Scrollcraft Agent Skill

scrollcraft

Scroll is the only input every visitor already knows how to use. This skill

treats it as a timeline: the wheel is a scrubber, the page is a film with real

text on top, and each section behaves differently enough that the visitor keeps

going to find out what the next one does.

What you produce: an interview brief, a page grammar, a customer-journey map,

a feeling curve with one engineered peak, a scroll score, one signature move,

generated assets, one real HTML page on a token-driven design floor, and a strip

of screenshots proving it holds up at every scroll position.

What this is not

It is not "generate a flythrough and drop text on it." That approach produces

one device applied to a whole page, and every site built that way is

recognisable at a glance: same claymation diorama, same centred copy, same

01 / 06 counter, same "scroll to explore" nudge. Five sections that behave

identically are one section shown five times.

Four rules follow from that, and they are the spine of this skill:

• Variety is the product. A page uses at least four device families and

never the same device twice in a row. Read references/devices.md.

• The world is photographic unless the brand is genuinely illustrated.

Soft matte low-poly clay diorama is banned as a default. Read

references/worlds.md.

• No continuous chain. A single unbroken camera flight is the most

expensive and most fragile thing you can build, and it exists only to hide

cuts between scenes. Vary the device instead and the cut disappears for free,

because the visitor is not watching one film. Chain only when the brief is

literally "one continuous journey."

• A different world is not a different page. The device kit varies how a

page looks. Structure is a separate axis, and it has to be decided

deliberately or every build inherits the same skeleton. The first four builds

did exactly that. Read references/uniqueness.md.

Step 0: The interview

Always interview the human before generating anything. Not a brief you

inferred from the brand name, not a plan you present for approval. Actual

questions, asked, answered, written down. A page built from assumptions comes

back looking like the last page built from assumptions.

The skill is a range instrument, not a house style. The human brings intent and

whatever assets they own; the interview is where that turns into the right kind

of page: one unbroken world, distinct scenes, printed chapters, a live surface.

The skill can do any of them. The interview decides which.

Keep it short. Eight questions, asked in one pass:

• Vibe in three to five words, plus up to three references from any medium.

A film, an album cover, a shop, a magazine, a game. Not "sites you like":

naming sites is how a page ends up looking like an existing site.

• The scroll journey, section by section, in their words. What the visitor

should hit first, what comes next, what the last thing is. Their sequence,

not a menu you offered.

• The energy curve. Where it should feel calm, where it should feel

intense. A page that is loud the whole way is as flat as one that is quiet

the whole way.

• **How should someone feel while scrolling, stage by stage, and what is the

ONE moment they should remember?** Energy is loudness. This is emotion, and

the two do not line up: on a loud page the quiet act can be the most intense.

The stage-by-stage answer becomes the feeling curve, the one moment becomes

the peak. Both are required in BRIEF.md. See references/feel.md.

• One thing this site should do that no site they have seen does. This is

the seed of the signature move. Push for a real answer; "be memorable" is not

one.

• How far from premium-minimal they want to go. Offer the range in

uniqueness.md §5: brutalist, maximalist, playful,

retro, dense, editorial, premium-minimal. Their answer governs the aesthetic

family, not your taste.

• One unbroken world, or distinct scenes? Should the whole page feel like

one continuous place the scroll flies through (worldflight, see

references/worldflight.md), or like separate

scenes, chapters, or cuts? This is the single biggest structural fork, and it

is their call, not a device you pick later. Offer both plainly; neither is

the default.

• What assets do they already have? Footage, photos, product shots, a

brand kit, clips of themselves. Real assets anchor the world and cut

generation cost; the answer decides what gets graded and encoded versus

generated. "Nothing" is a fine answer and means a fully generated world.

Write the answers into <workspace>/builds/<name>/BRIEF.md before any act planning, in

their words, not paraphrased into marketing prose. Everything downstream reads

from that file.

BRIEF.md must contain, at minimum:

• The eight interview answers, verbatim.
• The feeling curve. One line per act: the emotion, then what on screen

causes it. Written before the acts exist, added to as the score fills in.

• The peak. The one moment, written as the sentence a visitor would say to

a friend, plus which act it lives in.

• The completed tell-someone sentence. "It's the site where ___", filled

with an experience, not a device name.

• Any authored silence, so the verification pass can tell it from dead scroll.

references/feel.md is the spec for all four.

