced-motion fallback for every scroll-driven effect. When prefers-reduced-motion: reduce is active: disable pinning, disable parallax, show static compositions, set all transitions to instant. Reference performance-budget.md Section 3, Tier 4.

3. Give mobile a touch-safe cinematic experience — not a dead page. Below 768px, unpin every chapter and stack the layout, BUT keep motion scroll-coupled: a lerped/damped image parallax (one transform-only mover per section) plus scroll-linked entrance reveals (transform + opacity). A flat, motionless mobile page is a failure mode for this skill. Drive it with JS (rAF reading scrollY in Mode A, framer useScroll/useSpring in Mode B) — never CSS animation-timeline, which iOS Safari reports as supported but does not actually run. No pinning/scroll-jacking on touch, no 3D tilt on touch. Reference references/mobile-motion.md for the recipe and performance-budget.md Section 3 (Mobile Degradation Matrix).

4. Name the cinematic technique in code comments. Every scroll-driven animation should carry a developer comment identifying the film technique it implements (from taste-guardrails.md Section 2). This is a code-comment convention for maintainers — it does not dictate the page's user-facing language or locale, which follow the user's request.

5. Only animate transform and opacity in hot scroll paths. Never width, height, top, left, filter, box-shadow. Reference performance-budget.md Section 1 (Permitted Properties).

6. Use will-change strategically — 200ms before animation starts, 200ms after it ends, max 3 simultaneous elements. Never globally. Reference performance-budget.md Section 2.

7. Optional accelerator — compile from a choreography document. If the technical spec is expressed as a scroll-choreography.json, run the bundled compiler to emit the GSAP ScrollTrigger + Lenis code instead of hand-writing it: node compile-choreography.mjs my-scene.json --out scene.js. The compiler maps the schema's CSS property names to GSAP shorthand (translateX→x, rotateZ→rotation, …) — a mapping that is easy to get wrong by hand and silently no-ops in GSAP if you do. See scroll-choreography-compilation.md.

8. One choreography, two media. The same document also compiles to a fixed-time video timeline (Basic): node compile-choreography.mjs my-scene.json --target video. Scroll progress maps to seconds; the DOM contract is identical, so one skeleton serves the page and its launch film. The full render-ready web→video pipeline (HyperFrames/Remotion composition, npx hyperframes render → MP4, avatar walkthroughs) ships in Cinematic Scroll Studio.

9. Follow the technical-spec.md exactly. Do not improvise animation configs that differ from the approved spec.

10. If using fal.ai assets, follow the server-side generation pattern, never expose FAL_KEY in client code. Reference MODELS.md for model selection and cost.

Mode A vs Mode B

This phase operates in two modes. Follow the mode specified in the technical-spec.md.

If the request is ambiguous, default to Mode A for a single section and Mode B when the user says "site", "page", "release", "launch", or "landing".

Mode A build rules

• Single self-contained HTML file: <!DOCTYPE html> ... </html>, inline CSS + JS, renders immediately with no build step and no npm packages.
• Keep it truly single-file and file://-safe (guardrail 1.13): inline all JS — no sibling <script src="./x.js"> — and avoid JS-loaded local assets (runtime-fetched textures/images), since the browser's file:// cross-origin policy blocks both and the page would fail off a double-click. If a build genuinely needs external modules or texture files it is a served page (tell the user to "run python3 -m http.server"), not a file:// one. Ship no dead code (guardrail 1.12): every CSS selector must match an element, every attribute-branch must fire. cinematic-doctor's advisory hygiene pass flags external local scripts, dead selectors, and dead attribute-branches — drive it to zero.
• Load GSAP + ScrollTrigger from a pinned CDN with SRI (exact version, integrity hash, crossorigin — as every examples/* page does). Third-party CDN disclosure: generated Mode A pages load GSAP from unpkg.com, fonts from fonts.googleapis.com / fonts.gstatic.com, and @google/model-viewer from cdn.jsdelivr.net. These are outbound network requests made by the browser when the page is opened. If the user's deployment policy restricts third-party origins, self-host these assets and replace the CDN URLs — mention this proactively. For sandboxes where the CDN is unreachable, fall back to requestAnimationFrame-throttled scroll handlers: identical motion grammar, zero dependencies. Either way there is no bundler and nothing to npm install. (Lenis is Mode B only.)
• perspective: 1200px on chapter wrapper. 3D transforms on at least one layer (rotateX ±4deg max, rotateY ±2deg max).
• Minimum 5 depth layers per chapter.
• Type reveal: use one of mask reveal, word stagger, letter stagger, vertical mask, or scrub letter-spacing.
• clamp() for all typography. No fixed px for font-size.
• Progress HUD in top-right for sandbox/iframe environments.
• Reduced-motion check: prefers-reduced-motion: reduce → static composition, no scroll binding.

