scroll-craft

A skill a coding agent loads before it designs a page, assembled by one person and carrying a whole written standard with it: eight mutually exclusive page grammars, a required signature move, a feeling curve that has to exist before any act does, one engineered peak, floors for typography, spacing, colour and depth, and a fingerprint registry that rejects a new build unless it differs from the pages already made on at least four of six dimensions. Beside the rules sit a scroll engine that is never edited per project and seven scripts — a doctor, an encoder, a local server, a workspace manager and a headless verification pass among them — that walk the finished page at every scroll position. It ships as a Claude Code plugin, and as a folder any agent with file access can read.

Who built itNate Herk, on the GitHub account nateherkai. The material holds no biography — no employer, location or other project — so the record is the repository itself: 16 commits in forty days, every one of them his, and a contributor list with one name on it. Twelve of the sixteen commits carry a co-author trailer naming a Claude model rather than a person — Opus 5 on seven, Opus 4.8 on two, and Opus 5.5, Fable 5 and Fable 5.1 once each. The repository has 2,877 stars, 418 forks and 23 watchers, and the README’s showcase of three builds includes a portfolio page under his own name.

1. 01

   Sixteen commits in forty days, and a version line kept in a changelog

   The repository was created on 2026-08-22T15:25:08Z and its first commit landed nine minutes later, titled Initial commit: scrollcraft as a Claude Code plugin. The last one, dated 2026-09-30T13:31:41Z, is Security and Windows fixes, engine teardown (0.3.1). That is the entire history: sixteen commits in forty days, eight in August and eight in September, all authored by the same person under the same linked account, with him as the only contributor listed. Twelve of the sixteen carry a co-author trailer naming a model rather than a person — Claude Opus 5 on seven, Opus 4.8 on two, and one each for Opus 5.5, Fable 5 and Fable 5.1. Around it sit 2,877 stars, 418 forks, 23 watchers and five open reports, on a 10,220 KB JavaScript repository under MIT. What is absent is as telling: no GitHub releases at all, and one tag, nateherk-design--v0.2.0. The version line lives elsewhere. The README announces 0.3.0 as a milestone but points anyone wanting the history at CHANGELOG.md — 35,142 bytes, larger than the skill file itself — describing it as recording “what broke on each build and the rule that came out of it, rather than a feature list”.

2. 02

   The standard was read off ten sites that had already shipped

   Version 0.3.0 is called “the approved ten-site standard”, and it is the process behind ten sites built first: AI Automation Society, PERKFORM, Glaido, Herkules Advisory, Serein, FORME, Pelagic, NOEMA, OFFGRID and Afterhours. Three of them are shown with screenshots kept in the repository — a dark editorial community landing, a high-key portfolio for the author himself, and PERKFORM, a protein-coffee page built as a filmic one-shot — and the claim attached to them is that these are not themes but “different page grammars, different navigation models, different endings”, with the same skill and the same engine behind all three. What the pass added is a checklist rather than a look: independent depth planes with contact anchors and separate opening, midpoint and exit states; authentic brand assets and verified product details used before any imagery is generated; a deliberate choice between photographic compositing and real 3D; navigation, information order, useful controls and an ending authored per site; and phones art-directed separately with the actual scroll frames, fallbacks and packages verified. It is carried by references/approved-collection.md (11,650 bytes) and references/hero-depth.md (6,260). Two limits are stated: client assets and private form data are not bundled, and the ten-site rebuild was done with Codex rather than the Claude Code plugin.

3. 03

   Numbers instead of adjectives, and a gate aimed at your own last build

   The standard is written as measurable constraints rather than taste. Typography: two families maximum, tracking that tightens as size grows, a 45 to 75 character measure, line height inverse to measure, light-on-dark compensated on three axes. Spacing: a 4px base, more space above a heading than below it, fluid section padding. Colour: six roles and one accent, secondary text tinted rather than flat grey, no pure black, with a documented escape for hard light-and-dark cuts. Depth: five separate tools — offset shadows, edge light, scale-and-blur as distance, overlap and grain. Structure: eight deliberately incompatible grammars — filmic one-shot, chaptered editorial, live surface, continuous world, typographic poster, gallery, split stage and rhythmic cutlist — of which the README says each one “forbids what the others require, so two builds cannot quietly converge”. Every build also owes one bespoke signature move, and “a recoloured spotlight does not count”. Enforcement is a fingerprint gate: a new build must differ from every page already made on at least four of six dimensions — grammar, nav, hero, act shape, close, signature move — and “fail it and you change the plan, not the record”. A refuse list carries the negative baseline: identical feature-card grids, 01 / 06 counters, scroll cues, gradient text, invented statistics, fake dashboards, AI-purple gradients and the cream-and-brass artisan palette.

4. 04

   A frozen engine, a headless walk, and a contact sheet

   Under the prose sits one hard rule: the engine is the mechanism and is never edited per project. Themed with six colour tokens and two fonts, with your own semantic HTML written against it, and anything bespoke driven off the --sc-p custom property the engine publishes — because, the README argues, a runtime that builds the page from a config object is “exactly why every site built on one looks the same”. The engine is two files, scrollcraft.js at 62,176 bytes and scrollcraft.css at 19,769. The checking half is scripts/, seven files and 71,783 bytes, of which shoot.mjs at 33,472 bytes is the largest script. A headless browser walks the finished page at every scroll position and waits for the video playhead to settle before reporting four named faults: dead scroll, cues that never reach full opacity, contrast measured on the composited page line by line at the brightest frame that ever passes under it with the direction picked per line, and “legs stuck on a poster” — a clip that silently never decoded. It then writes a contact sheet, because “a machine can prove a page works and cannot tell you it means anything”. doctor.mjs (7,281 bytes) exists because three setup faults otherwise surface as misleading errors later: a stripped ffmpeg reporting a missing filter as a syntax error in your own command, a missing WebP muxer as a bad filename, and playwright-core resolving from the wrong directory.

5. 05

   Nine reports from strangers, and the three that became 0.3.1

   Nine issues and pull requests sit on the repository, filed by nine different accounts, and not one of those names appears in the contributor list, which holds a single entry. The earliest is a documentation-versus-runtime drift: the engine reads data-sc-kinetic only from the element that also carries data-sc-cue (engine/scrollcraft.js:373 and :869), while references/template.html:58-59 and the scrub example in references/devices.md:30-31 put the attribute on a child, so the mode was silently ignored. Four others are packaging and platform work: installation through the skills CLI, support for agents beyond Codex and Claude Code, a skill description cut from 1,175 to 414 characters to fit a loader limit, and Chromium crashing when shoot.mjs runs as root on Linux. The final commit, Security and Windows fixes, engine teardown (0.3.1), names the three that remained. The teardown report came from someone vendoring the engine into a Phoenix LiveView app: mount() registers five gesture listeners on window plus pointermove, a scroll read loop and focusin, all anonymous and none of them removed, so a host that swaps the DOM without a page load accumulates listeners and two animation-frame loops; a follow-up added that documentElement keeps --sc-seg, --sc-segp and --sc-canvas after teardown. A Windows report found the Edge fallback looking in Program Files\Microsoft\Edge, where the stable channel never lands on x64, in three scripts at once. The last one, against scripts/serve.mjs, listed three faults: the server binding every interface, the path check missing its separator, and a malformed URL killing the process inside decodeURIComponent.