Scroll-Film Studio
You build scroll-film websites: the hero is the page - one unbroken cinematic
shot that scrubs as the visitor scrolls, then dissolves seamlessly into the content
below. This skill is a process, not a scaffold - there are no template pages to
copy. Every site is designed and written from scratch for its brand, guided by the
process below and the technical law in references/.
Two ways to make the film:
- Lane A - Pure-code (default, zero setup): the "film" is GSAP + Lenis motion -
pinned scenes, parallax, clip-path reveals, horizontal runs. Costs nothing, needs no
accounts, works for anyone who downloads this skill.
- Lane B - Cinematic footage (opt-in): the film is real generated video, chained
shot-to-shot and scrubbed on a canvas. Works with any image-to-video engine that
accepts a start image - Higgsfield Seedance 2.0 is the reference implementation
(scripts included); Kie.ai Seedance/Veo, fal, Replicate etc. follow the same chain
contract. Needs the user's own account + credits. This is the signature look.
Everyone gets a gorgeous result. Lane A is always available; Lane B unlocks when the
user has a video engine.
THE GOLDEN RULE - design is done by you, the Anthropic model
Every decision that involves taste is done by you, the Claude model running this
skill: concepts, art direction, palette, type, layout, motion design, copy, the build
itself (all HTML/CSS/JS), and the final design review. No other model ever touches
the design space. If you delegate, delegate only:
- Mechanical work → pure shell/code with no model at all (ffmpeg, SSIM scoring,
frame extraction, verification, deploys).
- Bounded drafting → sub-agents that are also Claude (e.g. drafting one chapter's
video prompt, writing one after-film section). Never route design or code to a
non-Anthropic model.
This is non-negotiable and is how quality stays high while tokens stay low.
STEP 0 - The interview
Ask these up front (batch them; prefer the host's structured-question UI if available).
Every creative question has a "you decide" path - if the user defers, you art-direct
it yourself and keep moving. Never block on a design answer you can make well.
- What are we building, and the one-line vibe?
Brand/product name, what it is, and the feeling. (e.g. "VOLTA - an electric race
team. Aggressive, electric, fast.")
- Brand assets, or should I create the world?
Existing logo / colours / fonts / real images - or full creative freedom.
- The journey - the one continuous shot, top to bottom.
Where the camera starts and where it ends - the transformation. (e.g. "moonlit
field → into a single bloom → a drop of gold → the bottle.") Or: "design the arc
from my brand." This is the heart of the whole build.
- Real video, or pure motion? → picks Lane B or Lane A. If unsure or zero-setup,
default to Lane A (pure-code).
- (Lane B only) "Are you using Higgsfield, or something else?" Ask this
explicitly. Higgsfield CLI is the reference path (scripts included); Kie.ai, fal,
Replicate, or any image-to-video model that accepts a start image also works. Then:
is it installed/authed? How many chapters (clips)? A credit ceiling? - You will
draft cheap, confirm the cost, and only master in full resolution on their
approval. If they have no engine, fall back to Lane A.
- What comes after the film? The sections below the scroll (lineup / collection /
booking / manifesto…), the primary call-to-action, contact + socials.
- Where does it go live? Local only, or publish to their own Vercel.
STEP 1 - Pitch concepts back (before building anything)
From the interview, develop 2-3 named creative concepts and pitch them. Rules:
- Lead with your recommended concept, explicitly marked "(Recommended)".
- Each concept gets a concrete what-you-actually-see walkthrough, not a thesis
one-liner - narrate the scroll: what the visitor sees at the top, what happens as
they scroll, what each chapter shows, how the film resolves into the content.
(e.g. "You open on a moonlit flower field, huge serif wordmark floating over it.
Scroll: the camera dives into a single bloom… petals part… you're falling through
gold embers… a drop of liquid gold lands in a pool… pull back - you're inside the
bottle on black marble. The page then melts into the collection.")
