hyperframes-best-practices

name	hyperframes-best-practices
description	Best practices for programmatic video creation using HyperFrames, plain HTML compositions with GSAP animations rendered to MP4, with full Hebrew and RTL support. Covers composition authoring, data-* timing attributes, GSAP timeline contract, layout-before-animation methodology, visual identity gate, Hebrew fonts via Google Fonts (Heebo, Rubik, Assistant), RTL text rendering with dir="rtl", Hebrew TikTok/Reels-style captions via Whisper, audio-reactive visuals, scene transitions, and bidirectional Hebrew+English text. Use when building HTML-based video content or Hebrew social/marketing videos without React. Do NOT use for Remotion or general React video work, use remotion-best-practices for that.
license	Apache-2.0

HyperFrames Best Practices

Adapted from the upstream HyperFrames skill at heygen-com/hyperframes (Apache-2.0). Hebrew and RTL adaptations by skills-il.

HTML is the source of truth for video. A composition is an HTML file with data-* attributes for timing, a GSAP timeline for animation, and CSS for appearance. The framework handles clip visibility, media playback, and timeline sync.

Problem

Building HTML-based videos with Hebrew text requires the compiler to fetch Hebrew Google Fonts on demand, explicit dir="rtl" on Hebrew containers, mirrored GSAP entrance directions, and Hebrew caption sync via Whisper, none of which HyperFrames documents out of the box. Hebrew voiceover is a separate gap: the local Kokoro fallback does not support Hebrew (only 8 languages: en-us, en-gb, es, fr-fr, hi, it, pt-br, ja, zh), so Hebrew narration must come from a cloud TTS provider, either by pointing hyperframes tts at ElevenLabs via $ELEVENLABS_API_KEY (when your CLI build supports cloud providers), or by generating the file with any external service and importing it as an <audio> element.

Hebrew and RTL

For Hebrew and RTL compositions, load references/hebrew-rtl.md. It covers Hebrew font loading (the compiler auto-fetches from Google Fonts), dir="rtl" scoping, GSAP x-axis mirroring, Hebrew caption sync via hyperframes transcribe --language he, Hebrew voiceover via external TTS, and bidirectional text with <bdi>.

Approach

Before writing HTML, think at a high level:

1. What, what should the viewer experience? Identify the narrative arc, key moments, and emotional beats.
2. Structure, how many compositions, which are sub-compositions vs inline, what tracks carry what (video, audio, overlays, captions).
3. Timing, which clips drive the duration, where do transitions land, what's the pacing.
4. Layout, build the end-state first. See "Layout Before Animation" below.
5. Animate, then add motion using the rules below.

For small edits (fix a color, adjust timing, add one element), skip straight to the rules.

Visual Identity Gate

Before writing ANY composition HTML, you MUST have a visual identity defined. Do NOT write compositions with default or generic colors.

Check in this order:

1. DESIGN.md exists in the project? → Read it. Use its exact colors, fonts, motion rules, and "What NOT to Do" constraints.
2. visual-style.md exists? → Read it. Apply its style_prompt_full and structured fields. (Note: visual-style.md is a project-specific file. visual-styles.md is the style library with 8 named presets, different files.)
3. User named a style (e.g., "Swiss Pulse", "dark and techy", "luxury brand")? → Read visual-styles.md for the 8 named presets. Generate a minimal DESIGN.md with: ## Style Prompt (one paragraph), ## Colors (3-5 hex values with roles), ## Typography (1-2 font families), ## What NOT to Do (3-5 anti-patterns).
4. None of the above? → Ask 3 questions before writing any HTML:
   • What's the mood? (explosive / cinematic / fluid / technical / chaotic / warm)
   • Light or dark canvas?
   • Any specific brand colors, fonts, or visual references? Then generate a minimal DESIGN.md from the answers.

Every composition must trace its palette and typography back to a DESIGN.md, visual-style.md, or explicit user direction. If you're reaching for #333, #3b82f6, or Roboto, you skipped this step.

For motion defaults, sizing, entrance patterns, and easing, follow house-style.md. The house style handles HOW things move. The DESIGN.md handles WHAT things look like.

Layout Before Animation

Position every element where it should be at its most visible moment, the frame where it's fully entered, correctly placed, and not yet exiting. Write this as static HTML+CSS first. No GSAP yet.

Why this matters: If you position elements at their animated start state (offscreen, scaled to 0, opacity 0) and tween them to where you think they should land, you're guessing the final layout. Overlaps are invisible until the video renders. By building the end state first, you can see and fix layout problems before adding any motion.

