noti-tiktok-full-text
Build a Vietnamese vertical TikTok explainer in the "MỔ XẺ PAPER AI" (AI paper dissection) format with HyperFrames (HTML/CSS/GSAP → MP4), Noti.vn style. N text scenes + PDF paper-page scenes, every pi
SKILL.md.claude/skills/noti-tiktok-full-text/SKILL.md
GitHubTikTok Full-Text - the "MỔ XẺ PAPER AI" explainer (Noti.vn)
⚖️ STYLE DESIGN - PRIORITY RULE (read this first)
If the edit prompt contains a "STYLE DESIGN (BẮT BUỘC TUÂN THỦ 100%)" section: the palette, fonts and tone in that section COMPLETELY REPLACE every color/font/branding rule in this skill (including "dark fintech blue", the GĐT hex values, the default gradient...). Keep only: animation technique, layout, cutting rhythm, render workflow, the bug fixes. Illustrations generated through /api/illustrations must pass the correct styleId in the prompt. NO Style Design section → use the skill's default branding.
If the edit prompt ALSO contains a "PHONG CÁCH DỰNG" (video style) section: that section WINS over this skill for ALL visual and motion language - materials, transitions, camera moves, effects. This skill then keeps only PROCESS: step order, cutting, captions/keys, draft→final, QC. Do not mix the skill's default visuals into the video style - half-and-half is exactly the failure this rule exists to prevent.
Build a Vietnamese vertical explainer with HyperFrames (HTML/CSS/JS + GSAP → MP4). All text is a real HTML element so it can be animated and stays razor sharp. There are text scenes and PDF paper-page scenes.
⚡ Read this whole file before building. Every ⚠️ section is a bug already hit & already fixed - do NOT repeat it. Project structure: scaffold with
npx hyperframes init, then follow the layout in thevideo-pipelineskill (index.html+compositions/+assets/+renders/+hyperframes.json+meta.json). (video-projects/sapo/is an example project on the author's machine, NOT shipped with the repo -video-projects/is gitignored.)
ROLE
HyperFrames video editor. Compose HTML/CSS/JS, animate with GSAP, render a vertical MP4. Genre: "MỔ XẺ PAPER AI" - a vertical explainer, Noti.vn style.
OUTPUT SPEC
- 9:16 - 1080×1920, 30fps.
- N scenes (12 by default). Each scene = its own HTML sub-composition (
<template>+data-composition-src), declaringdata-start/data-duration/data-track-index. - Safe zone: important text + captions inside the central 1080×1400 area, leaving 15% at the bottom (~288px) + the right corner (~130px). Put captions at
bottom: ~372px.
⚠️ VIETNAMESE FONT - NO MISTAKES ALLOWED (tofu / missing diacritics)
- Font "Be Vietnam Pro" or "Inter" vietnamese subset,
display=block. Embed via Google Fonts - the HyperFrames renderer fetches + caches it and injects a deterministic @font-face on its own:<link href="https://fonts.googleapis.com/css2?family=Be+Vietnam+Pro:wght@400;500;600;700;800;900&display=block&subset=vietnamese" rel="stylesheet"> <meta charset="UTF-8">,<html lang="vi">,font-display: block.- End the fallback chain with a font that has diacritics:
font-family: "Be Vietnam Pro","Inter",sans-serif;- do NOT end on a baresans-serif. - This test string must render 100% correctly:
Vấn đề: càng siết càng tệ - ƯỢ Ễ Ỹ ặ ậ ề thiết kế ngữ cảnh. If you see boxes or missing marks, fix the font before anything else.