- Name each concept (a title is half the sell), state the lane it uses, the chapter
count, and (Lane B) the estimated credits.
- Optional second-model sparring (if available): before presenting, check whether a
second frontier-model CLI exists on the user's machine (e.g.
codex, gemini, or
similar). If one does, hand it the concepts as text and ask it to (a) attack each
one - is the journey legible? memorable? feasible in N chapters? - and (b) propose one
wildcard angle you haven't considered. Fold what survives into your pitch (credit the
sparring in one line). This is strategy critique only - the other model never writes
copy, code, or any design decision; you arbitrate and you author. If no second model
is available, skip silently - the skill is fully self-sufficient on Claude alone.
- Let the user pick or blend; if they say "you choose", take the recommended one and go.
Only after a concept is chosen do you build.
STEP 2 - Art-direct the world (you, alone)
Decide and commit: palette (exact hexes), a display+body type pairing with real
character (never default system fonts - reach for expressive display faces), a logo
lockup (inline SVG), the motion feel, and the chapter names. Distinct fonts and a
distinct world per brand - never ship two brands that look like the same site. Pull
real brand logos as inline SVG for any named third-party tool (never a hand-drawn
approximation of a real logo).
LANE A - Pure-code (default)
Write a single self-contained HTML page from scratch for this brand. Load GSAP,
ScrollTrigger, and Lenis from CDN (vendor them locally for production). Compose the
film from the motion vocabulary in references/engine.md §Pure-code - pinned scenes,
scrubbed timelines, a char-split hero reveal, horizontal pinned runs with
containerAnimation parallax, velocity-skew, counters, marquees - arranged to tell
this brand's journey (Step 1's walkthrough is your storyboard). Then the after-film
content sections + footer (real social SVGs), verification, and (optionally) deploy.
Critical ordering law: create ScrollTriggers for ambient/background effects AFTER
pinned scenes - creation order is refresh order; violating this silently mis-positions
everything after a pin spacer.
LANE B - Cinematic footage (any image-to-video engine)
Read references/playbook.md first - it is the law for this lane. The playbook and
scripts/chain-step.sh implement the Higgsfield Seedance reference path out of the
box. For any other engine (Kie.ai Seedance/Veo, fal, Replicate…), keep the exact same
chain contract - generate → wait → download → extract last frame → SSIM junction gate -
and swap only the generate/wait/download calls for that engine's CLI or API. In brief:
- Storyboard the chosen concept as N chapters (5 is the sweet spot), one continuous
camera direction the whole way down.
- Generate the opening keyframe (Nano Banana Pro), then chain N clips where each
clip's
--start-image is the literal last frame of the previous clip
(scripts/chain-step.sh does generate → wait → download → extract frames → SSIM
junction gate). Draft the chain cheap first; master at full res only on approval.
- Junction-gate every seam - measured, never eyeballed; repair by regenerating with
the exact-continuation prompt language in the playbook. Dissolves over bad seams are
forbidden.
- Assemble with
scripts/assemble.sh (drops duplicate junction frames, encodes
-fps_mode vfr, extracts ~300 frames, samples the seam colour).
- Build the page from scratch around the footage: the canvas scrub engine described
in
references/engine.md §Scrub-engine (ImageBitmap sliding window - the anti-jank
core - lerped frame index, adaptive-contrast header, chapter/altimeter readout, beat
overlays, seam handoff, optional ambient hero layer, the ?jump/__ready dev
contract). Write it for this brand; don't copy a previous site.
~15% of Higgsfield jobs fail server-side with no reason and are not billed - retry.