Mode B build rules

• Scaffold from templates/nextjs/ — copy bundled files verbatim. Do NOT regenerate package.json, ChapterScene.tsx, fal-models.ts, fal-generate.ts, or API routes from memory. The templates contain tested, production-safe code.
• choreo-3d for motion primitives: ScrollChoreography, ScrollLayer, ScrollDepthImage, ScrollBackgroundMorph, useTilt3D, useMouseSpring.
• GSAP plugins (all free): ScrollTrigger, SplitText, ScrollSmoother. Register once: gsap.registerPlugin(ScrollTrigger, SplitText, ScrollSmoother).
• @gsap/react's useGSAP() hook with a scope for cleanup.
• Lenis (lenis package — NOT @studio-freight/lenis) for smooth scroll. Forward Lenis RAF tick to ScrollTrigger.update.
• lib/editions-manifest.ts — 6-12 chapters, each with: id, eyebrow, title, summary, features, accent, background, foreground, poster, video.
• ChapterScene.tsx — the 7-layer cinematic scene. Do NOT downgrade it: never collapse to 2 layers, never remove perspective: 1200px, never replace word-stagger with plain opacity fade, never drop mobile fallback.
• lib/fal-models.ts adapter for all image generation — never inline image_size, aspect_ratio, or negative_prompt.
• fal.ai key stays server-side only. Never in client components or .env.

3D / WebGL / XR build rules (Tier B/C/D only)

These apply only when the technical spec declared Tier B, C, or D. Tier A (GSAP/CSS) ships none of this. examples/flagship/ is the worked 4-chapter reference; references/3d-stack.md (renderer + caps) and references/webxr.md (Tier D sessions) are the authority. Each rule below is enforceable — a miss is a bug, not a style choice.

1. Cap devicePixelRatio low for live scenes — Math.min(devicePixelRatio, isMobile ? 1.0 : 1.5) (raymarchers/fullscreen shaders can go 0.85–1.3; the blur hides it). Uncapped or a flat 2 is a Retina GPU tax — a 4K display renders 4–9× the pixels and is the #1 cause of 3D scroll jank. One renderer per page — never a second WebGL context per chapter. See performance-budget.md §9. (doctor-enforced)

1b. Light budget: ~2–4 dynamic lights — prefer emissive + IBL. Every real-time PointLight/SpotLight/DirectionalLight costs per-fragment on every lit mesh; one light per object (e.g. a spotlight per painting) is a frame-rate cliff. Get the look from scene.environment (an equirectangular HDRI via PMREMGenerator — image-based fill + real reflections), emissive materials/emissiveMap for self-lit art, and a couple of camera-following lights. Keep MSAA off on fog/fill-heavy scenes. performance-budget.md §9. (doctor-enforced)

2. Feature-detect WebGL before creating a context. Probe for a context; if it fails, render the permanent poster / CSS fallback — never a blank canvas. The fallback is a first-class deliverable, not an afterthought.

3. Handle context loss. Add a webglcontextlost listener that calls e.preventDefault(), plus a webglcontextrestored handler that rebuilds the scene. Without preventDefault() the context never comes back.

4. prefers-reduced-motion: reduce → render a single static frame. Draw once, then stop. No continuous rAF loop, no auto-rotate, no idle animation.

