Communitygithub.com

svg-voiced-video-studio

将口播稿、分镜脚本或选题制作成带本地个人音色配音的 60fps 原生几何 SVG 动态视频,并配套四尺寸防裁切高清封面、同步 SRT 字幕与五平台(B站、小红书、抖音、视频号、快手)过检标题简介及置顶首评。当用户要求制作哲思或商业解说 SVG 有声视频、排查 GPT-SoVITS 长句吞字或微调新音色、批量生成防宫格切字四规格封面时使用;不处理实拍素材剪辑或写实人物插画生成。

svg-voiced-video-studio 是什么?

svg-voiced-video-studio is a Claude Code agent skill that 将口播稿、分镜脚本或选题制作成带本地个人音色配音的 60fps 原生几何 SVG 动态视频,并配套四尺寸防裁切高清封面、同步 SRT 字幕与五平台(B站、小红书、抖音、视频号、快手)过检标题简介及置顶首评。当用户要求制作哲思或商业解说 SVG 有声视频、排查 GPT-SoVITS 长句吞字或微调新音色、批量生成防宫格切字四规格封面时使用;不处理实拍素材剪辑或写实人物插画生成。.

兼容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/zou172866/svg-voiced-video-studio/tree/main/skills/svg-voiced-video-studio

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档

svg-voiced-video-studio: 60fps Voiced SVG Video & Multi-Platform Delivery Workshop

Transform scripts, storyboards, or verified cases into local personal voiceover with zero word swallowing + 4K 60fps native geometric SVG motion video + @2x four-size crop-safe covers + five-platform titles, descriptions, and pinned comments.

Key Invariants

  1. Inspect the script and voice environment first, then split sentences and synthesize with multi-pass caching. Based on: Chinese disyllabic rhythmic word theory, audience common ground theory, Grice's conversational maxims, and content-addressed idempotent construction. Before synthesis, always run lint_script.py to check: single-word officialese, binary reversal, five types of AI voice fingerprints (false empathy, false profundity, self-weighting, universal templates, triple parallelism), cold-open announcements, parentheses fillers, overly long breath gaps, sentence length coefficient of variation (≥ 0.25), polyphonic characters, and number/English reading styles; for measured step-by-step content, retain audience-common measurement units (e.g., token) and unify throughout, address objects and sources by real names, and converge selection chronologically; lock misread words using {pronunciation=>subtitle} or dual-track ttsText. The synthesis script supports --check-only pre-check and out/.tts_cache/ hash-based incremental continuation, physically splitting sentences at ;:—— (breath gap 0.18s) and 。!? (breath gap 0.30s) with multi-pass best selection (the sole parameter source is in Voice & Script Craft Manual).
  2. Reverse-drive the 60fps timeline from measured seconds, with fast exits and reveal frames aligned to key points. Based on: normalized time parameterization and audiovisual synchronization perception window. The synthesis script writes scenesData.json; components must never hardcode frame numbers; springs use SPRING_PRESETS four-tier unified for similar elements (mode C locks to no-bounce gear), exit 0.25s faster than entry; mask transitions use revealOn() to align reveal frames with emphasized words or cut points; Chinese on-screen text holds for minReadHoldSec() (chars÷4.5+1.5s, key digits ≥ 2.5s).
  3. Extract dual-state static frames for visual inspection first, then run full video encoding. Based on: defect detection shift-left and Gestalt figure-ground separation. For each scene, extract at 25% (entry state) and 76% (text-heavy state); scenes with overlapping lines or trajectory crossings additionally extract at 48%; confirm lines do not intersect text, paths are closed, text frame clearance ≥ 8px, and at most 1 accent color focus per screen.
  4. Generate all four cover sizes at once, strictly lock central 3:4 safe zone for 9:16. Based on: multi-viewport safe frame and mixed Chinese-Western visual character width norms. Cover main titles: pure Chinese characters ≤ 6 per line, mixed Chinese-Western visual character width ≤ 6.2 em (unbreakable long proper nouns use --allow-7 special approval up to 7.2 em with automatic font size reduction 8%); 9:16 version top/bottom margin ≥ 240px.
  5. Copywriting must guard language and factual boundaries. Based on: expectation violation check and curse of knowledge. Pure text, zero formatting, remove quotation marks, Kuaishou max 4 hashtags, forbid binary reversal sentences, forbid presenting hypothetical deductions as author's established facts, creator's fixed closing phrase preserved verbatim—detailed rules unified in Five-Platform Copy & Four-Size Cover Spec section 2, article 0; applies equally to narration and covers.
  6. Philosophical/humanistic pieces: final scene preserves creator's fixed closing phrase verbatim. With unceasing inquiry, find the living self; within limited consensus, achieve infinite possibilities. If already present in the submitted script, leave unchanged; when generating a script from a topic, default as the final scene closing; split at semicolon into two sentences for synthesis; leave 4.5s~5.5s silence after the final period. Pure commercial quick explainers follow workspace rules or user instructions.
  7. Scenes dynamically expand/contract based on information points; visual aesthetics, data integrity, and Swiss layout follow the manual. Based on: Mayer's multimedia learning principles, Tufte's graphical honesty, and Swiss modular grid. Strictly forbid rigidly fitting or filling a fixed number of scenes: naturally split by "one core information point per scene" with complexity determining length (short commentary 3~5 scenes, standard analysis 5~9 scenes, deep long-form 10~16+ scenes all freely scalable); data charts (unverified numbers never shown on screen, coordinates before data, bars start from 0, verify multipliers, source small text), horizontal 12 / vertical 6 column grid, and 2~3 anti-homogenization style signals per video, all follow SVG Motion Aesthetics Manual.
  8. Final video must pass -14 LUFS technical inspection and independent blind review before delivery. Based on: ITU-R BS.1770 loudness standard and independent verification to prevent confirmation bias. Final mix adds percussion sound effects avoiding voice (sfx_synth.py), levels to -14.0 LUFS (true peak ≤ -1.2 dBTP); after rendering, always run inspect_video.py to strictly check frame count, zero black frames, loudness, and true peak (unchecked items marked NOT_CHECKED do not count as pass); also execute independent blind review.