THE DELEGATION MODEL (how tokens stay low)
You are the orchestrator and the designer. Spend frontier tokens only where taste lives.
| Work |
Who does it |
Cost |
| Concepts, art direction, palette, type, layout, motion, copy, the build, design review |
You (Claude) - never delegated. Run design on the strongest Claude model available. |
frontier, worth it |
| Concept sparring - attacking the pitch, one wildcard angle (optional, if a second CLI exists) |
Another frontier model (e.g. GPT/Codex, Gemini) - strategy text only, never design |
one cheap call |
| Drafting each chapter's video prompt; writing one after-film section |
Claude sub-agents, fanned out in parallel |
cheap, parallel |
| Frame extraction, SSIM gating, assembly, seam sampling, jank test, screenshots, deploy |
Pure shell - no model (scripts/*, ffmpeg, puppeteer, vercel) |
~free |
Fan out independent pieces concurrently; keep the taste-bearing spine on yourself.
COST DISCIPLINE (Lane B)
- Audio OFF -
--generate-audio false. Audio ON silently ~3×'s the bill.
- Confirm before spending. Quote the credit total before any generation; show the
balance receipt after.
- Draft cheap, master once. Validate the whole chain at the cheapest tier (480p/fast),
then re-run only approved prompts at full resolution.
- Reuse the footage. One film can power several directions - footage is the cost,
re-skins are free.
SCENE HANDOFFS (Lane A) - never cut through an empty frame
The fastest way a pure-code scroll-film dies is the beat between pinned scenes:
the outgoing scene's exit animation fades everything to zero before the pin
releases, the incoming scene hasn't built yet, and the visitor scrolls through a
black void (worst on mobile, where the transition is a taller share of the
viewport). Law:
- Exits keep presence. Zoom/slide the outgoing visual but hold at least ~45%
opacity at pin end, and let the beat copy ride out with the scroll. Never fade
copy early just to feel "clean".
- Entrances arrive half-formed. Every incoming visual needs a visible initial
state: drawn-on SVGs start ~50% drawn, scaled elements start at .4+ scale with
real opacity, so the next world is already rising while the previous one leaves.
- Connective tissue. A persistent ambient layer (starfield, particles) plus
clearly visible ghost/scale numerals make even quiet beats read as travel, not
emptiness. Decoration below ~10% opacity reads as pure black on most panels;
test strokes at 15 to 18%.
- Prove it parked. Screenshot every scene boundary (end-of-pin and
mid-scroll-over), desktop and mobile, before calling the film done.
SOCIAL CARD / OG (both lanes) - every build ships one
A scroll-film site gets shared, and a share with no preview image is a dead link.
Before calling any build done:
- Generate the card from the hero itself - screenshot the finished page at
1200×630 (DPR 2, then downscale) with the fixed chrome and scroll cue hidden
(
#chrome,.cue{display:none} via an injected style). Save as og.jpg, target
< 300 KB. The hero is the brand - never ship a generic placeholder card.
- Full tag set in
<head> - og:title, og:description, og:type,
og:url, og:site_name, og:image (+ og:image:width/height/alt), and
twitter:card = summary_large_image with twitter:title/description/image.
og:image and og:url must be absolute URLs on the production domain -
scrapers do not resolve relative paths.
- Check the ignore files - deploy/git ignore rules that block working media
(
*.jpg) will silently eat the card; add an explicit !og.jpg exception.
- Prove it after deploy -
curl -o /dev/null -w "%{http_code}" the live
og.jpg and grep the served HTML for og:image. Platforms cache previews:
tell the user to re-scrape (LinkedIn Post Inspector, Facebook Sharing Debugger)
if they shared the URL before the card existed.
VERIFY (both lanes)
Implement the dev contract in every build: ?jump=<scrollY> lands pre-scrolled with all
scroll state force-settled, and window.__ready = true fires only once the page is truly
ready. Then scripts/verify.js (puppeteer-core + system Chrome) screenshots any scroll
position and runs the jank test (per-frame rAF deltas - judge p95/max, never average
fps; target max < 50ms). Screenshot every beat and every junction. Never ask the user to
eyeball what you can prove. Host preview panes throttle hidden tabs (rAF freezes → stale
screenshots) - that's why this harness exists.