The process

1. Identify the hero frame for each scene, the moment when the most elements are simultaneously visible. This is the layout you build.
2. Write static CSS for that frame. The .scene-content container MUST fill the full scene using width: 100%; height: 100%; padding: Npx; with display: flex; flex-direction: column; gap: Npx; box-sizing: border-box. Use padding to push content inward, NEVER position: absolute; top: Npx on a content container. Absolute-positioned content containers overflow when content is taller than the remaining space. Reserve position: absolute for decoratives only.
3. Add entrances with gsap.from(), animate FROM offscreen/invisible TO the CSS position. The CSS position is the ground truth; the tween describes the journey to get there.
4. Add exits with gsap.to(), animate TO offscreen/invisible FROM the CSS position.

Example

/* scene-content fills the scene, padding positions content */
.scene-content {
  display: flex;
  flex-direction: column;
  justify-content: center;
  width: 100%;
  height: 100%;
  padding: 120px 160px;
  gap: 24px;
  box-sizing: border-box;
}
.title {
  font-size: 120px;
}
.subtitle {
  font-size: 42px;
}
/* Container fills any scene size (1920x1080, 1080x1920, etc).
   Padding positions content. Flex + gap handles spacing. */

WRONG, hardcoded dimensions and absolute positioning:

.scene-content {
  position: absolute;
  top: 200px;
  left: 160px;
  width: 1920px;
  height: 1080px;
  display: flex; /* ... */
}

// Step 3: Animate INTO those positions
tl.from(".title", { y: 60, opacity: 0, duration: 0.6, ease: "power3.out" }, 0);
tl.from(".subtitle", { y: 40, opacity: 0, duration: 0.5, ease: "power3.out" }, 0.2);
tl.from(".logo", { scale: 0.8, opacity: 0, duration: 0.4, ease: "power2.out" }, 0.3);

// Step 4: Animate OUT from those positions
tl.to(".title", { y: -40, opacity: 0, duration: 0.4, ease: "power2.in" }, 3);
tl.to(".subtitle", { y: -30, opacity: 0, duration: 0.3, ease: "power2.in" }, 3.1);
tl.to(".logo", { scale: 0.9, opacity: 0, duration: 0.3, ease: "power2.in" }, 3.2);

When elements share space across time

If element A exits before element B enters in the same area, both should have correct CSS positions for their respective hero frames. The timeline ordering guarantees they never visually coexist, but if you skip the layout step, you won't catch the case where they accidentally overlap due to a timing error.

What counts as intentional overlap

Layered effects (glow behind text, shadow elements, background patterns) and z-stacked designs (card stacks, depth layers) are intentional. The layout step is about catching unintentional overlap, two headlines landing on top of each other, a stat covering a label, content bleeding off-frame.

Data Attributes

All Clips

Attribute	Required	Values
id	Yes	Unique identifier
data-start	Yes	Seconds or clip ID reference ("el-1", "intro + 2")
data-duration	Required for img/div/compositions	Seconds. Video/audio defaults to media duration.
data-track-index	Yes	Integer. Same-track clips cannot overlap.
data-media-start	No	Trim offset into source (seconds)
data-volume	No	0-1 (default 1)

data-track-index does not affect visual layering, use CSS z-index.

Composition Clips

Attribute	Required	Values
data-composition-id	Yes	Unique composition ID
data-start	Yes	Start time (root composition: use "0")
data-duration	Yes	Takes precedence over GSAP timeline duration
data-width / data-height	Yes	Pixel dimensions (1920x1080 or 1080x1920)
data-composition-src	No	Path to external HTML file

Composition Structure

Sub-compositions loaded via data-composition-src use a <template> wrapper. Standalone compositions (the main index.html) do NOT use <template>, they put the data-composition-id div directly in <body>. Using <template> on a standalone file hides all content from the browser and breaks rendering.