⚠️⚠️ MANDATORY FIX - GRADIENT TEXT GETTING ITS DIACRITICS CLIPPED
Text using background-clip:text; color:transparent (blue gradient headings) CLIPS TALL STACKED DIACRITICS (Ẫ Ể Ấ Ữ Ổ Ợ…) because the marks stick out above the fill box. padding-top: 0.14em IS NOT ENOUGH - it must be 0.5em. Apply it to EVERY gradient text line:
.heading-grad {
background: linear-gradient(100deg,#0061ff,#00c2ff);
-webkit-background-clip: text; background-clip: text; color: transparent;
line-height: 1.0; padding-top: 0.5em; padding-bottom: 0.14em;
}
Plain white/solid-color text (no background-clip) does NOT have this bug - only gradient text needs fixing. When verifying: zoom 3× into the glyphs with stacked marks to check.
⚠️ Captions / multiple lines at the same spot - must NOT OVERLAP
Captions (and any text placed at the same spot) when the video is spoken continuously with no pauses: the previous line has not finished fading out when the next one fades in → they stack, the text turns to mush. Fix: PRE_ROLL = 0, the previous line goes fully off exactly when the next appears:
const inAts = SEGMENTS.map(s => Math.max(s.words[0].start, 0));
// fadeInAt = inAts[i];
// fadeOutAt = (i<n-1) ? Math.max(segEnd+0.02, inAts[i+1] - FADE_OUT) : segEnd + POST_HOLD;
Pull an extra frame right at the BOUNDARY between two caption lines.
⚠️ TEXT MUST BE A REAL HTML ELEMENT
- ALL Vietnamese text (headings, figures, captions, labels, quotes) is an
<h1>/<p>/<span>- do NOT bake text into images, no PNGs with text already in them. - Images are only used as background: the original PDF paper page (with no Vietnamese text pre-overlaid).
- Why: you need to animate it (reveal, count-up, per-word color change) and the glyph edges stay sharp when rendered.
⚠️ Sub-compositions render IN ISOLATION
Every sub-comp (data-composition-src) renders in its own context, inheriting NOTHING from the master:
- Every sub-comp has its own
<script src=".../gsap@3.14.2/dist/gsap.min.js">. - Every sub-comp with text embeds its own Google font
<link>(inside the<template>). - Do NOT use the master's
:rootCSS vars inside a sub-comp - use literal values (hex/gradient written out).
DESIGN SYSTEM (Noti.vn / GĐT)
Backgrounds:
- Text scenes: dark
#0A0E1A → #0F1629(radial/linear gradient), 1–2 blue blobs (#0061ff,#00c2ff) at blur 120px drifting slowly in parallax, grain noise at opacity ~0.04, vignette. - Paper scenes: the PDF page on a light background, 16px radius + soft shadow; the edge drawn by box-shadow (NOT a hard border).
Color:
- Primary text white
#F5F7FA. Secondary/already-read text grey#6E7787. - Brand accent (heading keywords): gradient
#0061ff → #00c2ffviabackground-clip:text(remember the 0.5em padding fix). - ACCENT: orange-red
#FF4D2Efor bad numbers, hot keywords in captions, arrows. - HIGHLIGHT: light red background
rgba(255,77,46,0.18)with a 6px radius (replacing the yellow highlighter), animated as a left→right sweep over 0.6s withclip-path. - Glass card:
rgba(255,255,255,0.06),backdrop-filter: blur(20px), 16–20px radius, edge drawn by box-shadow:inset 0 1px 0 rgba(255,255,255,0.18), inset 0 0 0 1px rgba(255,255,255,0.05), 0 18px 50px rgba(0,0,0,0.45).
Type: headings 800–900 (keywords in blue gradient/white); figures 900, 140–200px (bad numbers in orange-red); captions 700 + a light black text-shadow.
COMPONENTS (reusable)
- Page counter top left "01 / 12", grey, wide letter-spacing.
- Never add your own corner logo/watermark - the Remotion assembly layer stamps the Style Design logo automatically when the style has one (adding your own means two logos stacked). No "@noti.vn" text, no TikTok text; leave the bottom right corner empty.
- Animated caption bottom center, chunks of 3–6 words on their timestamps, one keyword in orange-red, fade+slide-up 0.25s (obeying the no-overlapping-lines rule above).
