| name | dasheng-html-video-bridge |
| description | Use when Dasheng transwrite needs no-human or human-material talking-head videos rendered through the local html-video project. |
Dasheng HTML Video Bridge
Role
Use the external html-video repository as the default renderer for Dasheng video transwrite. This skill turns a confirmed Draft into an executable html-video project package; it does not research facts or rewrite the article thesis.
html-video is an external dependency, not vendored into this repo and not version-locked. Default path: /Volumes/PSSD/html-video; override with HTML_VIDEO_ROOT.
Check or install before first use:
python3 scripts/ensure_video_external_deps.py --dep html-video --mode check
python3 scripts/ensure_video_external_deps.py --dep html-video --mode install --install-node-deps
For Dasheng no-human explainer scenes, external html-video must also provide gsap and lottie-web; ensure_video_external_deps.py --install-node-deps checks and installs them in the external repo.
Optional Lottie authoring uses external diffusionstudio/lottie (text-to-lottie) as a verified Skia Skottie player and scene workspace. It is an external dependency, not vendored or version-locked:
python3 scripts/ensure_video_external_deps.py --dep text-to-lottie --mode check
python3 scripts/ensure_video_external_deps.py --dep text-to-lottie --mode install --install-node-deps
Inputs
talking_head_video_manifest.jsonvideo_storyboard.jsontalking_head_script.mdstoryboard_template_review.html or equivalent approved pre-render scene/template contact tablestoryboard_review_decision.json plus a passing storyboard_review_gate.json- Optional human media: real video, real audio, or subtitle/SRT files
Default Renderer
Resolve the root from HTML_VIDEO_ROOT or /Volumes/PSSD/html-video, then run:
HTML_VIDEO_ROOT="${HTML_VIDEO_ROOT:-/Volumes/PSSD/html-video}"
node "$HTML_VIDEO_ROOT/packages/cli/dist/bin.js" doctor --cwd "$HTML_VIDEO_ROOT"
node "$HTML_VIDEO_ROOT/packages/cli/dist/bin.js" search-templates --intent "<intent>" --aspect 9:16 --top 5 --cwd "$HTML_VIDEO_ROOT"
Prefer these templates for market commentary:
frame-liquid-bg-hero: hook/title/opening atmosphereframe-data-chart-nyt: one clear data comparisonframe-electric-studio: quote or conflict frameframe-light-leak-cinema: cinematic transitionframe-logo-outro: ending card
Bridge Command
Generate a project plan without mutating html-video:
python3 scripts/transwrite_html_video_bridge.py \
--video-manifest <talking_head_video_manifest.json>
Create and preview a real html-video project only when the user asks to render:
python3 scripts/transwrite_html_video_bridge.py \
--video-manifest <talking_head_video_manifest.json> \
--execute create
Render MP4:
python3 scripts/transwrite_html_video_bridge.py \
--video-manifest <talking_head_video_manifest.json> \
--execute render
Two Production Modes
No Human Material
Flow:
draft -> short script -> storyboard beats -> TTS audio -> html-video template vars -> preview -> MP4
Rules:
- Do not use macOS
say for production voice unless explicitly accepted as a rough preview.
- Use MiniMax CLI (
mmx) as the default production provider for final narration, generated口播音频, background music, and AI image/video assets.
- Default synthetic narration voice is MiniMax
tianxin_xiaoling at 1.2x. After voice speed changes, rebuild or compress the visual timeline to keep audio and scene cuts aligned.
- For final/review videos, drive the timeline from real audio timing: per-scene TTS durations, provider alignment metadata, or ASR/forced-alignment backfill. Continuous TTS plus text-length estimation is preview-only.
- If a continuous TTS file has already been generated, backfill subtitle timing with
scripts/align_video_subtitles_to_asr.py before final render, so display text remains the approved script while timing follows the real voice.
- Coze or direct MiniMax API calls are fallback paths only when the CLI cannot express the required operation.