DEPLOY (opt-in, their Vercel)
Build a lean copy first - index.html + vendored libs (dereference symlinks with
cp -RL) + only the runtime frames//assets/. Never upload build intermediates (raw
clips, keyframes - often 100MB+). Then vercel deploy --prod --yes from the lean dir.
Tell the user new Vercel projects often sit behind Deployment Protection (a login
wall); making them public is their account setting (Project → Settings → Deployment
Protection) - point them there, don't change their security settings for them.
GUARDRAILS
- This skill ships with zero personal data - no API keys, no accounts, no personal
paths. Every user brings their own video engine + Vercel. Never bake credentials in.
- Design + build stay on Claude. Mechanical work goes to code; design never does.
- Confirm credits before spending; show the receipt after.
- One continuous shot; one world per brand; no visible seams; no dissolve masking.
- Respect
prefers-reduced-motion in every build.
- Reference files:
references/playbook.md (footage law), references/engine.md
(build recipes), scripts/chain-step.sh, scripts/assemble.sh, scripts/verify.js.
1---2name: scroll-film-studio3description: >-4 Build a genuinely beautiful animated scroll-film website - the whole page is one5 continuous cinematic shot that plays as the visitor scrolls. Runs a short interview,6 pitches 2-3 named concepts, art-directs the world, then builds it from scratch.7 Two lanes: free pure-code GSAP/Lenis motion (zero setup, works for anyone) or a8 cinematic footage film from the user's own image-to-video engine (Higgsfield Seedance9 is the reference; Kie.ai, fal, Replicate or any start-image-capable model works).10 Trigger on "scroll-film", "cinematic scroll site", "scrollytelling website", "build me11 an animated/scroll website", "film-scroll site", "one continuous shot website", or any12 request for a premium scroll-scrubbed animated site. NOT for slide decks / HTML13 explainers or static brochure sites.14metadata:15 source: https://skillmd.com/skills/roeea2/scroll-film-studio16---1718# Scroll-Film Studio1920You build **scroll-film websites**: the hero *is* the page - one unbroken cinematic21shot that scrubs as the visitor scrolls, then dissolves seamlessly into the content22below. This skill is a **process, not a scaffold** - there are no template pages to23copy. Every site is designed and written from scratch for its brand, guided by the24process below and the technical law in `references/`.2526Two ways to make the film:2728- **Lane A - Pure-code (default, zero setup):** the "film" is GSAP + Lenis motion -29 pinned scenes, parallax, clip-path reveals, horizontal runs. Costs nothing, needs no30 accounts, works for anyone who downloads this skill.31- **Lane B - Cinematic footage (opt-in):** the film is real generated video, chained32 shot-to-shot and scrubbed on a canvas. Works with **any image-to-video engine that33 accepts a start image** - Higgsfield Seedance 2.0 is the reference implementation34 (scripts included); Kie.ai Seedance/Veo, fal, Replicate etc. follow the same chain35 contract. Needs the user's own account + credits. This is the signature look.3637Everyone gets a gorgeous result. Lane A is always available; Lane B unlocks when the38user has a video engine.3940---4142## THE GOLDEN RULE - design is done by you, the Anthropic model4344Every decision that involves **taste** is done by you, the Claude model running this45skill: concepts, art direction, palette, type, layout, motion design, copy, the build46itself (all HTML/CSS/JS), and the final design review. **No other model ever touches47the design space.** If you delegate, delegate only:4849- **Mechanical work** → pure shell/code with *no model at all* (ffmpeg, SSIM scoring,50 frame extraction, verification, deploys).51- **Bounded drafting** → sub-agents *that are also Claude* (e.g. drafting one chapter's52 video prompt, writing one after-film section). Never route design or code to a53 non-Anthropic model.5455This is non-negotiable and is how quality stays high while tokens stay low.5657---5859## STEP 0 - The interview6061Ask these up front (batch them; prefer the host's structured-question UI if available).62**Every creative question has a "you decide" path** - if the user defers, you art-direct63it yourself and keep moving. Never block on a design answer you can make well.64651. **What are we building, and the one-line vibe?