douyin-one-second-word-video
Create and revise 9:16 Chinese Douyin “一秒钟记一个单词” mnemonic videos for one English word using deterministic HyperFrames HTML/CSS/SVG/GSAP and Edge TTS. Use when a user supplies an English word and wants
SKILL.mdSKILL.md
GitHubDouyin One-Second Word Video
Create one complete vertical vocabulary video per English word. Preserve the approved structure; vary only the word-specific mnemonic, SVG example, copy, and speech-aligned timing.
Required companions
Read and apply hyperframes, hyperframes-core, hyperframes-cli, hbg-douyin-code-explainer-video, GSAP, and code-video-visual-qa. Use code-only HTML/CSS/SVG/GSAP.
Non-negotiable format
- Render 1080×1920 at 24 fps.
- Use Edge TTS
zh-CN-XiaoxiaoNeuralat+20%unless the user overrides it. - Generate one continuous semantic narration turn. Never use Qwen.
- Do not use image-generation or video-generation models.
- Do not add BGM unless requested.
- Create a new project or revision path; never overwrite an approved MP4.
- Read references/proven-structure.md before writing the storyboard or composition.
Workflow
- Generate at least two candidate sound-chunk splits. Prefer two or three chunks that map to Chinese units and can form a short causal story, even when the English word is technically one syllable. Call them sound chunks, not syllables. Use a single whole-word mnemonic only when every plausible split makes the pronunciation unrecognizable or cannot produce a natural mini-story; record the rejected candidates in
BRIEF.md. - Choose short Chinese sound mnemonics that combine into a meaningful action, image, or mini-event. Prefer a memorable natural phrase over a forced transliteration; admit approximation through the visual mapping instead of adding explanatory clutter.
- Write a conversational semantic bridge. Use a concrete action or object that naturally produces the English meaning. Reject sentences that merely concatenate the mnemonic and definition.
- Write
BRIEF.md,SCRIPT.md, andSTORYBOARD.mdbefore the composition. - Generate Edge TTS with scripts/generate_edge_tts.py. Use the emitted word boundaries as the timing source.
- Build the static hero frames first, then add seek-safe GSAP motion.
- Run HyperFrames
check, inspect source snapshots, render a new high-quality MP4, then inspect the same timestamps from the encoded MP4. - Run scripts/final_media_qa.sh on the final revision.
Copy contract
Use this narration arc, adapting punctuation for natural Edge TTS delivery:
一秒钟记一个单词,今天要记的是 <word>,<meaning>。
<Chinese mnemonic units, spoken in Chinese only>。
窍门:<mnemonic>!你看,<natural concrete example that ends in the meaning>!
所以,<word>!<word>!<word>!<word>!<meaning>。
Keep the example short enough to fit in at most four large lines. Put the pronunciation hint directly beside the Chinese mnemonic and put (<word>) beside the result meaning. Ensure parentheses remain visible during scale emphasis by giving annotation spans explicit horizontal margin.
Completion gate
Deliver only after all of these pass:
- Frame zero contains only the word, IPA, and meaning—never the mnemonic.
- The mnemonic uses at least two sound chunks whenever a defensible split exists. If a rare one-chunk fallback is used, record in
BRIEF.mdwhy the attempted splits failed pronunciation or story naturalness. - English chunks are visible before mnemonic speech, while Chinese mnemonic units appear one by one on their Chinese TTS boundaries.
- The narration never reads English split chunks.
- Assembly reuses the same live chunk elements, closes every gap, and holds the complete word clearly before the next beat.
- The example reads naturally and its annotations are unobscured.
- The final frame retains four word echoes, the meaning, and the IPA.
- Final MP4 is H.264/yuv420p, 1080×1920, 24 fps, AAC stereo 48 kHz, faststart, with no black frames or unexplained long silence.
- Full-resolution frames were inspected from the encoded MP4, not only from source snapshots.