5. Gate the rAF loop. Run frames only when the document is visible (document.visibilityState) AND the canvas is on-screen (IntersectionObserver). Stop the loop otherwise. On teardown, dispose every geometry, material, and texture (and the renderer) — leaked GPU resources accumulate per chapter.

6. All runtime 3D asset paths come from a manifest — never hardcode model, USDZ, or poster paths in code. Read them from examples/flagship/assets-3d/manifest.json (shape: version (2), basePath, chapters.{id}.{label, runtime, iosAr, fallbackPoster, height, lift, spin, pivot, cameraNodes, shader, scale, animations, stripRootMotion} — runtime is the glb path or "procedural", iosAr the optional .usdz).

7. XR (Tier D) is feature-gated. Check navigator.xr and await navigator.xr.isSessionSupported('immersive-vr' | 'immersive-ar') before rendering any Enter-VR / Enter-AR button. If unsupported, the button never appears. The 2D page must be complete and shippable without XR — XR is a session the user explicitly enters, never the default render path.

Output: Mode A (single file) or Mode B (project directory)

Phase 5: Polish

Purpose: Performance audit, accessibility check, mobile verification, and final quality gate.

Agent instructions

1. Run the performance-budget.md monitoring checklist (Section 6). All 11 pre-launch checks must be verified:

   • Chrome DevTools Performance: 10s scroll recording, < 5% red frames
   • Lighthouse Performance score > 90
   • WebPageTest filmstrip: smooth visual progression during scroll
   • iPhone 12 Safari: no visible stutter during fast scroll
   • iPhone SE: content accessible, no broken layout on budget tier
   • Reduced-motion test: all content visible, no broken layout
   • Battery test: 5min continuous scrolling drains < 3% battery
   • Memory test: tab memory does not grow > 50MB after 5min scrolling
   • Layer count: < 10 layers desktop, < 4 on mobile
   • No layout thrashing: no purple "Layout" bars during scroll
   • Network: no images load during scroll animation

2. Verify no banned patterns survived. Re-check the code against taste-guardrails.md Section 1 (Banned Patterns).

3. Confirm emotional arc matches Phase 1 audit. Scroll through the entire experience and verify the emotional progression matches the cinematic-audit.md definition.

4. Verify all reduced-motion fallbacks. Test with macOS → Accessibility → Reduce Motion ON. All content must be visible and usable.

5. Verify mobile degradation. Test at 375px viewport. All pinned sections must be converted to stacked layout. No broken tap targets.

6. Verify accessibility: All images have meaningful alt text (or alt="" if decorative). All interactive elements have focus states. aria-label on visual navigation controls. Keyboard navigation works.

7. Measure scroll jank using the protocol from performance-budget.md Section 4 (Scroll Jank Measurement Protocol).

8. 3D / XR polish checks (Tier B/C/D only). Skip for Tier A. Verify:

   • Context loss tested — force a webglcontextlost, confirm e.preventDefault() fires and webglcontextrestored rebuilds the scene.
   • Fallback verified — disable WebGL (or block the context) and confirm the permanent poster / CSS fallback renders, never a blank canvas.
   • Mobile devicePixelRatio lowered (≤ 2, lower on phones) and a single renderer is used for the whole page.
   • No per-frame allocation — no new geometries/materials/vectors inside the rAF loop; no raycasting or heavy work every frame.
   • Teardown disposes all geometries/materials/textures (no GPU leak across chapters); rAF is gated on visibility + on-screen.
   • prefers-reduced-motion renders a single static frame (no loop).
   • XR feature-gated — Enter-VR/AR only appears after navigator.xr.isSessionSupported(...); the 2D page is complete without XR. See references/3d-stack.md and references/webxr.md for the authority.

9. Run the cinematic-doctor quality gate — the polish phase is not complete until it passes. Every build SHOULD pass this executable gate before shipping:

   npm run doctor -- examples/your-build/index.html
   # equivalently (the gate's direct entry point):
   node tools/cinematic-doctor/cli.mjs examples/your-build/index.html
   

   It statically scores the build 0–100 across taste, performance, a11y, mobile, tokens, and (when 3D is detected) 3D categories, prints a scorecard, writes cinematic-report.json, and exits non-zero below the default threshold of 80 — so it is CI-blockable / pre-commit-hook ready. Treat a failing score as a list of concrete fixes to apply, then re-run until it passes. Do not call the build polished while cinematic-doctor is red.