Sub-composition structure:

<template id="my-comp-template">
  <div data-composition-id="my-comp" data-width="1920" data-height="1080">
    <!-- content -->
    <style>
      [data-composition-id="my-comp"] {
        /* scoped styles */
      }
    </style>
    <script src="https://cdn.jsdelivr.net/npm/gsap@3.15.0/dist/gsap.min.js"></script>
    <script>
      window.__timelines = window.__timelines || {};
      const tl = gsap.timeline({ paused: true });
      // tweens...
      window.__timelines["my-comp"] = tl;
    </script>
  </div>
</template>

Load in root: <div id="el-1" data-composition-id="my-comp" data-composition-src="compositions/my-comp.html" data-start="0" data-duration="10" data-track-index="1"></div>

Video and Audio

Video must be muted playsinline. Audio is always a separate <audio> element:

<video
  id="el-v"
  data-start="0"
  data-duration="30"
  data-track-index="0"
  src="video.mp4"
  muted
  playsinline
></video>
<audio
  id="el-a"
  data-start="0"
  data-duration="30"
  data-track-index="2"
  src="video.mp4"
  data-volume="1"
></audio>

Timeline Contract

• All timelines start { paused: true }, the player controls playback
• Register every timeline: window.__timelines["<composition-id>"] = tl
• Framework auto-nests sub-timelines, do NOT manually add them
• Duration comes from data-duration, not from GSAP timeline length
• Never create empty tweens to set duration

Rules (Non-Negotiable)

Deterministic: No Math.random(), Date.now(), or time-based logic. Use a seeded PRNG if you need pseudo-random values (e.g. mulberry32).

GSAP: Only animate visual properties (opacity, x, y, scale, rotation, color, backgroundColor, borderRadius, transforms). Do NOT animate visibility, display, or call video.play()/audio.play().

Animation conflicts: Never animate the same property on the same element from multiple timelines simultaneously.

No repeat: -1: Infinite-repeat timelines break the capture engine. Calculate the exact repeat count from composition duration: repeat: Math.ceil(duration / cycleDuration) - 1.

Synchronous timeline construction: Never build timelines inside async/await, setTimeout, or Promises. The capture engine reads window.__timelines synchronously after page load. Fonts are embedded by the compiler, so they're available immediately, no need to wait for font loading.

Never do:

1. Forget window.__timelines registration
2. Use video for audio, always muted video + separate <audio>
3. Nest video inside a timed div, use a non-timed wrapper
4. Use data-layer (use data-track-index) or data-end (use data-duration)
5. Animate video element dimensions, animate a wrapper div
6. Call play/pause/seek on media, framework owns playback
7. Create a top-level container without data-composition-id
8. Use repeat: -1 on any timeline or tween, always finite repeats
9. Build timelines asynchronously (inside async, setTimeout, Promise)
10. Use gsap.set() on clip elements from later scenes, they don't exist in the DOM at page load. Use tl.set(selector, vars, timePosition) inside the timeline at or after the clip's data-start time instead.
11. Use <br> in content text, forced line breaks don't account for actual rendered font width. Text that wraps naturally + a <br> produces an extra unwanted break, causing overlap. Let text wrap via max-width instead. Exception: short display titles where each word is deliberately on its own line (e.g., "THE\nIMMORTAL\nGAME" at 130px).

Scene Transitions (Non-Negotiable)

Every multi-scene composition MUST follow ALL of these rules. Violating any one of them is a broken composition.

1. ALWAYS use transitions between scenes. No jump cuts. No exceptions.
2. ALWAYS use entrance animations on every scene. Every element animates IN via gsap.from(). No element may appear fully-formed. If a scene has 5 elements, it needs 5 entrance tweens.
3. NEVER use exit animations except on the final scene. This means: NO gsap.to() that animates opacity to 0, y offscreen, scale to 0, or any other "out" animation before a transition fires. The transition IS the exit. The outgoing scene's content MUST be fully visible at the moment the transition starts.
4. Final scene only: The last scene may fade elements out (e.g., fade to black). This is the ONLY scene where gsap.to(..., { opacity: 0 }) is allowed.

WRONG, exit animation before transition:

// BANNED, this empties the scene before the transition can use it
tl.to("#s1-title", { opacity: 0, y: -40, duration: 0.4 }, 6.5);
tl.to("#s1-subtitle", { opacity: 0, duration: 0.3 }, 6.7);
// transition fires on empty frame

RIGHT, entrance only, transition handles exit:

// Scene 1 entrance animations
tl.from("#s1-title", { y: 50, opacity: 0, duration: 0.7, ease: "power3.out" }, 0.3);
tl.from("#s1-subtitle", { y: 30, opacity: 0, duration: 0.5, ease: "power2.out" }, 0.6);
// NO exit tweens, transition at 7.2s handles the scene change
// Scene 2 entrance animations
tl.from("#s2-heading", { x: -40, opacity: 0, duration: 0.6, ease: "expo.out" }, 8.0);

Animation Guardrails

• Offset first animation 0.1-0.3s (not t=0)
• Vary eases across entrance tweens, use at least 3 different eases per scene
• Don't repeat an entrance pattern within a scene
• Avoid full-screen linear gradients on dark backgrounds (H.264 banding, use radial or solid + localized glow)
• 60px+ headlines, 20px+ body, 16px+ data labels for rendered video
• font-variant-numeric: tabular-nums on number columns

When no visual-style.md or animation direction is provided, follow house-style.md for aesthetic defaults.

