Nby Auto Edit

AI Auto Edit — Raw Recording to Finished Video

Automate video production: from raw green-screen recordings through ASR transcription, script matching, best-take selection, precise slicing, background replacement, to Remotion rendering.

Language

Match the user's language: respond in the same language the user uses.

Configuration

Preferences (EXTEND.md)

Check existence (priority order):

test -f .nby-skills/nby-auto-edit/EXTEND.md && echo "project"
test -f "$HOME/.nby-skills/nby-auto-edit/EXTEND.md" && echo "user"

Value Priority

CLI arguments > EXTEND.md (project) > EXTEND.md (user) > Skill defaults

First-Time Setup

If no EXTEND.md exists, ask:

1. Resolution: "视频分辨率？" → 720x1280 (竖屏) / 1920x1080 (横屏)
2. TTS Voice: "默认配音声音？" → Edge TTS 声音列表，或跳过
3. Green Screen: "是否使用绿幕拍摄？" → chromakey / rembg / 不使用

Save to .nby-skills/nby-auto-edit/EXTEND.md (ask project or user level).

Prerequisites

Environment Check (Required)

command -v ffmpeg   && echo "ffmpeg: OK"   || echo "ffmpeg: MISSING → brew install ffmpeg"
command -v ffprobe  && echo "ffprobe: OK"  || echo "ffprobe: OK (included with ffmpeg)"
command -v node     && echo "node: OK"     || echo "node: MISSING → brew install node"

If any required tool is missing, print install commands and stop.

Remotion Project Check

[ -f "package.json" ] && grep -q "remotion" package.json && echo "Remotion: OK" || echo "Remotion: NOT FOUND"

If not found, guide initialization:

npx create-video@latest my-video  # Choose "Blank" template
cd my-video && npm install

Key directories:

• src/ — Component files (scenes)
• src/root.tsx — Composition registry (must register all scenes here)
• public/ — Asset files (videos, images, audio)

Optional Tools

Pipeline Overview

Phase 1: 素材准备         ← 必须
Phase 2: 语音转录         ← 可选(有剧本时间线可跳过)
Phase 3: 剧本对照+选Take   ← 核心步骤
Phase 4: 精确裁切         ← 可选(已有切片可跳过)
Phase 5: 绿幕背景替换     ← 可选
Phase 6: AI 素材生成      ← 手动
Phase 7: Remotion 合成    ← 必须
Phase 8: 多版本衍生       ← 可选

IMPORTANT: After Phase 1, auto-detect which phases are needed. Present the plan and ask for confirmation ONCE, then execute sequentially.

Phases

Phase 1: 素材准备 (Required)

Collect from user:

需要以下素材：

  1. 原始录制视频（可以多个，不同角色分开录制）← 必须
  2. 剧本/台词文本 ← 强烈建议提供
  3. 角色与视频对应关系（如：角色1=视频A，角色2=视频B）
  4. 背景图片/视频（可选，用于绿幕替换）

Register raw materials:

素材登记：
  视频1: {path} → 角色2（伊田助男）720x1280
  视频2: {path} → 角色1（审查员）544x960

剧本：
  1: 你是日本人，你为什么能上天堂啊？
  2: 你好
  1: 叫什么名字？
  ...

Phase 2: 语音转录与文字提取 (Optional)

Skip if user provides script with exact timestamps.

Critical: Use large model for word-level timestamps.

# Extract audio from each video
ffmpeg -i {video1} -vn -acodec pcm_s16le -ar 16000 /tmp/video1-audio.wav
ffmpeg -i {video2} -vn -acodec pcm_s16le -ar 16000 /tmp/video2-audio.wav

# Option A: mlx-whisper (Apple Silicon, fastest — recommended on macOS)
mlx_whisper /tmp/video1-audio.wav --model mlx-community/whisper-large-v3-mlx \
  --language zh --word_timestamps True --output_format all

# Option B: whisper (cross-platform)
whisper /tmp/video1-audio.wav --model large-v3 --language zh \
  --word_timestamps True --output_format all

