Communitygithub.com

Scene (Storyboard)

Add B-roll and on-screen graphics to recorded voiceover: overlay HTML layers (annotation highlights, full-screen animations, zoom) onto screen recordings, bound to subtitle cues and driven frame-by-frame by the player, with live preview.

What is Scene (Storyboard)?

This AI agent skill, named "Scene (Storyboard)", is designed for creating visual storyboards and adding B-roll footage to pre-edited talking-head videos or screen recordings. It only activates when the user specifically requests tasks like storyboarding, matching visuals, adding animations, highlighting key points, or adding B-roll—never for trimming, subtitling, physical cuts, or final rendering. The skill operates on a strict two-pronged prerequisite: an edit list and subtitle tracks must already exist. It then reads subtitles.json to semantically segment the script, samples 8–12 frames of the current video to visually assess the scene, measures motion windows via pixel difference to identify stable overlaying periods, and finally generates visual modules (HTML/CSS/JS) that are bound to specific subtitle cues. These modules are placed as transparent HTML layers over the screen recording, driven frame-by-frame by the player, allowing the user to preview the composition directly without rendering. The skill supports platforms where a `videocut-cli` and `ensure-running` service are available (desktop-managed environments). It enforces a rigorous review cycle: never skip frame extraction or motion analysis, never visually estimate coordinates (use pixel statistics), never cover genuine evidence, and never use placeholder images for missing content. It also requires zoom to be used sparingly (≤1.6x, centered 62.5% of original frame) and brand icons to be rendered as official Unicode glyphs or downloaded from simple-icons—never hand-drawn substitutes.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/2789927464king-jpg/video-skill-ku/tree/HEAD/chengfeng-videocut-skills/plugins/chengfeng-videocut/skills/chengfeng-visual

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

What does Scene (Storyboard) do?

This is one task, not a step in a workflow. Call it whenever the user asks—come back to it after making edits.

There are only two prerequisites: the edit list and subtitles already exist (layers are placed per subtitle screen, timing is computed from the edit list).

Requires   edit-list.json, transcript.json, subtitles.json
Produces   visuals.json + modules/<number-name>/index.html

Once done, it stops; do not direct the user to the next step.

First, read and execute the Business Skill Phase Contract. Hard rules for module writing are in the Module Contract, judgment rules are in Visual Judgment.

0. Readiness

First execute the "Readiness Check" from Check Updates—whether skills are up-to-date and the Runtime is compatible; the plugin root is also located there (all <plugin-root> in this document refer to that literal path). Only proceed if "ready"; if "requires new session" or "stop", follow its instructions (including the "no self-built replacement UI" prohibition). The business skill does not include environment logic.

If the readiness result is runtime.kind=desktop-managed, directly reuse the stable CLI already installed by the desktop app and the same launchd/windows-task service; do not resolve Electron paths, install additional dependencies, or start a second Runtime.

Commands

node "<plugin-root>/scripts/ensure-running.cjs" --json
node "<plugin-root>/scripts/videocut-cli.cjs" visual get   <project> --json
node "<plugin-root>/scripts/videocut-cli.cjs" visual frame <project> --cues sub-0004,sub-0005 --count 12 --out <dir> --json
node "<plugin-root>/scripts/videocut-cli.cjs" visual add   <project> --module modules/01-xx/index.html --cues sub-0004,sub-0005 [--zoom x,y,w,h] [--id vis-0001] --json
node "<plugin-root>/scripts/videocut-cli.cjs" visual remove <project> --id vis-0001 --json

ensure-running must first prove that canonical port 5190 belongs to the platform's managed service; if it fails, pass through the error and stop—do not fall back to foreground or a custom port.

Layers are bound to subtitle screens (--cues); the product itself converts them to word IDs—layers are never bound to words that have not been spoken, and if the edit changes, layers move automatically. Do not store seconds.

Four Steps, No Skipping

① Text defines task      Read subtitles.json, segment by semantics, answer per segment:
                         Tutorial / demonstration / proof, or concept / logic / transition?
② Grab frames for visuals visual frame grabs 8–12 frames; really look with your eyes.
                         In tests, 2 out of 12 segments had visuals overturn the text prediction — this step is not a formality
③ Measure motion window   Pixel-diff adjacent frames to find moments of scrolling/switching/action.
                         Layers only overlay stable windows; annotations exit before motion starts
④ Create & place layers   Write module → visual add → verify each layer in preview

Judgment Rules for Step ②

Visual exactly matches what was said     → Annotate: circle, circle where the word is spoken. Do not cover real evidence.
Visual is weakly related / blank          → Full-screen animation: solid white background, from the style library.
Visual is clear enough on its own         → No action. This is the most common correct answer (in tests, 4 out of 13 segments required no action).

Step ③ Output Is Numbers, Not Feelings

Circle coordinates      Pixel statistics (grayscale scan to find text lines); never estimate by eye — eye estimation has been off by 50px.
Interaction moments     Per-frame pixel difference to find sudden changes (tests recorded jumps from 4 to 335); annotation exits 0.2s before the change.
Stable window           Scrolling is common at the start (user looking for content); bind the layer from the cue where the frame stabilizes.