Typography and Assets

• Fonts: Just write the font-family you want in CSS, the compiler embeds supported fonts automatically. If a font isn't supported, the compiler warns.
• Add crossorigin="anonymous" to external media
• For dynamic text overflow, use window.__hyperframes.fitTextFontSize(text, { maxWidth, fontFamily, fontWeight })
• All files live at the project root alongside index.html; sub-compositions use ../

Editing Existing Compositions

• Read the full composition first, match existing fonts, colors, animation patterns
• Only change what was requested
• Preserve timing of unrelated clips

Output Checklist

• npx hyperframes lint and npx hyperframes validate both pass
• Contrast warnings addressed (see Quality Checks below)
• Animation choreography verified (see Quality Checks below)
• Rendered to MP4 with npx hyperframes render (lint and validate run on the HTML; render is the step that actually encodes the video, see /hyperframes-cli upstream for flags)

Quality Checks

Contrast

hyperframes validate runs a WCAG contrast audit by default. It seeks to 5 timestamps, screenshots the page, samples background pixels behind every text element, and computes contrast ratios. Failures appear as warnings:

⚠ WCAG AA contrast warnings (3):
  · .subtitle "secondary text", 2.67:1 (need 4.5:1, t=5.3s)

If warnings appear:

• On dark backgrounds: brighten the failing color until it clears 4.5:1 (normal text) or 3:1 (large text, 24px+ or 19px+ bold)
• On light backgrounds: darken it
• Stay within the palette family, don't invent a new color, adjust the existing one
• Re-run hyperframes validate until clean

Use --no-contrast to skip if iterating rapidly and you'll check later.

Animation Map

After authoring animations, run the animation map to verify choreography:

node skills/hyperframes/scripts/animation-map.mjs <composition-dir> \
  --out <composition-dir>/.hyperframes/anim-map

Outputs a single animation-map.json with:

• Per-tween summaries: "#card1 animates opacity+y over 0.50s. moves 23px up. fades in. ends at (120, 200)"
• ASCII timeline: Gantt chart of all tweens across the composition duration
• Stagger detection: reports actual intervals ("3 elements stagger at 120ms")
• Dead zones: periods over 1s with no animation, intentional hold or missing entrance?
• Element lifecycles: first/last animation time, final visibility
• Scene snapshots: visible element state at 5 key timestamps
• Flags: offscreen, collision, invisible, paced-fast (under 0.2s), paced-slow (over 2s)

Read the JSON. Scan summaries for anything unexpected. Check every flag, fix or justify. Verify the timeline shows the intended choreography rhythm. Re-run after fixes.

Skip on small edits (fixing a color, adjusting one duration). Run on new compositions and significant animation changes.

---

References (loaded on demand)

• references/captions.md, Captions, subtitles, lyrics, karaoke synced to audio. Tone-adaptive style detection, per-word styling, text overflow prevention, caption exit guarantees, word grouping. Read when adding any text synced to audio timing.

• references/tts.md, Text-to-speech with Kokoro-82M. Voice selection, speed tuning, TTS+captions workflow. Read when generating narration or voiceover.

• references/audio-reactive.md, Audio-reactive animation: map frequency bands and amplitude to GSAP properties. Read when visuals should respond to music, voice, or sound.

• references/css-patterns.md, CSS+GSAP marker highlighting: highlight, circle, burst, scribble, sketchout. Deterministic, fully seekable. Read when adding visual emphasis to text.

• references/typography.md, Typography: font pairing, OpenType features, dark-background adjustments, font discovery script. Always read, every composition has text.

• references/motion-principles.md, Motion design principles: easing as emotion, timing as weight, choreography as hierarchy, scene pacing, ambient motion, anti-patterns. Read when choreographing GSAP animations.

