Herk Edit Kit by Nate Herk: edit videos, reels, and YouTube Shorts with Codex or Claude Code. 15 skills, transcript-driven cuts, storytelling, 406 motion-graphics cards, and HyperFrames templates.
Style Library — Guide
The library of motion-graphic card styles that Agent 3 pulls from to add tiered cards to a video. Each "style" is a named visual identity (e.g. a Vox-inspired explainer look, or a smooth-animated YouTuber look) with its own set of card variants. This file is the contract: read it before adding a style or a card.
Mental model
style-library/
├── GUIDE.md ← this file (the rules)
├── registry.json ← GENERATED master index of every style + card (Agent 3 reads this)
├── _blueprint/ ← copy-me skeleton for a new style (never edit in place; clone it)
└── NN-slug/ ← one folder per style, e.g. 01-vox-explainer/
One style = one folder. We build one style fully before starting the next.
Identifiers
Three-part scheme (numeric prefix + slug + dotted IDs):
- Style folder:
NN-slug— e.g.02-vox-explainer.NNis a zero-padded order/ID,slugis human-readable kebab-case. - Style id: the slug without the number — e.g.
vox-explainer(stable even if the number changes). - Card id:
<style>.<tier>.<purpose>[.<variant>]— e.g.vox-explainer.t1.stat,vox-explainer.t2.list.compact. Stable, unique, self-describing.
A style folder
NN-slug/
├── style.json ← manifest: identity + palette + fonts + motion + the list of cards (with slots, tiers, purposes)
├── DESIGN.md ← the design spec: palette, typography, motion rules, "what not to do" (the shared design format)
├── tokens.css ← style-scoped CSS variables (colors, fonts, easings). Cards reference these, never hardcode.
├── references/ ← the reference images you upload for this style (the source of truth for the look)
├── cards/
│ ├── tier1/ ← full-screen takeovers (thesis, big stat, section marker, quote)
│ ├── tier2/ ← supporting cards that keep the speaker visible (label, list, equation, lower-third)
│ └── <custom>/ ← optional "innovations" — any card kind that doesn't fit tier1/tier2
└── preview/ ← per-card preview: <card-id>.mp4 (2-3s loop) + <card-id>.png (poster)
Committed: style.json, DESIGN.md, tokens.css, the card .html files, the design specifications, and any original preview posters. The student release excludes third-party reference screenshots. Gitignored: the heavier preview MP4 loops (regenerable from the cards).
The card contract
Every card is a standalone HyperFrames sub-composition — a self-contained .html that previews and renders on its own, and that Agent 3 mounts into a video via data-composition-src. Each card MUST:
- Have a root
<div>withid,data-composition-id,data-start="0",data-width="1920",data-height="1080". - Register exactly one paused GSAP timeline on
window.__timelines["<data-composition-id>"](key === composition id). - Pull all colors / fonts / easings from
tokens.cssvariables — never hardcode hex or font names. (This is the "bespoke layout, shared tokens" rule: layouts are hand-built per style; the palette/type lives in tokens so it stays consistent and tweakable.) - Background discipline by tier:
tier1(takeover) → opaque full-frame background (it replaces the video).tier2(support) → transparent everywhere except the card itself (it overlays the video; never cover the speaker's face).
- Expose its text as named slots: elements carrying
data-slot="<name>"(e.g.data-slot="headline"). Agent 3 fills slot text per beat. Declare every slot instyle.json(name, type, max chars) so the agent knows what to fill. - Be deterministic — no
Date.now(), no unseededMath.random(), no network fetches.
A card is "done" when it lints clean, previews correctly in the Studio, and has a rendered preview MP4 + poster.
Standalone vs mounted form
Cards are authored as standalone compositions — the data-composition-id div sits directly in <body> and the card includes its own <script src=".../gsap.min.js">, so it previews/renders on its own. When Agent 3 mounts a card into a real video it wraps the card in a <template> and references it with data-composition-src (the HyperFrames sub-composition form); the parent provides GSAP and data-duration. The wrap is a mechanical transform Agent 3 performs at assembly time — library cards stay in standalone form.
Tiers & purposes (the taxonomy Agent 3 queries by)
- tier1 — takeover purposes:
thesis,stat,section,quote,overview - tier2 — support purposes:
label,list,equation,definition,lower-third - custom — anything else; give it a clear purpose tag.
Agent 3 asks the registry for "a tier1 stat card in style vox-explainer", fills the slots, times it to the transcript.
The build loop (per style)
- Scaffold:
node scripts/style-library/new-style.mjs <NN> <slug>→ clones_blueprint/toNN-slug/. - Upload references into
NN-slug/references/. - Define the look: fill
DESIGN.md+tokens.css+ the identity fields ofstyle.jsonfrom the references. - Build cards: author each card
.htmlundercards/, add its entry (id, tier, purpose, slots, file) tostyle.json. Lint each. - Preview: render a 2-3s loop + poster per card into
preview/(see below). - Index:
node scripts/style-library/build-registry.mjs→ regeneratesregistry.json. - Review & iterate against the references, then move to the next style.
Previews
Motion is half the identity, so previews are short MP4 loops (2-3s) + a poster PNG, one per card, in the style's preview/. Because each card is a valid composition, a preview is produced by rendering the card on its own (via the HyperFrames CLI) and grabbing frame 1 as the poster. The preview script is added once the first real card exists (so it's tested against something real).
Notes
- Nate's
style-templates/(dark-graph-paper, left-glass-popout) are reference only — not part of this library. Borrow ideas, don't port wholesale. MOTION_PHILOSOPHY.mdand.claude/skills/hyperframes/(palettes, transitions, typography, visual-styles) are shared craft resources every style can draw on.