If the human is genuinely unreachable and the run is fully autonomous, write

BRIEF.md yourself: answer all eight questions in the brand's voice, mark the file

Self-authored, not interviewed at the top, and say so in the final report. A

self-authored brief is a fallback, never the plan.

Bootstrap

Environment, not a stage of the work. Do it once the interview is answered and

before Step 1.

Run the preflight rather than checking by hand. It knows the failure modes

that otherwise surface later as misleading errors, chiefly a stripped ffmpeg

that reports a missing filter as a syntax error in your command:

node <skill>/scripts/doctor.mjs

It reports node, a full ffmpeg build, playwright and Chrome, the API key, and

the resolved workspace. Required failures exit non-zero. Say plainly which items

are missing rather than working around them silently.

The workspace

Builds and the fingerprint registry live in one directory, and **it is resolved,

never assumed**:

node <skill>/scripts/workspace.mjs --ensure     # prints it, creates it, seeds the registry

Resolution order, first hit wins:

• SCROLLCRAFT_HOME
• the nearest .scrollcraft.json walking up from the cwd, { "workspace": "..." }
• <project root>/scrollcraft, where the project root is the nearest ancestor

holding a .git

So a build folder is <workspace>/builds/<name>/ and the registry is

<workspace>/FINGERPRINTS.md. The registry starts empty: the gate exists to

stop you repeating yourself, so your first build has nothing to clear.

If you already keep builds somewhere else, drop a .scrollcraft.json at your

project root pointing at it and nothing moves.

The rest

• KIE_AI_API_KEY, only if you are generating assets. A build from the

user's own photos and footage needs no key and no spend, and that is a

first-class route, not a fallback. Confirm balance with

node <skill>/scripts/kie.mjs probe. A still costs cents and a 5s clip costs

more; a six-act page with two clips is a small spend, not a large one.

• A brand kit if one exists (colours, logo, type, existing product shots). If

the brand has a folder in this repo, read it before generating anything, and

obey its hard rules. A brand that forbids invented numbers means no stat

counters, however good they look.

Copy engine/scrollcraft.js and engine/scrollcraft.css into the build folder.

Never edit the engine per-project; it is the mechanism. Theme it with tokens and

write your own markup.

Step 1: The brief, journey first

The subject is the user's to state. Ask it open, in plain prose, never as a

fabricated multiple-choice list of industries: a made-up menu biases them and

reads as you deciding their business for them.

Step 0 already covered vibe, sequence, energy and range. Do not ask any of it

again. Ask only what you cannot sensibly default:

• What is this, and who is it for? One or two sentences in their words.
• What must the visitor believe by the end? The single sentence the page

exists to install. Not a feature list. If they give three, make them pick.

• What does the visitor do next? One action. One label for it, used

everywhere on the page.

• What do you already have? Logo, palette, photography, product shots,

footage, a brand doc. Real assets beat generated ones every time.

• Art direction: offer the worlds in references/worlds.md

as a real choice, and say they can go their own way.

Then write the journey before anything else: four to seven beats, each one a

shift in what the visitor knows or feels.

1  Recognition   they see their own morning
2  Tension       the cost of it, named plainly
3  Turn          the thing that changes
4  Substance     why it holds up
5  Range         what they can choose
6  Commitment    the one action

Beats are the spine. Sections serve beats; a section that serves no beat is cut,

however nice the shot is. Show the journey to the user and get it right before

generating a single asset, because assets are the expensive part and the journey

determines every one of them.

Step 2: Grammar, gate, then score

Three things in order, and the first two come before any act planning. Full

detail in references/uniqueness.md.

Pick a grammar. Eight of them, and they are mutually exclusive because each

one forbids things the others require. Filmic one-shot is the one the first four

builds all used, so choosing it again means saying in the report why the other

seven did not fit the interview. Nav, hero and close all follow from the

grammar; they are not decided separately.

Invent the signature move. One bespoke interaction that lives on this site

alone, coded in the page, not a parameter change to a kit device. Question 5 of

the interview is the seed. The engine stays untouched.

Run the fingerprint gate. Read your registry at

<workspace>/FINGERPRINTS.md (see The workspace in Bootstrap; run

node <skill>/scripts/workspace.mjs to print the path). The planned build must differ

from every existing row on at least 4 of 6 dimensions: grammar, nav

treatment, hero device, act-sequence shape, close pattern, signature move. Four

against each row individually. If it fails, change the plan, not the log.