# Option C: coli (if installed)
coli asr /tmp/video1-audio.wav --language zh --output-format srt

Output: Word-level transcript with classification:

| 时间 | 文字 | 类型 |
|------|------|------|
| 0.00 - 0.34 | 321走 | 打板 |
| 4.64 - 5.24 | 你好 | take1（声音小） |
| 9.56 - 10.70 | 321走 | 打板 |
| **10.70 - 11.42** | **你好** | **take2 (采用)** |
| 12.46 - 13.06 | 可以吗 | 确认 |

Classify each segment as:

• 打板: "321", "321走" — slate clap, skip
• 有效 take: Matches a script line
• NG: 口误, 不完整, "重来", "不对"
• 废话: "可以吗", 闲聊, 确认
• 静默: Gaps > 2 seconds

Key learnings:

• Small ASR models skip "silent" segments that actually have speech — always use large-v3
• Transcribe full video first, then classify — don't pre-cut
• For long recordings (>2min): split into segments first, transcribe each, then merge — more accurate
• On Apple Silicon, mlx-whisper is significantly faster than standard whisper
• Timestamp drift on long videos: Full-video ASR timestamps can drift 1-2s on recordings >1min. For critical cut points, re-transcribe just the surrounding 8-10s (ffmpeg -ss {t-2} -t 8) and use those local timestamps instead
• Long-offset seeking: When -ss is >60s, put it AFTER -i (decode-based seeking) instead of before -i (fast seek) to avoid audio-video desync

Phase 3: 剧本对照 & Take 选择 (Core Step)

This is the most critical phase — match ASR results against the script line-by-line.

3.1 Script-to-ASR matching:

For each script line, find all matching takes across all videos:

剧本第1句: "你是日本人，你为什么能上天堂啊"
  → 视频2 take1 (1.14-3.86) ✓ 完整
  → 视频2 take2 (6.80-9.28) ✗ 少了"啊"

剧本第2句: "你好"
  → 视频1 take1 (4.64-5.24) ✗ 声音小
  → 视频1 take2 (10.70-11.42) ✓ 采用

3.2 Best take selection criteria:

• Completeness: All words present, matches script
• Clarity: Clear pronunciation, no stammering
• Timing: Natural pacing
• If multiple valid takes, pick the later one (usually better after warm-up)

3.3 Determine precise cut points:

Rules:

• First word start time - 0.1s → Last word end time + 0.15s
• When a take follows a "321走" slate, the "走" end time is the anchor — cut starts AFTER "走" ends
• Same-character consecutive segments should have no gap between them

"你好" spoken at 10.70-11.42 (after "321走" ending at 10.70)
  → Cut range: 10.60 - 11.55 (0.1s before, 0.15s after)

3.4 Present the cut plan for confirmation:

裁切方案（{N} 个片段）：

| 片段 | 台词 | 来源 | 语音时间 | 裁切范围 |
|------|------|------|----------|----------|
| seg01 | 你是日本人... | 视频2 | 1.14-3.86 | 1.04-3.96 |
| seg02 | 你好 | 视频1 | 10.70-11.42 | 10.60-11.55 |
| seg03 | 你叫什么名字 | 视频2 | 12.66-14.24 | 12.56-14.40 |
| seg04 | 我叫伊田助男 | 视频1 take2 | 18.32-19.82 | 18.22-19.95 |
...

确认裁切方案？

Phase 4: 精确裁切 (Optional)

Skip if user has pre-cut segments.

# Cut each segment (use -ss before -i for fast seek, -t for duration)
ffmpeg -y -ss {cut_start} -i {source_video} -t {duration} \
  -c:v libx264 -crf 18 -preset slow \
  -c:a aac segments/seg{NN}.mp4

# Verify duration matches expected
ffprobe -v error -show_entries format=duration -of csv=p=0 segments/seg{NN}.mp4

Key rules:

• Use -t {duration} NOT -to {end_time} when -ss is before -i (otherwise cut will be wrong)
• Duration MUST match precisely — mismatches cause background flash between segments
• Single-pass encoding: When possible, combine cutting + green screen in one ffmpeg command to avoid quality loss from re-encoding
• Do NOT scale during cutting or keying — defer resolution unification to Phase 7 assembly
• Same-character consecutive segments should have no gap

Phase 5: 绿幕背景替换 (Optional)

Skip if not using green screen footage.