Think Before Using Zoom

No zoom by default.     960x720 screen recording becomes blurry when zoomed; the user will notice immediately.
When zoom is necessary  ≤1.6x, area 62.5% square centered — composition integrity > reading small text.
Text too small to read  Use a circle to guide the eye; if it must be legible, re-record at high resolution, not by zooming.

Review Cycle

Name module directories by number (01-daily-report, 05-anim-task-split); the timeline track displays the directory name. The user reports by number; you modify only that layer, leave others unchanged.

Two discipline rules for modifying layers, each corresponding to a real incident:

When modifying a layer, always check the module      After narrowing/extending a layer, the hardcoded relative times in the module do not automatically adjust —
                                                      the exit time of an annotation may fall after the layer ends; never let that happen.
Verify by pixel screenshot, not by inspecting DOM attributes. An iframe with "visibility toggled correctly" was once physically outside the viewport —
                                                      all attributes were correct, the screen was entirely blank.

What Not To Do

  • Do not skip frame extraction and motion analysis and go straight to creation (both mistakes have occurred: covering genuine evidence, placing a circle over scrolling content).
  • Do not try to do something for every segment — filling every second is not the goal; "no action" is a common correct answer.
  • Do not store seconds, do not estimate coordinates by eye, do not add a dark overlay on a white background page.
  • Without real content, do not fill with placeholder images.
  • Select styles from the registry; do not invent them.
  • Brand icons: use the official Unicode glyph; do not hand-draw substitutes (an 'X' was drawn with two crossed lines and caught by the user). If the local index lacks an icon, fetch a single-color official glyph from simple-icons (black/white works perfectly with the style), save the original SVG in the module directory for reference; within the same clip, the same brand must always look identical.

Individual skills in this repo

This repo contains 8 individual skills — each has its own dedicated page.

2789927464king-jpg/video-skill-ku

剪辑环境的唯一管理者:就绪检查(skills 是否最新 → Runtime 是否配套)、Skills 更新激活、Runtime 安装与体检。用户说检查更新、安装剪辑环境、装播放器、检查剪辑环境、剪辑环境就绪了吗、配置转录凭证时使用;业务 Skill(剪口播/字幕/画面/导出)第 0 步也引用本 Skill 的就绪检查。不用于剪辑、字幕、画面、导出本身或项目数据迁移。

2789927464king-jpg/video-skill-ku

剪辑中文口播原素材:逐词转录、词典修字出修字表、五轮扫描找口误与重复、汇总表与重复句子表、打开 Studio 让用户复核、复盘沉淀用户偏好与词典。只产出一份已复核的删词账本,不切媒体、不做字幕、不做分镜动画。用户说剪口播、处理口误、生成口播基础素材、继续剪口播,或确认卡回传 action=return_cut_review 时使用。不要用于执行物理剪切、导出剪后视频、单独安装、单独打开工作台或口播分镜成片。

2789927464king-jpg/video-skill-ku

把剪好的口播烧成一个成片文件:账本切片段、推近、字幕、HTML 画面层,一次全部烧进 mp4。用户说导出、出成片、烧字幕、渲染、导出视频、生成最终文件时使用。不要用于生成删词候选、写字幕、做画面动画。

2789927464king-jpg/video-skill-ku

整理、脱敏并上报 chengfeng-videocut 的 GitHub Bug。用户说上报 Bug、反馈剪口播问题、提交 GitHub Issue、这个问题告诉开发者,或要求继续提交已预览的 Bug 草稿时使用。不要用于功能建议、普通排错、代码提交或未获用户确认的自动上报。

2789927464king-jpg/video-skill-ku

给剪好的口播做字幕:直接用已有的逐词稿加账本算出剪后时间(不必导出、不必重新转录)、用词典和作者文稿改写听错的专名、按句子分屏、在 Studio 里逐屏复核。用户说做字幕、加字幕、改字幕、重新分屏、字幕不对时使用。不要用于删词剪辑、物理剪切、分镜动画或成片渲染。

2789927464king-jpg/video-skill-ku

操作 chengfeng-videocut 剪辑工作台的已有工程,通过 Runtime CLI 查工程与词句位置、添加或替换画面覆盖层、移除画面、删除明确指定的词句或片段、拆分裁剪移动已有片段、调整画布比例并回读核验。用户说插入片段、删掉这句、移除动画、换个画面、调整顺序或管理时间线时使用;先区分画面覆盖与主轨顺延插入,不负责自动判断口误、创作动画、制作字幕、导出或排查安装故障。

2789927464king-jpg/video-skill-ku

Create Ian Xiaohei-style HTML/SVG motion illustrations for Chinese articles, scripts, storyboard scenes, workflow explanations, concept metaphors, and short video visual aids. Use when the user asks for 小黑 SVG、漫画感 HTML、分层 SVG 动效、把小黑生图做成 HTML/SVG、可编辑矢量动效、口播流程动画、手绘漫画动效、正文配图动画, or wants a Xiaohei illustration style adapted into controllable HTML/SVG/GSAP output instead of raster image generation.

2789927464king-jpg/video-skill-ku

剪映 (JianYing) AI自动化剪辑的高级封装 API (JyWrapper),提供开箱即用的 Python 接口,支持录屏、素材导入、字幕生成、Web 动效合成及项目导出。全面适配 MacOS (Apple Silicon/Intel) 与 Windows,支持 v5.9+ (draft_info.json) 架构、工程自修复、智能配音字幕及录屏变焦。

Related Skills