- Stat callout "95.8% → 58.3%": big figures at 900, the bad number in orange-red, the → arrow in orange-red, count-up as it appears.
- List item: a huge "01" + title + grey sub-label, with a vertical blue gradient bar on the left.
- Comparison table: glass rows at 14px radius, grey label on the left + big white value on the right, the key row highlighted on a light red background; stagger 0.1s.
- Pull-quote: a big faded " mark on top, the quote bold and centered, fading in line by line, the payoff words highlighted light red.
- Pill badge: "ARXIV xxxx.xxxxx" on a rounded glass background, small white uppercase text.
- Outro: a question + a "Drop ý kiến vào comment" button (glass pill + SVG icon), a few floating emoji.
PAPER SCENE
- Show the real PDF page (the user supplies an image per page), slow Ken Burns zoom/pan over 6–8s. ⚠️ Wrap the
<img>in a<div>and animate the WRAPPER (never animate width/top/left directly on an img/video). - Vietnamese heading overlay + the ARXIV pill - both real HTML elements.
- Light red highlight over 1–2 payoff sentences: an HTML highlight box laid on top, animated with a horizontal
clip-pathsweep.
EFFECTS (GSAP)
- Text reveal: y 40→0 + opacity 0→1, stagger 0.06s, ease
power3.out. - Figures: count-up (tween 1 object + onUpdate) + scale-pop 0.9→1.03→1.
- Light red highlight:
clip-path: inset(0 100% 0 0)→inset(0 0% 0 0). - Scene change (blur-dissolve - verified, see the ⚠️ TRANSITION section below); paper scenes zoom-fade.
- ⚠️ NEVER use
setTimeout/Date.now()/Math.random()- GSAP timeline only (deterministic). Every animation locked to the FPS.
⚠️ SCENE TRANSITION (blur-dissolve - the pattern that works)
Standard layout: the background (ambient-bg) is one sub-comp running continuously end to end on its own track behind everything; each text scene is a transparent layer (no background of its own). Because the background is continuous, the transition only animates the text, no background crossfade needed.
- Every scene handles its own entrance + exit inside its own timeline. Do NOT overlap clips, do NOT switch tracks, do NOT animate wrapper opacity at the master level → this keeps the caption/SFX sync intact.
- The exit (end of each scene, ~0.3s before the cut) = fade + slide up + blur in a single tween, ease
power2.in:
Set// applies to the scene's content wrapper + page counter (EXCEPT the final outro scene - no exit there) tl.to([sel(".sX-wrap"), sel(".counter")], { opacity: 0, y: -16, filter: "blur(10px)", duration: 0.32, ease: "power2.in" }, EXIT_AT);EXIT_ATso thatEXIT_AT + 0.32 ≈ data-durationof the scene. Still pad withtl.set({}, {}, DUR)at the end (Law 11). - The entrance (start of the next scene) = fade + staggered line-by-line reveal (already part of the standard reveal) → the eye reads it as "the old text blurs away → the new text arrives".
- ⚠️ This is a dissolve over a shared background, NOT a crossfade of two overlapping scenes. A true overlapping crossfade (the new scene's text appearing while the old text is still dissolving) requires alternating tracks (e.g. 2↔14) + a ~0.35s time overlap - heavier, only do it if the user asks.
- Captions live on their own track → they never pick up the scene blur → the subtitles stay sharp right through the transition (that is intentional, do not "fix" it).
- Strength:
blur(8–12px). Verify: extract a frame mid exit move with an accurate seek,ffmpeg -i in.mp4 -ss <t> -frames:v 1(put-ssAFTER-i;-ssbefore-ifast-seeks to a keyframe → you miss the move and wrongly conclude the blur is not running).
