Subtitles and audio in,
nine B-roll segments out.

Watch first.
All nine segments play here.

This is the preview page I reviewed. It runs the same scene code Remotion used for the 4K render, so what you see is the final film. Pick a segment; tick 带上配乐音效 for the sound design and music alongside the narration, or 叠上头像和字幕 to see where my avatar and subtitles sit in the edit.

00Last episode00:00–00:40 · 40.8s 01Windows is here00:40–01:00 · 19.4s 02Daily work & website01:19–01:51 · 32.4s 03Two tasks at once02:30–03:04 · 33.5s 04Human and agent apart03:45–04:06 · 20.5s 05Results & the idea04:16–04:58 · 41.7s 06Comments & bot checks04:58–05:54 · 56.9s 07Spaces, takeover, speed05:54–06:50 · 55.8s 08Wrap-up06:50–07:40 · 49.2s

For smooth playback the preview uses MP3 audio and compressed MP4 clips; the delivered film is 3840 × 2160 at 30fps. The first load is about 30 MB. Illustrative (not recorded) frames are listed at the bottom of the preview.

I handed over four things.

Everything else had to be covered: the master only shows a teleprompter and my avatar. That came to nine segments, about 350 seconds, over 70% of the film.

Before starting, the agent used the Skill's rules to cut my recording list from six clips to three. Anything animation could explain, such as the side-by-side 'take over halfway' comparison, did not need recording.

The agent started with 纸白信号橙 (paper white and signal orange), the template I had saved earlier, and built segment 02 as a sample. My reply: 'This template is too ugly. I don't want it.'

↗

So I had it run two subagents against each other. Each built the same 19-second segment, then reviewed the other's work: three fatal problems, two things the rival did better, then another revision.

↗

↗

↗

↗

With the template chosen, four agents built the other eight segments in parallel. The first full cut arrived about half an hour later. My note: sometimes what I say does not match the picture.

The agent first read this as 'titles appear early', compared 102 on-screen text items with the subtitles and fixed a dozen. That was not my point, so I sent a screenshot:

↗

Second point: too many images, the same few screenshots cycling. Where they made no sense, build the interface and motion in HTML instead.

The rebuild did four things:

1. A line-by-line storyboard: about 230 lines across nine segments, each with its picture and whether it is real footage or HTML. Screenshots only where proof is needed (comments, view counts, source, Excel results, website, recorded clips), at most twice per film.
2. Word anchors: every icon, window and screenshot appears when a specific word is said. Times are looked up from the subtitles, 0.12s early by default, never typed by hand.
3. A shared UI kit: browser windows, Space tabs, cursor clicks, terminal typing, tables filling row by row, CAPTCHAs, token meters. All four agents used the same parts.
4. Two checks: a script compares each element's entrance with the moment its word is spoken and flags anything over 0.45s; then a line-by-line 'words | picture' contact sheet is reviewed frame by frame.

// scenes.js 头部（节选）：按 SRT 查出某个词开口的时间
function SAY(seg, word, nth) { … }
// 相对当前场景的秒数，默认比开口早 0.12 秒
function AT(sc, seg, word, off, nth) {
  return +Math.max(0, SAY(seg, word, nth) - (sc && sc.from || 0) + (off == null ? -0.12 : off)).toFixed(2);
}

// 第 01 段「把 Chrome 替换掉」这一页（节选）
// 说到“Chrome”时 Chrome 图标才出现，“替换掉”时被划掉、向左推走，ego 图标从右边顶进来。
build: function () {
  var tC = at(this, 'Chrome'), tX = at(this, '替换'), tR = at(this, '替换掉');
  return … + TT('把 <b class="hl">Chrome</b> 替换掉了', tC) +
    UI.icon('chrome', { t: tC, strike: tX, out: tR + .15, outFx: 'left', label: 'Google Chrome' }) +
    UI.icon('ego', { t: tR + .2, fx: 'right', label: 'ego (lite)' }) + …;
}

↗

$ node broll-sync-audit.mjs http://localhost:8877/ 0,1,2,3,4,5,6,7,8
-- seg0: 29 个锚定元素
-- seg1: 1 个锚定元素
-- seg2: 19 个锚定元素
-- seg3: 28 个锚定元素
-- seg4: 14 个锚定元素
-- seg5: 25 个锚定元素
-- seg6: 40 个锚定元素
-- seg7: 25 个锚定元素
-- seg8: 28 个锚定元素
NOSAY seg7 [04 速度与省 Token] 省很多 Token @54.43
（NOSAY：有入场动画、但没锚到台词的元素，逐个人工确认，节选一条）
锚定元素全部同步

In the rebuilt v4, 208 elements were anchored to words, all within 0.45s (209 in the final version). At the opening: 'using it heavily' opens tab after tab in Ego Lite; 'made it my default browser' selects ego lite in system settings; Chrome appears only when I say 'Chrome', then gets struck out and pushed aside.

↗

v4 was synced line by line. After watching it all, I said: 'Let's go with 便签复盘 (sticky audit).' The sky template had only won a 19-second contest; across nine segments I wanted graph paper, heavy black headlines and orange highlights.

Because scenes describe content, the switch was cheap. The agent added a compatibility layer on top of sticky audit so the keynote-era scenes picked up the new style. Content and timing carried over, and the audit still passed.

↗

↗

↗

↗

Only the 'Windows is here' page kept the keynote layout. After the switch I preferred the original, so its launch-page layout was ported back. That is the image at the top of this article.

The last fixes
came from my eyes.

① A screenshot didn't fill its window

The source screenshot filled only a corner of its window. It looked fine in the agent's headless checks but came up short in my browser, because a 3840px image was cropped by offset. The fix: crop it first with PIL.

↗

↗

② An icon looked AI-made

A gradient tile with a four-point sparkle: 'this looks too AI-generated.' Gradients went everywhere, and the agent icon became a terminal glyph. The rule went into the agent's long-term memory too.

↗

↗

③ Only four comments

The 'no Windows version' page showed four comments. I asked for a pile that builds up with sound. Now, from the word 'comments', ten cards drop 0.26s apart, each with a card sound, older ones darkening underneath.

↗

↗

④ The preview needed a zoom after switching

Switching to segment two broke the scale until I zoomed the browser and back. Rebuilding the stage reset its styles, including the scale; calling fit() after each build fixed it. The fix lives in the Skill's preview shell, so every future project gets it.

Once approved,
render straight through.

After I said 'render it', I did not step in again. The agent rendered nine 4K segments with Remotion, named each file after the range it covers, and added a replacement table. I dropped them onto the timeline.

3840 × 2160 30fpsMusic 16 dB under voicePeaks ≤ -18 dBFSNo narration in files

Want to try?
Start here.

npx skills add NanmiCoder/ajiang-vibe-coding-skills --skill ajiang-video-broll

1. 01

   Install the Skill and give your agent the final SRT plus the rough cut or narration audio.

2. 02

   Mark the ranges you already recorded. Only the rest needs B-roll.

3. 03

   Build one sample segment first. If the template is wrong, let two agents compete on it.

4. 04

   Then develop the other segments in parallel. Anchor every icon, window and screenshot to the word that names it, and run the sync audit.

5. 05

   Decide the template after watching the whole film. Pre-crop screenshots and skip gradients and sparkle icons.

6. 06

   Render 4K only after approving the preview. The agent can measure loudness; you still need to listen.