Output: polish-report.md

→ Full template: references/artifact-templates.md. Copy the polish-report.md section and fill every field.

Mandatory Motion + Craft Requirements

Every artifact MUST satisfy ALL of these. No exceptions for "demo simplicity" — the demo IS the product.

1. Multi-depth field — minimum 5 layers

Two-layer parallax is amateur. A real depth field uses 5-7 layers at distinct depth multipliers. Pick at least 5 of these 7 slots:

2. 3D perspective camera

Set perspective: 1200px on the chapter wrapper. Use scroll-driven 3D transforms on at least one layer: rotateX(±4deg max), rotateY(±2deg max), translateZ(0px → -80px). Disable all 3D rotation on touch devices AND when prefers-reduced-motion: reduce.

3. Type reveal patterns

Plain opacity: 0 → 1 on oversized titles is lazy. Use one of: word stagger, letter stagger, mask reveal (clip-path: inset), vertical mask, scrub letter-spacing. Combine with translateY() and opacity.

4. Smooth scrolling — mandatory in production

• Mode A: requestAnimationFrame-throttled scroll handlers (not raw scroll events). No packages — dependency-free by design.
• Mode B: Lenis (lenis npm — NOT @studio-freight/lenis) OR GSAP ScrollSmoother (preferred when GSAP is already in the build). Forward Lenis RAF tick to ScrollTrigger.update if using both.

5. GSAP is now free — use the premium plugins in Mode B

As of the Webflow acquisition (2025), GSAP is 100% free including every former Club plugin. In Mode B, prefer:

Register once: gsap.registerPlugin(ScrollTrigger, SplitText, ScrollSmoother).

6. Mobile-responsive — mandatory

• <meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover">
• Typography in clamp(min, fluid, max) — never fixed px for font-size
• Disable pinning below 768px, but keep motion scroll-coupled and touch-safe: lerped image parallax (one transform-only mover per section) + scroll-linked entrance reveals (transform + opacity). JS-driven, NOT CSS animation-timeline (iOS Safari reports support but doesn't run it). See references/mobile-motion.md.
• env(safe-area-inset-*) padding on fixed nav / overlays
• Tap targets ≥ 44px square
• Mobile-first: design at 375px viewport FIRST, then scale up

7. Loading sequence

• Preload critical backgrounds with <link rel="preload" as="image">
• Show poster / blurred LQIP placeholder during decode
• First paint readable within ~1.5s on simulated 4G
• In Next.js, <Image> with priority on above-the-fold imagery

8. Performance — compositor-only paths, designed for 60fps (benchmark your targets)

• Only transform and opacity mutate per scroll frame
• will-change: transform on animated layers ONLY (never globally)
• translate3d(0,0,0) to force GPU compositing where needed
• Cache getBoundingClientRect() once on init + resize, never per frame
• No layout reads in scroll handlers
• Chrome DevTools Performance flame chart = all green (composite only)
• Lighthouse Performance ≥ 90

9. Component rules

• Every full-screen chapter: id + single <section> wrapper + eyebrow, title, summary, features, asset, accent
• All text overlays = selectable HTML, never baked into images
• aria-label on visual navigation controls
• Avoid scroll hijacking — pin per chapter, not the whole page
• On mobile: collapse pinned scenes into stacked vertical cards
• Prefer 16:9 backgrounds, 4:5 foreground figures

Core Principles

1. Reduced motion first. Every effect degrades gracefully when prefers-reduced-motion: reduce is set. Pin hooks skip GSAP, layers snap to stable mid-keyframe, tilt returns zeros.

2. iOS WebKit video safety. Safari freezes <video> frames inside a transform-style: preserve-3d ancestor that updates. Detect touch and bypass the 3D wrapper for video.