SOUND EFFECT (SFX) - pick them yourself & wire them to timestamps
Add SFX to punctuate the video (figures, reveals, scene changes, punchlines). Take them from the pre-curated library:
assets/sound-effects/ - from inside a project: ../../assets/sound-effects/. The recommended set = the entries tagged hay-dung in assets/sound-effects/library.json. Skip any entry whose file is not on disk - library.json still lists a few files that are not shipped with the repo.
The exact <audio id="sfx-*"> wiring is in "Wiring technique (HyperFrames)" below.
Picking SFX yourself (by content + timestamp)
- Read the transcript + caption timestamps + storyboard → find the "accent points" that need sound: figures/stat count-ups, keyword reveals, a paper scene opening, the red highlight on a payoff line, hard scene changes, the outro CTA, a punchline/shock line.
- Map each accent point → 1 semantically fitting SFX (table below).
data-start= the exact EVENT beat (the number starting its count-up / the text starting its reveal / the scene cut frame), NOT the vague start of the scene. - ⚠️ Restraint: about 1 SFX every 3–6s, not one hit per line of text (it dilutes them and tires the ear). A ~60s video wants ~6–12 SFX. Vary the files, avoid repeating a single sound too densely.
- Present the SFX table to the user for approval (timecode | event | file | volume) together with the storyboard before wiring anything in.
Semantic map → file (in assets/sound-effects/)
| Content context | File |
|---|---|
| good numbers / score / positive reveal / correct tick | ding.mp3, ting.wav |
| money / profit / cost / "tỷ", "doanh thu", "$" | ka-ching.mp3, money.mp3, buy.mp3 |
| text/element popping in, light transition | pop.mp3 |
| snapshot/image/screenshot / a PDF paper scene appearing | camera-snap.wav, camera-flash-1.wav |
| bad numbers / wrong / warning / "tệ", "giảm", "lỗi" | error.mp3 |
| shock reveal / wow / surprise | anime-wow-1.mp3 |
| typing / code / data entry | mechanical-keyboard.mp3, iphone-typing.mp3 |
| UI click / selection / button press | click-button.mp3, mouse-click.mp3 |
| comedic punchline / sarcasm / light fail | dry-fart.mp3, fart-echo.mp3, duck-toy.mp3, jontron-what.mp3 |
| "ăn" / swallowing / merging / consuming | chomp.mp3 |
(.wav and .mp3 both work - the right meaning matters more than the format.)
Wiring technique (HyperFrames)
- Copy the chosen file into the project's
assets/sfx/(keeps the project portable - do NOT letsrcpoint outside the project folder). - Each SFX = its own
<audio>in the root composition, next to the narration<audio>. ⚠️ NOclass="clip"on<audio>(Render Contract - it breaks audio):<audio id="sfx-stat1" data-start="12.46" data-duration="2.0" data-track-index="6" data-volume="0.5" src="assets/sfx/ka-ching.mp3"></audio> - ⚠️ Every SFX gets its OWN
data-track-index(e.g. 4,5,6,…) - clips on the same track must NOT overlap in time. Narration is on track 0; SFX use tracks 4 and up (so they never collide with the visual scene tracks). - ⚠️ Volume: SFX
data-volume0.4–0.6 (under the narration at 1.0) so they never bury the voice. Comedic/punch SFX can go 0.55–0.65. data-durationlong enough for the SFX to finish ringing (1–3s); too short and it gets cut off. It does not need to match the scene length.
