This is one thing, not a stage of a process. Do it whenever the user calls for it; you can come back and do it after changing the edit.
There are only two prerequisites:the ledger and subtitles already exist(layers are placed according to the subtitle track, and times are calculated from the ledger).
需要 edit-list.json、transcript.json、subtitles.json
产出 visuals.json + modules/<序号-名字>/index.html
Stop when done,do not direct the user's next step。
First read and execute the business Skill's stage contract. For the hard rules of module writing, see module contract, and for judgment rules see visual judgment。
0. Ready
First execute the "readiness check" of check for updates — whether skills are
up to date and whether the Runtime is matched;plugin rootis also located there (in the commands in this document, <插件根> all substitute
that literal path). Continue only if "ready"; for "new session required" or "stop", follow its handling
(including the "no self-made substitute interface" prohibition). Business Skills do not carry their own 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 parse the Electron path, install additional dependencies, or start a second
Runtime.
command
node "<插件根>/scripts/ensure-running.cjs" --json
node "<插件根>/scripts/videocut-cli.cjs" visual get <project> --json
node "<插件根>/scripts/videocut-cli.cjs" visual frame <project> --cues sub-0004,sub-0005 --count 12 --out <dir> --json
node "<插件根>/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 "<插件根>/scripts/videocut-cli.cjs" visual remove <project> --id vis-0001 --json
ensure-running must first prove that canonical 5190 belongs to the managed service of the current platform; if it fails, pass it through and
stop, do not fall back to foreground or choose your own port.
Layer bound tosubtitle track（--cues), the product converts it into word ids itself — a layer can never be bound to a word nobody said,
and when the edit changes the layer moves along with it.Do not store seconds.
Four steps, do not skip a single one
① 文字定任务 读 subtitles.json，按语义分段，每段回答：
教程/演示/证明，还是概念/逻辑/过渡？
② 抽帧看画面 visual frame 抽 8-12 帧，真的用眼睛看。
实测 12 段里有 2 段被画面推翻文字预判 —— 这步不是走形式
③ 量运动窗口 相邻帧做像素差，找出滚动/切换/操作的时刻。
层只盖稳定窗口，圈在画面开始动之前退场
④ 做层放层 写模块 → visual add → 在预览里逐层核对
the judgment rules for ②
画面正是他说的东西 → 标注：圈住、说到哪个词圈哪里。真证据不许盖
画面弱相关/空白 → 小黑动画：整屏白底，从风格库做
画面自身足够清楚 → 不动。这是最常见的正确答案（实测 13 段里 4 段不动）
the output of ③ is numbers, not feelings
圈的坐标 像素统计量出来（灰度扫描找文字行），不许目测 —— 目测差过 50px
互动时刻 逐帧像素差找突变（实测一次从 4 跳到 335），圈在它之前 0.2s 退完
稳定窗口 开头常有滚动（他在找内容），层从停稳那屏的 cue 开始绑
Before using push-in (zoom), think
默认不推近。 960x720 的录屏放大就糊，用户会立刻看出来
非推不可时 ≤1.6 倍、区域取 62.5% 见方居中 —— 构图完整 > 读清小字
字太小读不清 用圈引导视线就够了；真要读清，答案是高分辨率重录，不是放大
review loop
Module directories usesequential numbering（01-daily-report、05-anim-task-split),
the directory name is shown on the timeline track; when the user reports a number, you change that layer and leave the rest alone.
Two layer-editing disciplines, each corresponding to a real incident:
改层必查模块 层收窄/延长后，模块里写死的相对时刻不会跟着变 ——
圈的出场时刻可能落到层结束之后，永远不出现
验收看像素 截图看画面，不查 DOM 属性。「可见性翻转正确」的 iframe
曾经物理上不在画面里 —— 属性全对、屏幕全空
What is not allowed
- Do not skip frame extraction and motion windows and do it directly (each of the two mistakes has been made once: covering real evidence, and a circle hanging over scrolling)
- Do not do something to every segment — filling everything is not the goal; doing nothing is a common answer
- Do not store seconds, do not eyeball coordinates, do not add darkening on white-background pages
- Without real content, do not pad with placeholder images
- Choose styles from the registry; do not invent them out of thin air
- Use official glyphs for brand icons; do not hand-draw substitutes(X was drawn with two crossing strokes and the user caught it). If it is not in the local index, take the monochrome official glyph from simple-icons (the black-and-white style matches perfectly), and save the original svg into the module directory for reference; the same brand may only look one way in the same video