Visual Master & Canvas Mode Selection

Choose based on script style and primary publishing platform (if workspace AGENTS.md or user specifies colors, follow that):

ModeLogical Canvas → 4K Output (--scale=2)Core Color Palette (zero gradients, 0px sharp corners)Composition PresetCover ThemeSuitable Topics
A. Deep Charcoal Night Car Philosophical1920×1080 → 3840×2160Background #111110 · Main Ink #F2EFE9 · Accent #D94E28--preset long --speed 1.00--theme darkIntrospective analysis, existentialist long-form
B. Swiss Classic Warm Ledger1920×1080 → 3840×2160Background #F5F0E6 · Paper #FAF7F1 · Charcoal Ink #191715 · Wax Seal #C83E2B--preset long --speed 1.00--theme vellumTime & money ledgers, classical philosophy long-form
C. Swiss Warm Paper Media Vertical Quick Explainer1080×1920 → 2160×3840Background #F5F1EA · Paper #FAF8F3 · Ink Black #1A1A18 · Vermilion #D93B22--preset short --speed 1.07--theme editorialVertical business analysis, product reviews, real data short videos (scenes naturally expand/contract with information)

Numbers use font-variant-numeric: lining-nums tabular-nums; forbid default old-style numeral fonts (e.g., Georgia).

Mode C Additional Notes: Shell passes theme={EDITORIAL_PAPER_THEME}; vertical canvas automatically switches platform obstruction safe zones (top ≥ 200px, bottom ≥ 400px, right ≥ 140px, single-line subtitles ≤ 16 characters); graphics only drawn within safe rectangle using swissGrid(safe) with 6-column grid; verify number reading list from lint output item by item.


Standard Five-Step Workflow

If the workspace AGENTS.md has manual review gates (topic, final script review), stop at the corresponding gate; do not render before approval:

  • Pre-finalization internal blind review & decision recording: Before submitting the final script package, if the host supports sub-agents, first dispatch a sub-agent not involved in writing the script to check against the five AI voice types and factual boundaries, clean them up. The finalization package includes: "one-sentence core claim, mode and 2-3 chosen style signals (at least 1 breaking away from the most cliché version), voice and speed, scene-by-scene narration with *keywords* and nine-column storyboard self-check, four-size cover text, five-platform titles" numbered item by item; after user approval, immediately write to out/定稿决策.md; subsequent rendering and cross-session handoffs treat this file as the sole source of truth.

Step 1: Script Inspection & Zero-Word-Swallowing Cached Voiceover Synthesis & Mixing

