Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →多智能体协同的商品视频创作流水线，支持从商品信息到成视频的全流程创作：文案生成、故事策划、脚本创作、分镜设计、图片生成、字幕创作、音效推荐及视频合成；适用于电商商品宣传、品牌推广、社交媒体营销等场景
.claude/skills/anbeime-product-video-creator/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-12 | ✗→✓ | ▲ Improved | 62% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 103% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 47% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 30% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -1% | 0% |
moviepy>=1.0.3 pillow>=10.0.0 opencv-python>=4.8.0
bgm/：背景音乐文件夹（可选，脚本会推荐也可使用默认）product_images/：商品原始图片（可选，智能体可基于描述生成）职责：分析商品信息，生成吸引目标受众的核心文案
执行要点：
输出：核心文案文本
职责：基于文案设计连贯的场景故事线，让商品卖点自然融入
执行要点：
输出：场景故事大纲（含每个场景的情节和卖点关联）
职责：将故事转化为具体的分镜脚本
执行要点：
输出：分镜脚本表格，格式参考 references/storyboard_format.md
职责：为每个分镜生成详细的画面描述，指导图片生成
执行要点：
输出：每个镜头的画面描述词列表
职责：为每个分镜生成符合描述的图片
执行要点：
输出：图片序列，命名为 scene_001.jpg, scene_002.jpg, ...
职责：为每个镜头创作字幕文本，并规划显示位置和时间
执行要点：
输出：字幕文本与时间配置，格式参考 references/subtitle_format.md
职责：推荐背景音乐和音效方案，提升视频感染力
执行要点：
输出：音频方案说明文档
职责：调用脚本将图片、字幕、音频合成为最终视频
执行要点：
scene_*.jpgsubtitles.jsonbgm.mp3（或其他格式）调用脚本：
bashpython scripts/video_composer.py \ --images ./scene_*.jpg \ --subtitles ./subtitles.json \ --audio ./bgm.mp3 \ --output ./final_video.mp4 \ --fps 24 \ --duration 30
输出：最终视频文件 final_video.mp4
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-12 | fail→pass | 14,867 | 9,615 | -35% | 1 | 1 | 0% | 2,460 | 3,973 | +62% | 0 | 0 | — |
case-01 | fail→fail | 23,555 | 21,477 | -9% | 1 | 1 | 0% | 4,621 | 6,706 | +45% | 0 | 0 | — |
case-02 | fail→pass | 25,839 | 27,421 | +6% | 1 | 1 | 0% | 3,784 | 7,668 | +103% | 0 | 0 | — |
case-03 | fail→pass | 22,503 | 24,292 | +8% | 1 | 1 | 0% | 4,508 | 6,632 | +47% | 0 | 0 | — |
case-04 | fail→fail | 17,121 | 11,997 | -30% | 1 | 1 | 0% | 2,902 | 4,451 | +53% | 0 | 0 | — |
case-05 | pass→pass | 20,943 | 14,537 | -31% | 1 | 1 | 0% | 2,767 | 4,435 | +60% | 0 | 0 | — |
case-06 | pass→pass | 15,141 | 10,660 | -30% | 1 | 1 | 0% | 2,400 | 3,936 | +64% | 0 | 0 | — |
case-07 | pass→pass | 14,200 | 10,891 | -23% | 1 | 1 | 0% | 2,361 | 4,135 | +75% | 0 | 0 | — |
case-08 | fail→pass | 17,034 | 8,480 | -50% | 1 | 1 | 0% | 3,160 | 4,093 | +30% | 0 | 0 | — |
case-09 | pass→pass | 19,930 | 21,163 | +6% | 1 | 1 | 0% | 3,794 | 5,786 | +53% | 0 | 0 | — |
case-10 | pass→pass | 17,305 | 18,003 | +4% | 1 | 1 | 0% | 2,893 | 4,515 | +56% | 0 | 0 | — |
case-11 | fail→pass | 15,832 | 3,418 | -78% | 1 | 1 | 0% | 3,204 | 3,174 | -1% | 0 | 0 | — |
case-13 | pass→fail | 22,389 | 13,285 | -41% | 1 | 1 | 0% | 3,715 | 4,449 | +20% | 0 | 0 | — |
case-14 | pass→pass | 14,066 | 5,282 | -62% | 1 | 1 | 0% | 2,584 | 3,475 | +34% | 0 | 0 | — |
case-15 | fail→pass | 15,645 | 7,393 | -53% | 1 | 1 | 0% | 2,620 | 3,931 | +50% | 0 | 0 | — |
case-16 | pass→pass | 15,197 | 8,575 | -44% | 1 | 1 | 0% | 2,576 | 3,791 | +47% | 0 | 0 | — |
case-17 | pass→pass | 18,446 | 14,717 | -20% | 1 | 1 | 0% | 2,736 | 4,673 | +71% | 0 | 0 | — |
case-18 | fail→pass | 15,659 | 5,947 | -62% | 1 | 1 | 0% | 2,607 | 3,455 | +33% | 0 | 0 | — |
case-19 | fail→pass | 18,434 | 5,772 | -69% | 1 | 1 | 0% | 3,031 | 3,568 | +18% | 0 | 0 | — |
case-20 | fail→fail | 20,230 | 19,186 | -5% | 1 | 1 | 0% | 4,118 | 6,315 | +53% | 0 | 0 | — |
case-21 | fail→fail | 6,901 | 7,724 | +12% | 1 | 1 | 0% | 1,441 | 3,634 | +152% | 0 | 0 | — |
case-22 | fail→fail | 20,443 | 12,672 | -38% | 1 | 1 | 0% | 3,798 | 4,825 | +27% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +32 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.