Write the feeling curve before the score table. One line per act: the

emotion, then what causes it. Curve first, acts second, because a device chosen

before the feeling is a device looking for a reason. Two adjacent acts with the

same feeling means one is filler, and it is cheaper to cut it here than after

the assets exist. Name the peak in the same pass and give it the largest span on

the page. Full method in references/feel.md.

Then assign each beat a device. Do it deliberately and write it down as a table:

| Beat | Device | Why this one |

|---|---|---|

| Recognition | scrub | The camera moving under the reader's own hand is the strongest possible open |

| Tension | pin + kinetic | Copy assembles line by line while the frame holds still |

| Turn | reveal | A wipe is a change of state, which is what this beat is |

| Substance | scrub (macro) | Texture at a scale the eye cannot get otherwise |

| Range | pan | Lateral travel reads as "options", vertical reads as "argument" |

| Commitment | pin + pointer | The page stops moving and starts responding |

That table is a filmic score. It is the right shape for one grammar and the

wrong shape for the other seven, so read your grammar's leans-on and bans list

before filling in a row.

Checks before you build:

• The grammar's bans hold. A grammar that forbids pin forbids it here too,

however well it would have worked.

• Four or more distinct device families. Fewer means the page has one idea.
• No device family twice in a row.
• At most two scrub acts. Video is the heaviest thing on the page, and the

third one stops being a surprise.

• No two adjacent acts carry the same feeling. If they do, one is filler.
• One act is the peak and it has the largest span by a visible margin. The act

before it is quieter than it is.

• Every act earns its scroll span. Total page length 8 to 14 viewport-heights.

Longer is not more immersive, it is slower.

• The act count and total length do not land in the 6-to-7 acts at 13.6-13.8vh

band that all four prior builds hit. That band is a fingerprint dimension now.

Step 3: Generate the assets

Full pipeline, prompt scaffolds and model notes: references/assets.md.

Short version:

node <skill>/scripts/kie.mjs still "<style preamble>\n\n<scene>" out/01-hero.png --ar 16:9 [--ref brand-can.png]
node <skill>/scripts/kie.mjs shot  "<camera move>" out/01-hero.png out/01.mp4 --dur 5
bash  <skill>/scripts/encode.sh out/01.mp4 assets/01.mp4
bash  <skill>/scripts/encode.sh out/01.mp4 assets/01-m.mp4 mobile

Three things that decide whether this looks premium or generated:

• One style preamble, reused verbatim in every prompt. This is what makes

six separate images look like one shoot. Write it once, never paraphrase it.

• Look at every asset before you use it. Read the PNG. Generation is cheap

and rerolling is cheaper than shipping a bad frame.

• Encode for scrubbing, not playback. encode.sh sets a dense GOP because

seeking walks from the previous keyframe. A normal web encode plays perfectly

and scrubs like mud.

Step 4: Build the page

Write real HTML. Real <h1>, real <p>, real links, real reading order. The

engine reads data-sc-* attributes off your markup and drives it; it never

generates DOM. A runtime that builds the page from a config object is exactly

why every site built on one looks the same.

Start from references/template.html. The device patterns are in

references/devices.md; the spacing, type, depth and

colour rules are in references/taste.md. Read taste.md

before writing markup, not after, and build without announcing the checklist.

Theme by overriding tokens, six values and two fonts:

:root {
  --sc-canvas: #0A0806;  --sc-surface: #16110E;
  --sc-ink:    #F5EBDD;  --sc-ink-soft: #A2968A;
  --sc-accent: #FF5A3D;  --sc-accent-ink: #15110F;
  --sc-font-display: "Archivo", system-ui, sans-serif;
  --sc-font-text:    "Geist", system-ui, sans-serif;
}

Step 5: Verify by scrolling it

Not optional, and not "it should work." A scroll page has no single state:

every position is a different frame, and the failures live between the two you

happened to look at. Full procedure: references/verify.md.

cd <build project> && npm i playwright-core     # once
node <skill>/scripts/serve.mjs --root . --port 4500 &
node <skill>/scripts/shoot.mjs --url http://localhost:4500 --out lab/shots
node <skill>/scripts/shoot.mjs --url http://localhost:4500 --out lab/mobile --width 390 --height 844
node <skill>/scripts/shoot.mjs --url http://localhost:4500 --out lab/reduced --reduced-motion

The harness walks each act at six positions, waits for the scrub video to

actually settle, and reports dead scroll, **cues that never reach full

opacity, and contrast measured on the composited page at the brightest

frame under each line**. It writes a contact sheet.