Method selection:

FFmpeg chromakey:

ffmpeg -y \
  -i bg.png \
  -ss {start} -to {end} -i raw_video.mp4 \
  -filter_complex "
    [0:v]scale={video_w}:{video_h}:flags=lanczos[bg];
    [1:v]chromakey=0x00FF00:similarity=0.24:blend=0.06,
    despill=type=green:mix=0.6:expand=0.4,
    eq=saturation=1.05:contrast=1.02[fg];
    [bg][fg]overlay=0:0,format=yuv420p[out]
  " \
  -map "[out]" -map 1:a \
  -c:v libx264 -preset slow -crf 16 \
  -c:a aac -b:a 192k \
  -movflags +faststart \
  output.mp4

Important: Scale background to match video's native resolution ({video_w}x{video_h}), do NOT scale the video before chromakey. Add eq=saturation=1.05:contrast=1.02 after despill to compensate color loss.

rembg AI (recommended for poor green screen):

# Extract frames
ffmpeg -i seg.mp4 -vf fps=30 frames/%04d.png
# AI remove background
rembg p frames/ frames_nobg/
# With alpha matting (slightly better hair edges):
rembg p --alpha-matting frames/ frames_nobg/
# Composite with background + reassemble

Key learnings:

• Use standard green 0x00FF00 for chromakey, NOT sampled color values
• Poor green screen → just use rembg AI, don't fight with chromakey params
• Do NOT resize during keying — keep original resolution

Phase 6: AI 素材生成 (Manual Step)

User generates assets with their preferred AI tool. Provide guidance:

Consistency rule: All scene images should use the same AI model + style prompt to maintain visual coherence.

Prompt template:

{art_style} depicting {scene_description}.
Consistent {color_palette} across all scenes.
{era/period} setting. Format: {width}x{height}.

Asset checklist:

• Background images (one per character or scene)
• Insert images (historical photos, documents, etc.)
• Character portraits (if needed for picture-in-picture)
• Title / ending cards
• BGM audio file
• Sound effects (whoosh, etc.)

Save all to public/{project-name}/.

Phase 6b: FFmpeg 快速拼接 (Optional — 替代 Remotion)

当不需要字幕、特效、转场时，直接用 ffmpeg concat 拼接片段，速度最快：

# 1. 创建文件列表
cat > /tmp/concat_list.txt << EOF
file '/path/to/segments/seg01.mp4'
file '/path/to/segments/seg02.mp4'
...
EOF

# 2. 无损拼接（要求所有片段编码参数一致）
ffmpeg -y -f concat -safe 0 -i /tmp/concat_list.txt \
  -c copy -movflags +faststart \
  output/final.mp4

前提条件：所有片段必须使用相同的编码器、分辨率、帧率。Phase 4+5 产出的片段已满足此条件。

适用场景：快速出片、测试剪辑节奏、不需要字幕/BGM/特效的粗剪版本。

Phase 7: Remotion 合成 (Required)

7a. Create component — src/scenes/{project}/{ProjectName}.tsx:

import React from 'react';
import {AbsoluteFill, Audio, Video, Sequence, staticFile,
  useCurrentFrame, useVideoConfig, interpolate, spring} from 'remotion';

const FPS = 30;

interface SubEntry {
  text: string;   // Supports **keyword** highlighting
  en?: string;    // English translation (optional)
  start: number;
  end: number;
}

interface Segment {
  file: string;
  duration: number;
  character: string;
  subs: SubEntry[];
}

// ← Fill from Phase 3 cut plan
const SEGMENTS: Segment[] = [
  {file: 'seg01.mp4', duration: 2.92, character: 'judge', subs: [
    {text: '台词内容', start: 0, end: 2.92},
  ]},
  // ... more segments
];

const segmentFrames = SEGMENTS.map(s => Math.round(s.duration * FPS));
const CONTENT = segmentFrames.reduce((a, b) => a + b, 0);
const ENDING = 3 * FPS;
export const TOTAL_FRAMES = CONTENT + ENDING;

export const MyVideo: React.FC = () => {
  const starts: number[] = [];
  let off = 0;
  for (const f of segmentFrames) { starts.push(off); off += f; }

  return (
    <AbsoluteFill>
      {/* BGM */}
      <Audio src={staticFile('{project}/bgm.mp3')} volume={0.15} loop />