3. Animate transform + opacity only in hot scroll paths.

4. Pin chapters, not the page. Each cinematic block opts into pinning. The rest of the document scrolls normally.

5. Deterministic motion. Any procedural value must be stable across re-renders so SSR and resize don't shift layout.

Quality Bar

Output must compete with:

• Shopify Editions (Winter/Summer drops) — multi-chapter release worlds
• Apple product launch pages — pinned cinematic sequences
• Linear release notes — editorial typography + restraint
• Stripe Sessions — depth-of-field + atmospheric morphing
• Awwwards SOTD nominees in Editorial + Product Launch categories

"Looks like a Bootstrap landing page" or "looks like a Tailwind UI template" = failure. Output should look studio-crafted. If constraints prevent this tier, say so explicitly and deliver the highest-quality fallback the constraints allow — never ship mid-tier silently.

fal.ai Integration (Mode B)

This skill includes NO keys or credits. Every user creates their own fal.ai account. The page works without fal.ai — ChapterDemoVisual renders stunning CSS-only chapter visuals at $0.

Setup

1. Walk new users through examples/GETTING_STARTED.md
2. Sign up at fal.ai, create API key, add FAL_KEY to .env.local
3. Restart dev server after adding env vars
4. Never put FAL_KEY in client components or committed .env files
5. Mention they can skip fal.ai and use static images

Technical rules

1. Never expose FAL_KEY in browser code
2. Use @fal-ai/server-proxy/nextjs — export GET, POST, and PUT
3. Always go through lib/fal-models.ts — never inline image_size or negative_prompt
4. Use server routes for production asset generation
5. Use fal.subscribe for ≤5 chapters; fal.queue.submit + webhook for >5
6. Set allowedEndpoints on the proxy + allowUnauthorizedRequests: false
7. Model IDs configurable via environment variables

See MODELS.md for the full model menu, cost table, and per-model parameter differences. Default: fal-ai/flux-2-pro (~$0.06/img, ~4s).

Quick-Start (For Expert Users)

Experienced users can skip the full pipeline by providing a complete brief upfront. The agent runs all 5 phases internally and delivers the final output in one shot. Use these prompts as templates.

Quick-Start A: Single scroll section (Mode A)

Build a cinematic-scroll pinned hero chapter for my [brand/product]. Visual system: [Symmetric Monument / Clinical Noir / Storybook Geometry / Temporal Monument / Atmospheric Sublime / Warm Scrapbook / Naturalistic Drift]. [N] chapters, [color palette], [typography feel]. Pin duration [X]vh. Output: single self-contained HTML file.

The agent internally runs Phase 1-3 assumptions, builds (Phase 4), and delivers a performance-annotated file with inline polish notes (Phase 5 lightweight).

Quick-Start B: Full release site (Mode B)

Scaffold a complete Shopify-Editions-tier cinematic release page for [product]. Visual system: [name]. [N] chapters. Demo mode first — no fal.ai key required. Copy templates verbatim from templates/nextjs/.

The agent runs the full pipeline internally: cinematic audit (assumed), storyboard (assumed), technical spec (assumed), build (Mode B), and delivers with a lightweight polish checklist.

Quick-Start C: Existing project upgrade

Add a cinematic-scroll pinned chapter to my existing [React/Next.js] project. Use choreo-3d primitives. Pattern: [Pinned Hero/Chaptered Release/etc from scroll-patterns.md]. Pin [X]vh, [N] layers. Match my existing [palette/typography].

The agent runs Phase 2-4 only, integrating with the existing codebase.

Example Prompts — Full Pipeline (5-Phase)

These examples show how the complete gated pipeline works end-to-end.

Example 1: Fintech Trust Page (Clinical Noir system)

Phase 1: We're a fintech app that needs to communicate trust and precision. Our brand is clinical, data-driven, restrained. Audience: decision-makers on desktop. Build a cinematic-scroll experience using the Clinical Noir visual system from references/film-archetypes.md Section 2. Output: cinematic-audit.md.