Then do the part the harness cannot: read sheet.png. It proves a clip

advances; it cannot tell you the composition is good, the motion is smooth, or

the page means anything. Also tab through for focus order.

Then run the feel check (references/feel.md §6). Scroll

the page cold, write one word per act for what you felt, and only then open

BRIEF.md and diff it against the intended curve. Where they disagree the page is

wrong, not the brief. Confirm on the sheet that the peak is the largest visual

change and holds the most scroll room, and that the last screen resolves instead

of fading to nothing.

And say what a green run does not cover: a real phone. Headless Chrome

cannot reproduce an iPhone's video decoder, autoplay policy, Low Power Mode,

or touch scrolling; a build once shipped four green rounds while the hero clip

sat frozen on the actual device. Mobile is a first-class target throughout,

not a pass at the end: portrait phone clips, touch-tuned lerp, grown tap

targets are all authored (see assets.md and verify.md). When any mobile defect

is reported, deploy references/device-diag.html beside the site on the

first round and let the device answer, rather than theorising from a

machine that cannot reproduce the failure. The full iOS clip-lifecycle notes

live in verify.md, "The phone is a different machine".

Fix what you found and shoot it again. Report what you actually verified and

what you did not.

Hard rules

Ship-blockers, not preferences. Each one is a thing that makes a page read as

machine-made.

| Never | Instead |

|---|---|

| Clay diorama / low-poly / claymation as the default world | Photographic. See worlds.md |

| A "scroll" cue, arrow, or animated mouse icon | Nothing. They are looking at the hero; they know |

| 01 / 06 section counters | Delete them. Sequence is not information here |

| An eyebrow above every section heading | At most one per three sections. The heading carries itself |

| Em dash anywhere visible | Period, comma, colon, or parentheses |

| Centred copy in every act | Vary the anchor: lead, trail, centre, split |

| The same device twice in a row | Score the journey properly in Step 2 |

| Generating anything before the human has been interviewed | Run Step 0. Write BRIEF.md, or mark it self-authored |

| A page with no engineered peak, or with three competing ones | One peak. It gets the asset budget, the silence before it, and the most scroll room. See feel.md §2 |

| An ending that trails off, fades out, or just becomes a footer | The close resolves and holds. The last feeling is the one they carry |

| Planning acts before the feeling curve exists | Curve first, devices second. See feel.md §1 |

| Shipping without one bespoke signature move | Invent one. A recoloured spotlight or a retuned tilt is not one. See uniqueness.md §3 |

| A build that clears fewer than 4 of 6 fingerprint dimensions against any existing row | Change the plan, not FINGERPRINTS.md |

| Editing the engine to get a bespoke behaviour | Bespoke JS in the page, driven off --sc-p and your own data-sc-* |

| Reaching for filmic one-shot because it is what the last build did | Pick from all eight grammars, and say why the other seven lost |

| A full-frame dark overlay to fix contrast | A scrim only where the text sits |

| Text baked into a generated image | Real markup, always. It is selectable, translatable and sharp |

| Invented statistics in a counter | Only real numbers. No number, no counter |

| transition: all, or animating width/height/top/left | transform and opacity; clip-path for wipes |

| Gradient text, neon glow, zero-offset coloured halo shadows | Weight and size for emphasis; shadows with offset and blur |

| Autoplaying audio, or any audio at all on a scrub clip | Strip the track. encode.sh already does |

| Shipping without running Step 5 | Run Step 5 |

Output

The build folder, including BRIEF.md, then a short report: the grammar and why

the other seven lost, the signature move, the fingerprint gate result against

each existing row, the journey, the feeling curve and the peak, the feel-check

diff (intended curve against felt curve, and what you changed), the score table

(device per beat), what you

generated, what you verified with screenshots, and anything you could not

verify. Say if the brief was self-authored rather than interviewed. Give the

local URL. Keep it brief; the page is the deliverable.

Then append the build's row to <workspace>/FINGERPRINTS.md.

Changes to the skill itself, and the build findings that drove them, are logged

in CHANGELOG.md.