Video & Animation

「Vox 拼贴风科普短视频」工作流：给一个科普主题（新闻、诺奖、科学发现、科技事件），先上网查权威资料和可用的真实素材（NASA、Wikimedia Commons、官方新闻稿插图），写一份像聊天一样有语气、有强钩子和深度结尾的口播稿，用带语气提示的豆包复刻音色配音，再在 weilanx-claude-videos 工程里做成横屏 16:9 的 Vox 拼贴片：A-roll 是黑卡大字，B-roll 是钉在旧纸桌面上的真实视频和照片，按字幕硬切，配一条有律动的电子配乐。最后打开 Remotion Studio 给用户预览，确认后再导出。用户说「做一期 Vox 风格的科普」「用 Vox 风格讲讲 XX」「做一期 XX 的科普，要真实素材」「像诺奖那期一样做一期」时使用。

Generate narration locally in a consented speaker's voice using the bundled Chatterbox TTS model. Works with any language supported by the installed checkpoint. Audio finishing and soundtrack mixing are separate steps.

「参考视频 → 新风格素材包」蒸馏工作流：把用户给的一支或多支参考视频逐镜头拆解（硬切、镜头表、关键帧、联系表、主色板与强调色、运动强度、动效切片、配乐卡点），提炼出配色、字体、版式、入场 / 转场 / 镜头运动、节奏和招牌画面这些「画面语法」，再在 weilanx-claude-videos 工程里从零写成一套可复用的 Remotion 风格素材包：src/library 下一个以风格 ID 命名的组件目录、一个 library- 开头的预览作品、styles/我的风格 下的风格配方卡，并登记进素材库目录和风格总表。用户说「蒸馏这个视频的风格」「把这几个视频的风格整理成一套素材」「一比一复刻这个视频的样式」「拆解这个视频的动效和画面」「照这个视频做一套新风格」，或者丢来参考视频（template/、Downloads 里的 mp4）想学它的样子时都用这个 skill；只想复刻单个动画片段、不需要进素材库的，也先用它做拆解。

Review a finished video export for story coverage, readable text, framing, motion graphics, audio, and technical integrity before delivery.

Turn explicit creator feedback on pacing, framing, titles, transitions, and audio into a reusable style note without overgeneralizing one-off corrections.

Audit the picture and audio at edit boundaries in a rendered video, catching clipped speech, sudden sound changes, black frames, awkward dissolves, and visible jump cuts.

Prepare source-linked cutaway clips for any long or short video, with timecodes, rights, clear names, format-aware framing, and a contact sheet.

Measure and finish the mixed audio of a Reel, Short, long video, or audio-only deliverable with local FFmpeg and an encoded-output check.

Build a coherent travel, food, visit, or topic-driven video from inventoried footage, choosing story beats, clip order, pacing, still-photo placement, and an edit decision list.

Design and add short motion graphics inside an existing video, such as titles, maps, labels, lower thirds, photo moves, and transitions, using Remotion or HyperFrames where appropriate.

Compare multiple frames from videos or photos across cameras, lighting, and output formats, then review a deliberate color adjustment.

Prepare source-backed cover art, title or caption, credits, and optional long-form chapters for YouTube videos, Shorts, Instagram Reels, or other delivery formats.

Reframe a selected source range for wide YouTube, vertical Reels or Shorts, square, or portrait-feed delivery with fit or intentional crop.

Match narration or dialogue to indexed footage and audit a video timeline for off-topic shots using local, free tools. Use for footage searches and voiceover-to-picture alignment; do not treat filename similarity as visual proof.

动效音乐与音效方法论：BGM 三选一、拍点对齐、SFX 峰值对齐、响度 -14 LUFS。混音之前先用它定方案。

Inventory and classify footage, photos, audio, and factual evidence before a footage-based video edit; record required shots and source-backed labels in this repository.

纯代码渲染引擎方法论：seek(t) 确定性时间函数、静帧先行探针、子帧动态模糊、无头浏览器逐帧渲染。写渲染代码时用它。

动效导出与质检方法论：ffmpeg 导出参数、平台规格、成片自查流程。点"导出"之前先用它过一遍。

将人物演奏视频与单独录制的鼓声音频同步剪成架子鼓成片。适用于筛选正确录制版本、保留歌曲前奏、对齐落鼓动作与鼓声、删除准备或操作设备画面、检查漂移并轻度提亮。

Biến nguyên liệu thô (link, file PDF/ảnh/slide, đoạn mô tả của người dùng) thành videos/NNN-slug/idea.md có cấu trúc cho kênh devdraw. Dùng khi người dùng dán link, đính kèm tài liệu hoặc kể một ý tưởng và muốn "lên idea", "làm video về cái này", "lưu ý tưởng này". Không dùng để viết script.json.

动效导演 brief 方法论：需求 5 问澄清、分镜三段式、拍点网格、交付规格。写任何渲染代码之前先用它定方向。

Turn a vehicle on the Winn Kia of Fremont lot into three branded 10-second vertical reels (1080x1920) — "heartfelt" (warm family story), "sky-drop" (the car falls from the sky and lands), and "exploded" (the car bursts into floating parts and reassembles). Pulls the real listing photos from winnkiaoffremont.com, generates each clip with Higgsfield Seedance 2.5 using those photos as references, then adds the Winn logos, a hook, captions written for that car, a music bed and an end card with the dealership address. Use this whenever the user names a Kia model, stock number, VIN or Winn Kia listing URL and wants a reel, video, social content, "the three videos", or "do the same for this car" — and for follow-ups such as changing captions, re-running one style, or redoing a clip.

把真人口播竖屏短视频剪出「网感」——用 Remotion 做文字版式、换场画面、补停顿、配音效，并守住社媒安全区。用于：剪口播、口播精剪、抖音精剪（B-roll / 小人像窗 PIP / 分屏 / 海报大字）、抖音/Reels/Shorts 竖屏短视频、字幕太长或跑边、画面太平没网感、「用剪映那种花字/版式」、talking-head short video editing、Douyin-style 精剪 edit、B-roll、kinetic captions、safe zone。需要 1080×1920 的口播视频；有稿子更好。不用于：没有真人口播的品牌宣传片、多场景长片、纯素材混剪、静态标题卡。

多通道 AI 视频生成 + 配音合成引擎（fal.ai 队列：Wan 3.0 / Seedance 2.0+2.5 / MiniMax H3 / Gemini Omni Flash；TTS 用 ElevenLabs 英文 + 火山豆包中文；ffmpeg 合成）。当用户要求生成视频、文生视频、图生视频、视频编辑、配音、TTS、口播、短视频、BGM 混音，或提到 视频、视频配乐、视频号、抖音视频、短片、video clip、voiceover 时使用——即使没有说"生成"。需要 FAL_KEY（fal.ai）；可选 ELEVENLABS_API_KEY（英文配音）/ VOLC_TTS_API_KEY（中文配音，豆包 seed-tts V3）；REPLICATE_API_TOKEN 通道已脚手架化。零依赖，Node ≥ 18（合成与时长校验需要 ffmpeg/ffprobe）。模型路由、端点、参数枚举与定价集中在一个 JSON 注册表——加模型=改配置，不改代码。