Phase 2: Based on the audit, design a 6-chapter motion storyboard. Chapters: Authority, Problem, Solution, Product, Proof, CTA. Use Chaptered Release pattern for chapters 1 and 5, Sticky Narrative for chapter 2, Data Story for chapter 4. Reference references/scroll-patterns.md Sections 5, 4, and 9. Output: motion-storyboard.md.

Phase 3: Produce the technical spec. Use GSAP ScrollTrigger + Lenis + choreo-3d. Scrub 0.5, pin spacing true. Reference performance-budget.md Sections 1, 2, and 7 for all constraints. Output: technical-spec.md.

Phase 4: Build Mode B — Next.js project from templates. 6 chapters, Clinical Noir palette (ash grey, steel blue, sickly yellow-green, black). CSS-only demo mode for first run. Output: project directory.

Phase 5: Run the full polish checklist. Verify all 11 pre-launch checks from performance-budget.md Section 6. Confirm no banned patterns from taste-guardrails.md Section 1 survived. Output: polish-report.md.

Example 2: Wellness Brand (Warm Scrapbook + Naturalistic Drift)

Phase 1: We're a longevity science company. We want warmth, approachability, and land-connection. Primary system: Warm Scrapbook (references/film-archetypes.md Section 6). Accent system: Naturalistic Drift (Section 7) for the landscape chapters. Output: cinematic-audit.md.

Phase 2: Design a 5-chapter storyboard. Chapters: Welcome, Science, Nature, Product, Community. Use Pinned Hero for ch1, Editorial Longread for ch2, Parallax Gallery for ch3, Chaptered Release for ch4, Landing Sequence for ch5. Reference references/scroll-patterns.md Sections 1, 8, 6, 5, and 10. Output: motion-storyboard.md.

Phase 3: Technical spec with warm palette progression (rose → peach → amber → sage → cream). GSAP + Lenis. Mobile (≤768px): keep touch-safe scroll-coupled motion — lerped image parallax + scroll-linked entrance reveals (no pinning, no 3D tilt); see references/mobile-motion.md. Reference performance-budget.md Section 3 (Mobile Degradation Matrix). Output: technical-spec.md.

Phase 4: Build Mode B. 5 chapters, organic editorial aesthetic. fal.ai for chapter images: historicalLayer: 'atelier', painterly botanical subjects. Demo mode for first run. Output: project directory.

Phase 5: Polish. Verify emotional arc matches Phase 1: welcome (warmth) → science (curiosity) → nature (awe) → product (trust) → community (belonging). Run 11-point checklist. Output: polish-report.md.

Example 3: Sci-Fi Game Reveal (Temporal Monument, Mode A)

Phase 1: We're launching a sci-fi game expansion. We want cosmic scale, event-level drama, layered realities. Visual system: Temporal Monument (references/film-archetypes.md Section 4). Audience: gamers, 70% desktop. Output: cinematic-audit.md.

Phase 2: Design a 7-chapter storyboard. Chapters: Teaser, World, Lore, Characters, Gameplay, Release, CTA. Use Pinned Hero for ch1, Chaptered Release for ch2 and ch3, 3D Product Orbit for ch5, Landing Sequence for ch7. Reference references/scroll-patterns.md Sections 1, 5, 5, 7, and 10. Max 7 layers in ch2 (the deepest chapter). Output: motion-storyboard.md.

Phase 3: Technical spec. Mode A output (single HTML). rAF-throttled scroll, no packages. 5-7 depth layers per chapter. 3D camera: rotateX(±4deg), translateZ(0 → -80px). Performance budget: all transform + opacity only, will-change on 3 elements max. Reference performance-budget.md Sections 1 and 2. Output: technical-spec.md.

Phase 4: Build Mode A. Single self-contained HTML. Near-black backgrounds, deep teal and crimson accents, heavy grain overlay. 7 chapters, each pinned 200-300vh. Title reveals: mask wipe, word stagger, letter-spacing scrub, scale-down entrance — varied per chapter per taste-guardrails.md Section 4.5. Reduced-motion fallback: static compositions. Progress HUD included. Output: index.html.