• visual-styles.md, 8 named visual styles (Swiss Pulse, Velvet Standard, Deconstructed, Maximalist Type, Data Drift, Soft Signal, Folk Frequency, Shadow Cut) with hex palettes, GSAP easing signatures, and shader pairings. Read when user names a style or when generating DESIGN.md.

• house-style.md, Default motion, sizing, and color palettes when no style is specified.

• patterns.md, PiP, title cards, slide show patterns.

• data-in-motion.md, Data, stats, and infographic patterns.

• references/transcript-guide.md, Transcription commands, whisper models, external APIs, troubleshooting.

• references/dynamic-techniques.md, Dynamic caption animation techniques (karaoke, clip-path, slam, scatter, elastic, 3D).

• references/hebrew-rtl.md, Hebrew and RTL compositions: dir="rtl" scoping, Google Fonts auto-fetch for Heebo/Rubik/Assistant, GSAP x-axis mirroring, Hebrew captions via hyperframes transcribe --language he, Hebrew voiceover via external TTS, bidirectional text with <bdi>. Read for any composition with Hebrew text.

• references/transitions.md, Scene transitions: crossfades, wipes, reveals, shader transitions. Energy/mood selection, CSS vs WebGL guidance. Always read for multi-scene compositions, scenes without transitions feel like jump cuts.

  • transitions/catalog.md, Hard rules, scene template, and routing to per-type implementation code.
  • Shader transitions are in @hyperframes/shader-transitions (packages/shader-transitions/), read package source, not skill files.

For GSAP timeline patterns and easing, follow house-style.md and references/motion-principles.md in this skill, plus the official GSAP docs at https://gsap.com/docs/v3/.

Gotchas

These are agent failure modes specific to Hebrew/RTL HyperFrames work. Generic HyperFrames gotchas (see upstream) still apply.

• Don't add a Google Fonts <link rel="stylesheet"> tag or a CSS @import url(...) statement for Hebrew fonts. The compiler already fetches Google Fonts server-side via fetchGoogleFont() in packages/producer/src/services/deterministicFonts.ts, caches the WOFF2s at ~/.cache/hyperframes/fonts/<slug>/, and embeds them as base64 data URIs in the compiled HTML. An external stylesheet breaks determinism (network dependency at render time) and duplicates the font loading. Just write font-family: 'Heebo', sans-serif;.
• Don't reach for the built-in hyperframes tts command for Hebrew narration. The bundled Kokoro-82M supports 8 languages via voice-ID prefix, a=American English, b=British English, e=Spanish, f=French, h=Hindi, i=Italian, j=Japanese, p=Brazilian Portuguese, z=Mandarin. Hebrew is not included, so the local fallback can't voice Hebrew. Two paths: (a) set $ELEVENLABS_API_KEY so hyperframes tts routes to ElevenLabs (Hebrew voices) instead of Kokoro, if your installed CLI build supports cloud providers; or (b) generate the WAV/MP3 with any external service (ElevenLabs, OpenAI TTS, Google Cloud TTS Hebrew) and drop the file into the composition as a normal <audio> clip. The published hyperframes tts may synthesize locally with Kokoro only, so verify against your installed version.
• Don't use .en Whisper models on Hebrew audio. .en variants TRANSLATE non-English audio to English instead of transcribing it. For Hebrew captions use npx hyperframes transcribe audio.wav --model medium --language he (small is weak for Hebrew ASR; step up to large-v3 --language he for noisy audio). The .en suffix is only correct when the user explicitly says the audio is English.
• Don't forget dir="rtl" on Hebrew text containers, even inside a RTL-defaulted composition. HyperFrames sub-compositions set their own direction context. GSAP x: tweens also don't auto-mirror. A title that uses gsap.from({x: -80}) enters from the left in both LTR and RTL, for Hebrew, flip to x: 80 so it enters from the right, matching reading direction.
• Don't paste English brand names into Hebrew paragraphs without <bdi> or unicode-bidi: isolate. Without isolation, the Unicode bidi algorithm reorders mixed-direction runs and can place punctuation on the wrong side of the brand name or visually reverse it. Wrap brand names: הצטרפו ל־<bdi>HyperFrames</bdi> עכשיו.