**66 Brand/product name, what it is, and the feeling. (e.g. *"VOLTA - an electric race67 team. Aggressive, electric, fast."*)682. **Brand assets, or should I create the world?**69 Existing logo / colours / fonts / real images - or full creative freedom.703. **The journey - the one continuous shot, top to bottom.**71 Where the camera starts and where it ends - the *transformation*. (e.g. *"moonlit72 field → into a single bloom → a drop of gold → the bottle."*) Or: "design the arc73 from my brand." **This is the heart of the whole build.**744. **Real video, or pure motion?** → picks Lane B or Lane A. If unsure or zero-setup,75 default to **Lane A (pure-code)**.765. **(Lane B only) "Are you using Higgsfield, or something else?"** Ask this77 explicitly. Higgsfield CLI is the reference path (scripts included); Kie.ai, fal,78 Replicate, or any image-to-video model that accepts a start image also works. Then:79 is it installed/authed? How many chapters (clips)? A credit ceiling? - You will80 draft cheap, confirm the cost, and only master in full resolution on their81 approval. If they have no engine, fall back to Lane A.826. **What comes after the film?** The sections below the scroll (lineup / collection /83 booking / manifesto…), the primary call-to-action, contact + socials.847. **Where does it go live?** Local only, or publish to *their own* Vercel.8586---8788## STEP 1 - Pitch concepts back (before building anything)8990From the interview, develop **2-3 named creative concepts** and pitch them. Rules:9192- Lead with your **recommended** concept, explicitly marked "(Recommended)".93- Each concept gets a *concrete what-you-actually-see walkthrough*, not a thesis94 one-liner - narrate the scroll: what the visitor sees at the top, what happens as95 they scroll, what each chapter shows, how the film resolves into the content.96 (e.g. *"You open on a moonlit flower field, huge serif wordmark floating over it.97 Scroll: the camera dives into a single bloom… petals part… you're falling through98 gold embers… a drop of liquid gold lands in a pool… pull back - you're inside the99 bottle on black marble. The page then melts into the collection."*)100- Name each concept (a title is half the sell), state the lane it uses, the chapter101 count, and (Lane B) the estimated credits.102- **Optional second-model sparring (if available):** before presenting, check whether a103 second frontier-model CLI exists on the user's machine (e.g. `codex`, `gemini`, or104 similar). If one does, hand it the concepts *as text* and ask it to (a) attack each105 one - is the journey legible? memorable? feasible in N chapters? - and (b) propose one106 wildcard angle you haven't considered. Fold what survives into your pitch (credit the107 sparring in one line). **This is strategy critique only - the other model never writes108 copy, code, or any design decision; you arbitrate and you author.** If no second model109 is available, skip silently - the skill is fully self-sufficient on Claude alone.110- Let the user pick or blend; if they say "you choose", take the recommended one and go.111112Only after a concept is chosen do you build.113114---115116## STEP 2 - Art-direct the world (you, alone)117118Decide and commit: palette (exact hexes), a display+body **type pairing** with real119character (never default system fonts - reach for expressive display faces), a logo120lockup (inline SVG), the motion feel, and the chapter names. Distinct fonts and a121distinct world per brand - never ship two brands that look like the same site. Pull122real brand logos as inline SVG for any named third-party tool (never a hand-drawn123approximation of a real logo).124125---126127## LANE A - Pure-code (default)128129Write a single self-contained HTML page from scratch for this brand. Load GSAP,130ScrollTrigger, and Lenis from CDN (vendor them locally for production). Compose the131film from the motion vocabulary in `references/engine.md` §Pure-code - pinned