- Visuals must contain market/data metaphors, charts, and scene motion; do not render only enlarged outline bullets.
- Motion stack: HyperFrames scene composition + GSAP timing + optional Lottie accents. Lottie cannot replace real data charts or evidence.
- Final/review renders must record live HTML animation. Do not use screenshot PNG, static scene stitching, or FFmpeg
zoompan as a production renderer.
- Real-data scenes must render actual article/verified data. Do not use fake line paths, generic charts, or template-name-only visual switching.
- Do not create materials or render final video until the pre-render scene/template contact table has been reviewed and exported to
storyboard_review_decision.json.
- Validate the exported decision JSON with
scripts/validate_storyboard_review_gate.py; if any scene is pending, changed, edited, or deleted, return to storyboard revision instead of rendering.
- The table must include template screenshots or explicit missing-preview placeholders for every scene.
- When a TemplateShowcase MP4 exists, use
scripts/extract_template_preview_frames.py to build a template_previews/ directory and pass it into the review table generator.
- Final captions must be full timed subtitles, not shortened scene prompts. Generate JSON/SRT cues from the full narration, keep them synchronized to the voice timeline from real audio timing, and normalize years/quantities to Arabic numerals in display text.
Default voice render entry:
python3 scripts/render_html_anything_scene_pack_animated.py \
--manifest <scene_pack_manifest.json> \
--output-dir <render_dir> \
--with-voice \
--voice-provider mmx \
--voice "tianxin_xiaoling" \
--mmx-model speech-2.8-hd \
--mmx-speed 1.2
Human Material Present
Flow:
human audio/video -> Whisper/SRT -> beat alignment -> transparent/non-transparent HTML visual layer -> FFmpeg/html-video compose
Rules:
- Existing human audio is the master timeline.
- Visual frames follow the audio; do not force the speaker to follow animation timing.
- If only real video exists, extract audio first, then transcribe.
Text-to-Lottie Motion Layer
Use text-to-lottie when a scene needs reusable vector motion that is hard to author cleanly by hand in HTML/CSS:
- lower-third/name tag/term card accents
- warning badges, status feedback, market alert icons
- document scan, data flow, ticker, arrow/path motion accents
- chapter/outro symbols and compact social-video stickers
- transparent overlays for talking-head scenes
Do not use Lottie as the fact layer. Real charts, tables, screenshots, article images, and sourced data must still come from Draft/article assets or verified data. Lottie can decorate or direct attention to those facts.
Generation contract:
- Store generated Lottie JSON under the current task output folder, e.g.
~/Desktop/自媒体创作/<task>/lottie_assets/<scene_id>/lottie.json.
- Record each asset in the video manifest with
source_tool: text-to-lottie, background: transparent|full_frame, duration_frames, fps, intended_scene_ids, and fact_role: decorative|attention_cue.
- Verify in the upstream Skia Skottie player before using it in a render. At minimum inspect frame
0, midpoint, and op - 1.
- Import into html-video/Remotion with
lottie-web or @remotion/lottie; do not rasterize it into a static PNG for production.
- If verification fails, fall back to HTML/SVG/GSAP hand-authored motion, not static zoompan.
Output Contract
The bridge writes:
html_video_project_vars.jsonhtml_video_project_plan.jsonstoryboard_template_review.htmlstoryboard_review_decision.jsonstoryboard_review_gate.jsonhtml_video_commands.sh- Optional
html_video_execution.json when executed
talking_head_video_manifest.json must point to html_video_project_plan.json.
Hard Rules
- Draft owns facts, charts, and source claims.
- This skill owns video structure, visual metaphor, timing, and renderer handoff.
- Plans are not final videos; only mark
rendered after an MP4 exists.
- Keep output folders shallow: manifest, storyboard, script, vars, plan, renders.
- Never hardcode API keys or write credentials into manifests.
- Do not vendor or pin
html-video; install or update the external repo when the skill needs it.