Hebrew Bidi Details

The <bdi> rule above covers brand names. Mixed-direction Hebrew compositions need three more bidi habits:

• Hebrew line-breaking for long headlines. Hebrew does not hyphenate. A long Hebrew headline that overflows must wrap at word boundaries, never mid-word. Set max-width so it wraps naturally, and add word-break: keep-all (or overflow-wrap: normal) so the compiler does not break inside a Hebrew word. Do NOT use <br> to force breaks (see Rule 11). For deliberate one-word-per-line display titles, give each word its own element instead.
• Digits with adjacent symbols inside RTL runs. A bare digit run is handled correctly by the Unicode bidi algorithm: 2025 next to a Hebrew word stays left-to-right inside an RTL paragraph on its own, it does NOT reverse to 5202, so a plain integer needs no wrapper. The real hazards are a digit touching a symbol, a range, or a Latin token: a percent sign, currency sign, or range dash can detach and land on the wrong side. Wrap those with <bdi> or ⁦...⁩ (LTR isolate): <bdi>15%</bdi> הנחה, <bdi>₪199</bdi>, <bdi>10-20</bdi>. This keeps the symbol attached to its number. See references/hebrew-rtl.md.
• Punctuation mirroring. Parentheses, brackets, and quotes are mirrored characters: ( visually becomes ) in an RTL run. In Hebrew captions and paragraphs let the browser mirror them by keeping the text in a proper dir="rtl" container. Do NOT hand-swap ( and ) to "fix" it, and when a parenthetical contains LTR content (a brand, a URL, a number) wrap that inner content in <bdi> so only the inner run is LTR while the parentheses stay correctly mirrored.

Troubleshooting

hyperframes command not found / render fails immediately

HyperFrames requires Node 22+ and FFmpeg on PATH. Confirm node --version is 22 or higher and ffmpeg -version resolves. On macOS install FFmpeg with brew install ffmpeg; on Debian/Ubuntu use apt install ffmpeg. Without FFmpeg the compiler cannot encode the MP4 and aborts before rendering any frames.

Compiler warns "font not supported" or Hebrew renders in a fallback font

The compiler only embeds fonts it can fetch from Google Fonts. Use a Hebrew family that exists on Google Fonts (Heebo, Rubik, Assistant, Alef, Frank Ruhl Libre, Noto Sans Hebrew) and write it plainly in CSS: font-family: 'Heebo', sans-serif;. Do NOT add a <link rel="stylesheet"> or @import (see Gotchas), that breaks determinism without fixing the warning. If a custom non-Google font is required, the upstream docs cover local font embedding.

WCAG contrast audit fails (hyperframes validate)

validate samples background pixels behind each text element at 5 timestamps and flags ratios under 4.5:1 (normal text) or 3:1 (large text). Fix by adjusting the failing color WITHIN the palette family: brighten it on dark backgrounds, darken it on light backgrounds. Do not invent a new color. Re-run hyperframes validate until clean. Use --no-contrast only while iterating, never as the final state.

Composition renders blank or content is invisible

The most common cause is a <template> wrapper on a standalone composition. The main index.html must put the data-composition-id div directly in <body>, not inside <template>. Also check that every timeline is registered via window.__timelines["<composition-id>"] = tl and built synchronously (not inside async, setTimeout, or a Promise), the capture engine reads window.__timelines synchronously after page load.

Hebrew title enters from the wrong side

GSAP x: tweens do not auto-mirror for RTL. A gsap.from({x: -80}) enters from the left in both LTR and RTL. For Hebrew, flip to a positive value (x: 80) so the element enters from the right, matching reading direction. See references/hebrew-rtl.md.

Reference Links

Source	URL	What to Check
HyperFrames GitHub	https://github.com/heygen-com/hyperframes	Upstream repo, issues, releases
HyperFrames docs	https://hyperframes.heygen.com/quickstart	CLI, Node 22+, FFmpeg requirement
Compiler font logic	https://github.com/heygen-com/hyperframes/blob/main/packages/producer/src/services/deterministicFonts.ts	Canonical font list, Google Fonts fallback, cache path
Kokoro TTS voices	https://github.com/heygen-com/hyperframes/blob/main/skills/hyperframes-media/references/tts.md	Kokoro voices across 8 languages (no Hebrew)
Whisper model guide	https://github.com/skills-il/developer-tools/blob/master/hyperframes-best-practices/references/transcript-guide.md	.en vs multilingual models, --language flag
Google Fonts Hebrew	https://fonts.google.com/?subset=hebrew	Heebo, Rubik, Assistant, Alef, Frank Ruhl Libre, Noto Sans Hebrew
Unicode bidi spec	https://developer.mozilla.org/en-US/docs/Web/CSS/unicode-bidi	isolate, <bdi>, mixed-direction text