Phase 5: Polish the HTML. Verify: compositor-only scroll paths, < 5% dropped frames on 10s recording, reduced-motion shows all content, mobile <768px stacks with no pinning. No banned patterns from taste-guardrails.md Section 1. Output: polish-report.md.

What's in the Box

cinematic-scroll-skill/
├── SKILL.md                      # Agent contract (5-phase pipeline) [this file]
├── taste-guardrails.md           # Quality enforcement system (11 banned patterns,
│                                 #   cinematic vocabulary, pacing rules,
│                                 #   anti-convergence principles)
├── design.md                     # Design contract (token roles, motion, banned patterns)
├── tokens/                        # DTCG design tokens (core / motion / semantic) + build/
├── themes/                        # 11 per-visual-system theme overlays
├── manifest.json                 # Skill manifest (v2.3.5)
├── MODELS.md                     # fal.ai model menu and cost table
├── README.md                     # Human-facing overview
├── LICENSE                       # MIT
├── references/
│   ├── scroll-patterns.md        # 12 proven scroll patterns (Sections 1-12),
│   │                             #   each with use case, depth config, transition,
│   │                             #   mobile strategy, performance budget
│   ├── film-archetypes.md        # 11 visual systems (Sections 1-11):
│   │                             #   Symmetric Monument, Clinical Noir,
│   │                             #   Storybook Geometry, Temporal Monument,
│   │                             #   Atmospheric Sublime, Warm Scrapbook,
│   │                             #   Naturalistic Drift, Brutalist Kinetic,
│   │                             #   Liquid Chrome, Botanical Editorial,
│   │                             #   Data Cinematic — each → a themes/*.theme.json
│   ├── performance-budget.md     # 60fps production contract:
│   │                             #   frame budget, permitted/forbidden properties,
│   │                             #   will-change strategy, mobile degradation matrix
│   │                             #   (4 tiers), benchmark targets, asset budgets,
│   │                             #   11-point pre-launch monitoring checklist,
│   │                             #   GSAP-specific rules, failure modes
│   └── mobile-motion.md          # Touch-safe mobile motion recipe:
│                                 #   scroll-coupled lerped parallax + entrance
│                                 #   reveals, the iOS Safari animation-timeline
│                                 #   gotcha, vanilla (Mode A) + framer (Mode B)
│                                 #   sketches, reduced-motion gating
├── examples/
│   ├── PROMPTS.md               # 20+ trigger prompts (Mode A and B)
│   ├── GETTING_STARTED.md       # fal.ai setup walkthrough
│   └── KNOWN_ISSUES.md          # QA log of known issues and fixes
├── templates/nextjs/            # Next.js App Router template:
│                                 #   package.json, ChapterScene.tsx (7-layer scene),
│                                 #   ChapterDemoVisual.tsx (CSS-only fallback),
│                                 #   EditionsPage.tsx (orchestrator),
│                                 #   fal proxy routes, fal client/lib/scripts,
│                                 #   SmoothScrollProvider, use-device hooks,
│                                 #   globals.css with fluid type scale,
│                                 #   tailwind.config.ts, tsconfig.json
└── assets/                      # Shared static assets

Legal and Originality Rules

• Do not reproduce the Shopify logo, screenshots, copy, proprietary illustrations, exact section design, or exact visual scene.
• Do not generate images that imitate a living artist by name.
• Do not bake readable UI copy into generated images unless the user specifically asks and the target model supports reliable text.
• Build UI text, labels, nav, cards, numbers, and feature lists as HTML/CSS so they remain editable, accessible, and crisp.
• Use references only as art-direction benchmarks — chaptered release storytelling, not clone targets.
• If the user asks to clone a proprietary site exactly, respond by making an original system that uses the reference as inspiration.

Anti-Patterns

Do NOT use this skill for:

• "Build a basic hero + features + pricing landing page."
• "Generate a WordPress theme."
• Ordinary SaaS landing pages, CRUD dashboards, or simple brochure sites — unless the user explicitly asks for a cinematic / editorial treatment.
• "Regenerate all templates from scratch without reading bundled files."
• "Give me motion ideas only, no code." (The skill must output runnable artifacts.)