      {/* Video segments */}
      {SEGMENTS.map((seg, i) => (
        <Sequence key={i} from={starts[i]} durationInFrames={segmentFrames[i]}>
          <Video src={staticFile(`{project}/${seg.file}`)}
            style={{width: '100%', height: '100%', objectFit: 'cover'}} />
          {/* Add subtitles, character tags, image inserts here */}
        </Sequence>
      ))}
    </AbsoluteFill>
  );
};

7b. Register in src/root.tsx (essential — without this, render fails):

import {Composition} from 'remotion';
import {MyVideo, TOTAL_FRAMES} from './scenes/{project}/MyVideo';

// Add inside RemotionRoot:
<Composition
  id="MyVideo"
  component={MyVideo}
  durationInFrames={TOTAL_FRAMES}
  fps={30}
  width={720}    // or 1920
  height={1280}  // or 1080
/>

7c. Place assets:

mkdir -p public/{project}
cp segments/*.mp4 public/{project}/
cp bgm.mp3 public/{project}/

7d. Creative decisions checklist:

Before rendering, review these common decisions:

• Title card needed? (If seg01 is already attention-grabbing, skip it)
• Transitions between segments? (Hard cut often has better pacing than crossfade)
• Character flash/effect on speaker change? (Usually distracting — skip)
• Sound effects frequency? (Once is impactful, repeating is annoying)
• Subtitle coloring? (White for all is cleaner — use character tags to distinguish roles)
• Same-character consecutive segments: no effects between them (keep continuity)
• BGM volume? (10-15% is typical — too loud drowns dialogue)

7e. Preview:

npx remotion studio  # Select composition in browser

7e. Render:

npx remotion render MyVideo --concurrency=1 --codec=h264 --output=out/{project}.mp4

Note: --concurrency=1 required to prevent frame jitter.

Phase 8: 多版本衍生 (Optional)

From the same base, create variants:

Each variant should be a separate component file (e.g., MyVideoEn.tsx), not modifying the original.

TTS duration alignment (critical):

TTS_DUR=$(ffprobe -v error -show_entries format=duration -of csv=p=0 tts.mp3)
SEG_DUR=$(ffprobe -v error -show_entries format=duration -of csv=p=0 seg.mp4)
RATIO=$(python3 -c "print(${TTS_DUR}/${SEG_DUR})")

# Speed up/slow down (atempo range 0.5~2.0)
ffmpeg -y -i tts.mp3 -af "atempo=${RATIO}" -ar 24000 tts_aligned.mp3

# If TTS shorter, pad with silence instead
ffmpeg -y -i tts.mp3 -af "apad=whole_dur=${SEG_DUR}" -ar 24000 tts_aligned.mp3

Error Handling

Self-Evolution

After each run, save to .auto-edit/{project}/config.json:

{
  "project": "my-project",
  "segments": 10,
  "sources": {"video1": "角色A", "video2": "角色B"},
  "greenScreenMethod": "rembg",
  "renderConfig": {"concurrency": 1, "codec": "h264", "resolution": "720x1280"},
  "lessonsLearned": []
}

• EXTEND.md = user preferences (global, persists across projects)
• .auto-edit/{project}/ = per-project runtime state

Next run for same project loads config. For new: "是否基于已有项目配置？"

Remotion Reference

See references/remotion-guide.md for spring configs, reusable components, and project structure.

Composability

• Invokes: asr (transcription), tts (narration), image-gen (assets)
• Invoked by: User directly or via chat-based automation

Examples

Full pipeline (multiple green-screen recordings + script):

"我有2段绿幕视频和剧本，帮我自动剪辑" → Phase 1~7: ASR → Script match → Cut → Green screen → Compose → Render

Pre-cut segments:

"我已经有切好的片段了，帮我合成视频" → Skip Phase 2~4, jump to Phase 7

Script only (no recording):

"帮我用这个剧本生成一个视频" → Phase 1 → Phase 6 (guide asset creation) → Phase 7

Add variant:

"帮我做一个英文版" → Phase 8: Create separate component with English TTS + subtitles