⚠️ ENVIRONMENT (Windows) + TRANSCRIPTION
- Install FFmpeg with
winget install Gyan.FFmpeg. The PATH change only applies to NEW shells → when callingnpx hyperframesyou must prepend the bin to$env:Pathinline. - Dependencies (npm packages, Chromium, faster-whisper in
.runtime/venv, FFmpeg in.runtime/bin) are installed bystart/doctor.mjs(the Doctor page). Never runnpm install,pip installorwingetyourself - the agent policy blocks them (see video-pipeline, "Agent sandbox"); if something is missing, stop and ask the user to run Doctor.npx hyperframes browser ensureis fine. - ⚠️
npx hyperframes transcribeneeds an externalwhisper-cpp+ an.enEnglish-only model → do NOT use it for Vietnamese. - Vietnamese transcription: faster-whisper
large-v3,language="vi",word_timestamps=True(in Python:sys.stdout.reconfigure(encoding="utf-8")). Preferdevice="cuda", compute_type="float16"(on a GTX 1660: twice as fast, and it frees the CPU), with a try/except fallback todevice="cpu", compute_type="int8". Extract the audio:ffmpeg -i in.mp4 -vn -ac 1 -ar 16000 audio.wav. Fix Whisper's misheard proper nouns (e.g. Anthopix→Anthropic, Cloudy→Claude).
TRIMMING TO HIGHLIGHTS (if the source video is long)
ffmpeg filter_complex trim+concat several stretches into a re-encoded face.mp4; remap the word timestamps (new = orig - accumulated shift); generate captions grouped by punctuation / speech pause (gap > ~0.38s) / max 7 words.
RENDER CONTRACT (must)
- Root:
id,data-composition-id,data-start="0",data-width,data-height. - Any timed element needs
class="clip"(EXCEPT<video>/<audio>), plusdata-start/data-duration/data-track-index. Clips on the same track must NOT overlap in time (watch out for float rounding - leave a ~0.04s margin). <video>must bemuted; audio goes in its own<audio>. Do NOT animate width/height/top/left on a<video>/<img>- wrap it in a div and animate the wrapper (scale transforms are fine).- Every composition registers exactly 1
pausedtimeline onwindow.__timelines["<data-composition-id>"](the key must match exactly). Comp duration =tl.duration(). Pad the full slot:tl.set({}, {}, DUR)at the end of every timeline (Law 11 - shorter than data-duration → black frames).
WORKFLOW
- The user supplies: the Vietnamese transcript + timestamps + images of the paper pages + the figures to emphasize.
- Split into N scenes, map the components, set data-start/data-duration. Present the storyboard to the user for approval (table: scene | duration | component | content).
- Pick the SFX: scan the accent points by timestamp → map files from
assets/sound-effects/(prefer thehay-dungtag inlibrary.json) → copy intoassets/sfx/→ present the SFX table to the user for approval (timecode | event | file | volume). - Verify the font (test string, 1080×1920 screenshot) + confirm ALL text is a real HTML element.
npx hyperframes lint(0 errors; the self-selector/google-fonts/font-face warnings are benign).- Test-render 1 text scene + 1 paper scene (draft) → extract frames + zoom 3× to check the diacritics + check the caption boundaries + timing → listen back and check the SFX lands on its beat and never buries the voice → fix → render the full draft.
- User approves → render final with
--quality standard.
COMMANDS (inside video-projects/<project>/)
$bin="...\ffmpeg-...\bin"; $env:Path="$bin;$env:Path"
npx hyperframes lint
npx hyperframes compositions
npx hyperframes render --quality draft --output renders/draft.mp4
npx hyperframes render --quality standard --output renders/final.mp4
# verify diacritics: ffmpeg -ss <t> -i renders/draft.mp4 -frames:v 1 -vf "crop=700:340:60:260,scale=1400:680:flags=neighbor" out.png → Read out.png
# copy an SFX into the project: cp ../../assets/sound-effects/<file> assets/sfx/
# verify SFX: ffmpeg -ss <t> -i renders/draft.mp4 -t 4 -vn out.wav → listen that the SFX lands on its beat and never buries the narration
OUTPUT - hand off to the pipeline
renders/final.mp4 is NOT the finished product. Continue with the video-pipeline skill: automated QC on the draft → assemble final via the render queue (Remotion) → outputs/<id>-v<N>.mp4 → update meta.json (status + output) → thumbnail/publish. Report to the user: a timeline table: scene | duration | component | content + an SFX table: timecode | event | file | volume. Clean up the temporary draft/test files when you are done.