  1. Run python scripts/lint_script.py <script_path>, clean up sentence structures, five AI voice types, and officialese single words; verify polyphonic character and number/English reading list; if readings are wrong, write {pronunciation=>original}.
  2. Locate gpt-sovits-studio/voice_profile.json (prefer running with .venv\Scripts\python.exe dedicated environment; first time add --check-only for second-level pre-check of weights, CUDA, and ffmpeg). For training new voice or encountering Windows memory errors, first read GPT-SoVITS Windows Single GPU Fine-Tuning Troubleshooting Guide.
  3. Run synth_voiceover_template.py (select --preset, --speed, --sfx auto; if containing English proper nouns or technical terms, add --asr-verify; already synthesized sentences automatically reuse via out/.tts_cache/; full recalculation add --force-resynth). Outputs: public/audio/S01..SXX.mp3 (for pure voice quality check), ambient_pad.mp3, sfx_bus.mp3, mixed_master_ducked.mp3 (48kHz -14.0 LUFS voice + ducked background music + avoided sound effects final master; the only audio track for the final video), src/data/scenesData.json, out/ subfolder .srt, chapters.txt and sfx_cues.json.
  4. For 1-2 key sentences, optionally add "fx": "canyon" | "tape" | "water" (see voice_fx.py); before encoding video, must use multimodal read_file to listen to each public/audio/S01..SXX.mp3 by ear (key check: English proper nouns and polyphonic characters have no blurry sounds, no skipped words; for pure Chinese fine-tuning, use {touken=>token} for token, {P D F=>PDF} for PDF to dual-track lock).

Step 2: Build Decoupled Remotion 60fps Native SVG Project

  1. If reusable node_modules exist, mount via Windows directory junction.
  2. Copy motion_60fps.ts and SceneFrameShell_60fps.tsx, follow project conventions:
    • Composition IDs: Main movie MainEssay, scenes Scene-S01…Scene-SXX; use toShellTimeline() to convert timeline;
    • Audio track: only attach audio/mixed_master_ducked.mp3 once on MainEssay root composition; keep playSceneAudio shell closed;
    • All interpolate() explicitly clamp; uniformly use SPRING_PRESETS, EASE_SNAP / EASE_GLIDE / EASE_WHIP, default 0.25s fast exit, revealOn() alignment, minReadHoldSec() hold duration, and swissGrid(safe) modular grid;
    • Aesthetic and data implementation use accentAt() + mixHex() (red follows voice), cameraAt() (single canvas pan), taperStroke() / inkJitter() (modes A, B ink lines); commercial charts strictly follow aesthetics manual section 2 data standards.

Step 3: Dual-State Sampling Inspection, -14 LUFS Programmatic Final Check & Independent Blind Review

  1. npx tsc --noEmit zero type errors.
  2. Run extract_keyframes_template.mjs to export key frames (default binds port 18123 to avoid Windows Hyper-V/WinNAT reserved port range 3000–3100), visually inspect for line-text intersections, gaps, orphans, and platform obstructions; then run audit_aesthetics.py to check red focus, element count, box lines, protagonist height, density variation, and font size hints.
  3. Execute npx remotion render … --scale=2 --codec=h264 --crf=16 --pixel-format=yuv420p --port=18123 to directly render native true vector 4K 60fps final video (vertical 2160×3840, horizontal 3840×2160; completely eliminates yuv420p 4:2:0 chroma subsampling causing vermilion edge blur; studio preview also uses --port=18123 to avoid No available ports found).
  4. Always run inspect_video.py (pass --expect-size 2160x3840 [horizontal 3840x2160], --expect-fps 60, --scenes-json src/data/scenesData.json), confirm 4K resolution, frame count, zero black frames, -14.0 ± 1.5 LUFS loudness, and ≤ -1.0 dBTP true peak all PASS.
  5. If the host supports sub-agents, hand over qa_sheet.png, qa_phone.png and "one thing the audience should remember" to an independent reviewer sub-agent who does not look at the code; receive back the most severe ≤ 3 issues with timestamps; fix and re-check before proceeding.

Step 4: Render @2x Four-Size Crop-Safe HD Cover Set

  1. Draft cover text according to Five-Platform Copy & Four-Size Cover Spec and pass three hard checks.
  2. Run render_4covers_template.py (choose --theme per mode table) to output four .png files, verify each has no single-character orphans and 9:16 text fully within central 3:4 safe frame.

Step 5: Deliver Pure Text Five-Platform Copy & Structured Acceptance Checklist

Output pure text, zero-format copy per spec section 2 (Bilibili timeline directly from out/chapters.txt; project records remain in out/, not mixed into delivery directory). Final reply provides clickable links to .mp4, .srt, four .png, copy .md, npm run studio preview and re-render commands, and a fixed brief delivery acceptance checklist (verified items, unverified items, design exception reasons).


Boundaries, Authorization & Stop Conditions

  • Applicable Boundaries: For real-scene editing, live-person keying, or 3D realistic rendering requests, directly state that this Skill focuses on procedural native SVG voiced motion video and multi-platform packaging, not forcing templates.
  • External Action Authorization: Default only generates files locally; without explicit user instruction, do not invoke browser or API to automatically upload/publish.
  • Completion Acceptance Criteria: All hard-check items in inspect_video.py must be PASS (no FAIL or NOT_CHECKED), all four-size covers and pure text five-platform copy are complete, and the acceptance checklist truthfully records the verification scope and exception explanations.

相关技能