Communitygithub.com

SingleGrain-Team/video-skills

渲染一個 5-6 秒的動態圖形:超寬大字標題(2-4 個詞)在主持人背後突入,並將去背的主持人剪影疊在文字之上。接著一個硬切移除主持人,而相同標題保持完全相同的大小與位置,同時一行簡短輔助文字出現供閱讀停留。適用於談話頭介紹、章節開場、發布公告和社交鉤子。基於 HyperFrames 建置;輸出 1920x1080 H.264 MP4 及可編輯專案。

video-skills 是什麼?

這是一個專為 AI 代理(如 Claude Code、Cursor、Codex)設計的技能,用於自動化製作專業的動態圖形影片。它能夠產生一個 5-6 秒的片段:首先,一個超寬的粗體標題(2-4 個字)在主持人背後動畫進入,同時去背的主持人剪影覆蓋在文字上方,創造深度感;接著透過一個硬切瞬間移除主持人,但讓標題完全保持不動,然後再滑入一行輔助文字供觀眾閱讀。 此技能非常適合用於談話頭介紹、章節開場、產品發布或社群媒體鉤子。它支援 Claude Code、Cursor、Codex 等平台,並要求使用者提供一個主持人影片片段、一個大寫標題、一行輔助文字、一組顏色調色盤,以及兩個字型檔案。技能會自動執行影片裁剪、去背(使用 HyperFrames 的 remove-background 命令)、合成(使用 HyperFrames 的模板)和渲染,最終輸出一個 1920x1080、30 fps、H.264 編碼的 MP4 檔案以及完整的可編輯專案資料夾。其核心優勢在於確保剪輯前後標題完全一致,不會產生任何跳動,從而達到專業的視覺效果。

相容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/SingleGrain-Team/video-skills/tree/HEAD/skills/title-takeover

在你喜歡的 AI 中提問

開啟一個已預先載入此 Agent Skill 的新對話。

說明文件

title-takeover

一個想法,兩個節拍。節拍一:標題在講者背後猛然出現(深度)。節拍二:一個硬切移除講者,而完全相同的文字保持不動,同時一行文字提供清晰資訊(明確性)。剪輯前後使用相同的元素;唯一的變化就是這個剪輯。

輸入

輸入規則
主持人片段一個人,面向鏡頭,頭肩大致居中,1920x1080 橫向,至少 4 秒可用素材。靜音亦可。僅限自有或已授權素材。
標題2-4 個字,大寫,每行最多約 16 個字元。兩行效果最佳(例如 ADOPT AI / SKILLS)。可選一個強調詞。
輔助文字一個句子,理想為 4-9 個字,切勿超過約 50 個字元(必須在 60 px 字型大小下維持一行)。
調色盤背景、文字、強調色。深色背景搭配奶油色文字加上一個強調色效果最佳;保持對比度至少符合 WCAG AA 標準。
字型一個本地文件提供的粗體窄顯示字型(預設為 Anton,OFL 授權),加上一個乾淨的無襯線字型用於輔助文字(Inter)。

工作流程

  1. 準備片段。 裁剪出一個 6 秒的視窗,其中講者面向鏡頭且不離開畫面。移除音訊,標準化為 30 fps、1920x1080: ffmpeg -ss <in> -i src.mp4 -t 6 -an -r 30 -c:v libx264 -crf 14 -pix_fmt yuv420p assets/presenter.mp4
  2. 去背。 npx --yes hyperframes remove-background assets/presenter.mp4 -o assets/presenter-cutout.webm(透明 VP9 WebM,與來源相同的畫格數與解析度;在 Apple Silicon 上處理 180 畫格約需 25 秒)。若模型下載逾時,請重新執行;模型會快取在 ~/.cache/hyperframes 下。
  3. 合成。 複製 template/ 到專案中,將 presenter.mp4、presenter-cutout.webm、字型和標誌放在 assets/ 中,然後僅編輯標記的調色盤變數、兩行標題文字和輔助文字。圖層順序(由下至上):
    1. 背景(純色 + 柔和的強調色光暈)
    2. 主持人素材 + 暗色遮罩(使雜亂背景變暗,讓文字更突出)
    3. 標題
    4. 主持人剪影(與素材使用相同的 data-start / data-media-start / data-duration)
    5. 輔助文字、品牌標記
  4. 檢查並渲染。 npx --yes hyperframes check,然後執行 npx --yes hyperframes render --workers 2 --quality delivery --output final.mp4。

時間規則(6.0 秒影片,30 fps)

時刻時間規則
突入0.30 秒(第 1 行),0.45 秒(第 2 行)縮放從 1.4 到 1.0,透明度從 0 到 1,持續 0.45 秒,使用 expo.out。兩行間隔 0.15 秒。先顯示 0.3 秒單獨的主持人,讓標題落在一個即時的畫面上。
在主持人背後停留0.75 秒到剪輯點至少 2.4 秒讓講者可見於文字之上。
硬切3.4 秒(第 102 畫格)素材、遮罩和剪影在同一畫格結束。無淡入淡出、無轉場、無縮放。
輔助文字3.75 秒,上升 28 像素,持續 0.55 秒,強調色線條劃入在剪輯後留約 0.35 秒只有標題,以讓「停留」效果被感受到。
閱讀停留4.3 秒到 6.0 秒至少 1.5 秒完全靜止。以完整組合畫面結束;切勿淡出至黑畫面。

不變規則:

  • 標題是單一元素。在突入動畫穩定後(約第 24 畫格),永遠不要再調整其縮放、位置、大小或字距。剪輯必須不能移動任何一個字形。
  • 切勿在第二個場景中重新建立標題。
  • 剪影和素材必須畫格同步;兩者都加上 muted playsinline,不要使用 crossorigin。
  • 僅對標題行(它們位於一個靜態包裝器內)進行變換和透明度動畫。不要對 .clip 元素進行動畫。
  • 講者的臉會遮住部分文字。這是預期的效果;保持詞間距,並在兩側各保留一個可辨識的字母對,同時保持標題字型大小在 380-440 px,以便文字保持可讀。
  • 在第 1 行上方保留 140 px 的空間,確保字形永遠不會超出畫布。

驗證

  1. npx --yes hyperframes check 必須報告 0 個錯誤。標題在講者背後重疊的意圖已透過在剪影和標題行上宣告 data-layout-allow-occlusion 來表明(模板中已包含)。
  2. 從渲染的 MP4 中提取並檢查(不僅是計數)以下畫格:
    for n in 90 101 102 103 179; do ffmpeg -y -i final.mp4 -vf "select=eq(n\,$n)" -frames:v 1 qc_$n.png; done
    
    • 90:標題在講者背後可見,剪影邊緣乾淨(無光暈、無缺失的頭髮或肩膀)。
    • 101(最後一個講者畫格)和 102(剪輯後的第一個畫格):第 102 畫格上講者消失,標題在兩者中的大小和位置完全相同。
    • 103:仍然沒有輔助文字或剛開始出現;標題不變。
    • 179(最後):標題加上完整的輔助文字,完全清晰可見,無任何裁切。
  3. 如果剪影顯示邊緣條紋或閃爍,請使用 --quality best 重新執行 remove-background,或選擇一個更乾淨的來源視窗。如果標題在剪輯處跳動,表示仍有動畫在對其執行;請移除它。
  4. ffprobe 必須顯示 h264、1920x1080、30 fps、6.0 秒。

交付

交付 final.mp4 以及專案資料夾(index.html、assets/)。在 NOTES.md 中記錄素材來源 URL 和授權。切勿使用可辨識的創作者或名人,或第三方標誌的素材。

Individual skills in this repo

This repo contains 7 individual skills — each has its own dedicated page.

SingleGrain-Team/video-skills

Make a 30-45 second launch film for a product by staging it inside one persistent, neutral chat window. A user types a request, the assistant answers and calls the product as a connected tool, a progress card runs, a follow-up triggers a second tool action, and the result lands as a count-up headline number before closing on the logo, tagline and call to action. Use when you need a launch, feature or integration announcement video for something that plugs into AI assistants (a skill, MCP server, plugin, connector, CLI). Delivers a rendered 1920x1080 MP4 plus an editable single-file HyperFrames project.

SingleGrain-Team/video-skills

Build a motion graphic where a screen recording plays in a large frame beside a checklist, and each checklist item ticks off at the exact moment that step happens in the recording, ending with every item checked and a short hold. Use when someone has a screen recording (product walkthrough, tutorial, onboarding, install demo) plus a list of steps with timestamps and wants a "watch the steps get done" film. Input is a recording, step labels with timestamps, a title and an optional palette; output is a 1920x1080 H.264 MP4 and the editable HyperFrames project.

SingleGrain-Team/video-skills

Build a 15-20 second product launch film that shows the product's real screen on a laptop built from plain CSS (no 3D model, no generation model) in a dark studio, scrolls the page, zooms into the interface for one key action such as typing a search, and resolves to the logo lockup. Use when someone has a full-page screenshot or screen recording of their product, one action worth showing, and a logo, and wants a short silent-friendly launch teaser rendered as a 1920x1080 MP4 with an editable HyperFrames project.

SingleGrain-Team/video-skills

Builds a 6-8 second logo sting in which the brand mark rides inside a refracting liquid bead (distorted past legibility), the bead swells, holds still for a beat, then ruptures into droplets and leaves the logo lockup (mark plus wordmark) crisp and upright on screen. Delivered as a 1920x1080 H.264 MP4 plus the editable HyperFrames project, with no 3D model, stock footage or generation model. Use it for brand intros and outros, product-launch openers, channel stings and any "reveal our logo" moment, from an SVG or PNG logo (including raster-only logos), a wordmark string and a palette.

SingleGrain-Team/video-skills

Make a 20-30 second launch film for a product that is not tied to one interface, such as a CLI, package, plugin or agent skill. The identical prompt or command is typed into four visually different surfaces in turn (a wide terminal, a tall chat card, a floating palette, an editor cell), the film then hard-cuts into a fast reel of real outputs on a beat or fixed tempo grid, and it closes on the command in a large pill. Delivered as a rendered 1920x1080 MP4 plus the editable HyperFrames project. Music is optional - with a track the cuts land on detected beats, without one they land on a 120 bpm grid.

SingleGrain-Team/video-skills

Make a roughly 30-second film that explains a code change (a pull request, release or uncommitted diff) to people who will never read the diff. It opens by showing the real diff scrolling past too fast to follow, cuts to a plain-language summary, fires four capability beats each pinned to a real code snippet, and closes on how to get it (install command and URL). Delivered as a 1920x1080 H.264 MP4 plus the editable HyperFrames project. Use when asked to turn a PR, changelog or release into a launch or explainer video.

SingleGrain-Team/video-skills

Builds a 10-14 second product sting out of a short run of UI moments (an entry screen, a single counting metric, a navigation list, one click that opens a menu, then an endcard) cut together so every cut carries the previous motion forward and the piece reads as one continuous camera move. Delivered as a 1920x1080 H.264 MP4 plus an editable single-file HyperFrames project. Use it for app launches, feature teasers, social bumpers and site intros when you have a product's logo, a couple of screenshots, one headline number and its real nav labels, and want polished UI motion without narration.

相關技能