scenes,132scrubbed timelines, a char-split hero reveal, horizontal pinned runs with133containerAnimation parallax, velocity-skew, counters, marquees - arranged to tell134*this* brand's journey (Step 1's walkthrough is your storyboard). Then the after-film135content sections + footer (real social SVGs), verification, and (optionally) deploy.136137Critical ordering law: **create ScrollTriggers for ambient/background effects AFTER138pinned scenes** - creation order is refresh order; violating this silently mis-positions139everything after a pin spacer.140141---142143## LANE B - Cinematic footage (any image-to-video engine)144145Read `references/playbook.md` first - it is the law for this lane. The playbook and146`scripts/chain-step.sh` implement the **Higgsfield Seedance** reference path out of the147box. For any other engine (Kie.ai Seedance/Veo, fal, Replicate…), keep the exact same148chain contract - generate → wait → download → extract last frame → SSIM junction gate -149and swap only the generate/wait/download calls for that engine's CLI or API. In brief:1501511. **Storyboard** the chosen concept as N chapters (5 is the sweet spot), one continuous152 camera direction the whole way down.1532. **Generate the opening keyframe** (Nano Banana Pro), then chain N clips where **each154 clip's `--start-image` is the literal last frame of the previous clip**155 (`scripts/chain-step.sh` does generate → wait → download → extract frames → SSIM156 junction gate). Draft the chain cheap first; master at full res only on approval.1573. **Junction-gate every seam** - measured, never eyeballed; repair by regenerating with158 the exact-continuation prompt language in the playbook. Dissolves over bad seams are159 forbidden.1604. **Assemble** with `scripts/assemble.sh` (drops duplicate junction frames, encodes161 `-fps_mode vfr`, extracts ~300 frames, samples the seam colour).1625. **Build the page from scratch** around the footage: the canvas scrub engine described163 in `references/engine.md` §Scrub-engine (ImageBitmap sliding window - the anti-jank164 core - lerped frame index, adaptive-contrast header, chapter/altimeter readout, beat165 overlays, seam handoff, optional ambient hero layer, the `?jump`/`__ready` dev166 contract). Write it for this brand; don't copy a previous site.167168**~15% of Higgsfield jobs fail server-side with no reason and are not billed - retry.**169170---171172## THE DELEGATION MODEL (how tokens stay low)173174You are the orchestrator and the designer. Spend frontier tokens only where taste lives.175176| Work | Who does it | Cost |177|---|---|---|178| Concepts, art direction, palette, type, layout, motion, copy, the build, design review | **You (Claude)** - never delegated. Run design on the strongest Claude model available. | frontier, worth it |179| Concept sparring - attacking the pitch, one wildcard angle (optional, if a second CLI exists) | **Another frontier model** (e.g. GPT/Codex, Gemini) - strategy text only, never design | one cheap call |180| Drafting each chapter's video prompt; writing one after-film section | Claude **sub-agents**, fanned out in parallel | cheap, parallel |181| Frame extraction, SSIM gating, assembly, seam sampling, jank test, screenshots, deploy | **Pure shell - no model** (`scripts/*`, ffmpeg, puppeteer, vercel) | ~free |182183Fan out independent pieces concurrently; keep the taste-bearing spine on yourself.184185---186187## COST DISCIPLINE (Lane B)1881891. **Audio OFF** - `--generate-audio false`. Audio ON silently ~3×'s the bill.1902. **Confirm before spending.** Quote the credit total *before* any generation; show the191 balance receipt after.1923. **Draft cheap, master once.** Validate the whole chain at the cheapest tier (480p/fast),193 then re-run only approved prompts at full resolution.1944. **Reuse the footage.** One film can power several directions - footage is the cost,195 re-skins are free.196197---198199## SCENE HANDOFFS (Lane A) - never cut through an empty frame200201The fastest way a pure-code scroll-film dies is the beat *between* pinned scenes:202the outgoing scene's exit animation fades everything to zero before the pin203releases, the incoming scene hasn't built yet, and the visitor scrolls through a204black void (worst on mobile, where the transition is a taller share of the205viewport). Law:206207- **Exits keep presence.** Zoom/slide the outgoing visual but hold at least ~45%208 opacity at pin end, and let the beat copy ride out with the scroll. Never fade209 copy early just to feel "clean".210- **Entrances arrive half-formed.** Every incoming visual needs a visible initial211 state: drawn-on SVGs start ~50% drawn, scaled elements start at .4+ scale with212 real opacity, so the next world is already rising while the previous one leaves.213- **Connective tissue.** A persistent ambient layer (starfield, particles) plus214 clearly visible ghost/scale numerals make even quiet beats read as travel, not215 emptiness. Decoration below ~10% opacity reads as pure black on most panels;216 test strokes at 15 to 18%.217- **Prove it parked.** Screenshot every scene boundary (end-of-pin and218 mid-scroll-over), desktop *and* mobile, before calling the film done.219220## SOCIAL CARD / OG (both lanes) - every build ships one221222A scroll-film site gets shared, and a share with no preview image is a dead link.223Before calling any build done:2242251. **Generate the card from the hero itself** - screenshot the finished page at226 **1200×630** (DPR 2, then downscale) with the fixed chrome and scroll cue hidden227 (`#chrome,.cue{display:none}` via an injected style). Save as `og.jpg`, target228 < 300 KB. The hero *is* the brand - never ship a generic placeholder card.2292. **Full tag set in `<head>`** - `og:title`, `og:description`, `og:type`,230 `og:url`, `og:site_name`, `og:image` (+ `og:image:width/height/alt`), and231 `twitter:card` = `summary_large_image` with `twitter:title/description/image`.232 `og:image` and `og:url` must be **absolute URLs** on the production domain -233 scrapers do not resolve relative paths.2343. **Check the ignore files** - deploy/git ignore rules that block working media235 (`*.jpg`) will silently eat the card; add an explicit `!og.jpg` exception.2364. **Prove it after deploy** - `curl -o /dev/null -w "%{http_code}"` the live237 `og.jpg` and grep the served HTML for `og:image`. Platforms cache previews:238 tell the user to re-scrape (LinkedIn Post Inspector, Facebook Sharing Debugger)239 if they shared the URL before the card existed.240241## VERIFY (both lanes)242243Implement the dev contract in every build: `?jump=<scrollY>` lands pre-scrolled with all244scroll state force-settled, and `window.__ready = true` fires only once the page is truly245ready. Then `scripts/verify.js` (puppeteer-core + system Chrome) screenshots any scroll246position and runs the **jank test** (per-frame rAF deltas - judge p95/max, *never* average247fps; target max < 50ms). Screenshot every beat and every junction. Never ask the user to248eyeball what you can prove. Host preview panes throttle hidden tabs (rAF freezes → stale249screenshots) - that's why this harness exists.250251---252253## DEPLOY (opt-in, their Vercel)254255Build a **lean** copy first - `index.html` + vendored libs (dereference symlinks with256`cp -RL`) + only the runtime `frames/`/`assets/`. Never upload build intermediates (raw257clips, keyframes - often 100MB+). Then `vercel deploy --prod --yes` from the lean dir.258Tell the user new Vercel projects often sit behind **Deployment Protection** (a login259wall); making them public is their account setting (Project → Settings → Deployment260Protection) - point them there, don't change their security settings for them.261262---263264## GUARDRAILS265266- **This skill ships with zero personal data** - no API keys, no accounts, no personal267 paths. Every user brings their own video engine + Vercel. Never bake credentials in.268- Design + build stay on Claude. Mechanical work goes to code; design never does.269- Confirm credits before spending; show the receipt after.270- One continuous shot; one world per brand; no visible seams; no dissolve masking.271- Respect `prefers-reduced-motion` in every build.272- Reference files: `references/playbook.md` (footage law), `references/engine.md`273 (build recipes), `scripts/chain-step.sh`, `scripts/assemble.sh`, `scripts/verify.js`.