据 AlphaMoat 统计,截至 2026-09,UGC Campaign Pipeline 累计下载 2,240 次、获 3 个星标,在全部 5,078 个 AI 技能中排名第 4,262 位,在「内容创作」分类(362 个)中排名第 283 位。
| 平台 | 累计安装 | 累计下载 | 星标 | 评论 |
|---|---|---|---|---|
| ClawHub ↗ | 61 | 2,205 | 2 | - |
| SkillHub ↗ | 168 | 2,240 | 3 | — |
各平台的计数口径不同,统计时间不一致。
该流程涵盖主视觉图、图片变体、素材筛选、视频生成、剪辑成片等步骤,目标输出为含场景、字幕和品牌标识的 MP4。流程会调用引用的外部图像或视频生成工具,产品图片、品牌素材、提示词和脚本可能发送给这些服务;中间与成品文件也可能保留在本地目录。 流程需另行具备并配置morpheus-fashion-design、multishot-ugc、veed-ugc与Remotion;开始前须审核各工具设置。
资料发布于 2026-10-27
以下是根据技能文档生成的示例,不代表实际运行结果。
2026-09 起统计方法调整:同名技能按作者分别统计,并覆盖全部技能。图中虚线为调整前的月份,与之后的数据不直接比较,也不计算跨越调整的增长。
| 月份 | 累计下载 | 环比 | 星标 |
|---|---|---|---|
| 2026-09 | 2,240 | — | 3 |
| 2026-08 | 2,095 | +8.8% | 3 |
| 2026-07 | 1,926 | +19.5% | 3 |
| 2026-06 | 1,612 | +12.3% | 2 |
| 2026-05 | 1,435 | — | 2 |
| 2026-03 | 1,079 | — | 2 |
以下是作者发布的原始文件,按纯文本显示,本站未做修改或审核;其中的链接不可点击,命令请核实后再运行。抓取于 2026-10-23。
--- name: ugc-campaign-pipeline description: | Complete UGC video campaign pipeline: product → hero image → variations → videos → edited final. ✅ USE WHEN: - User says "crear campaña UGC" or "pipeline completo" - Need end-to-end UGC video production - Starting from product image/URL → final edited video - Want the full Doritos-style workflow
❌ DON'T USE WHEN:
- Just need one step (use individual skills)
- Already have final videos, just editing → use Remotion
- Only need images, no video → use Morpheus only
OUTPUT: Edited MP4 video with multiple scenes + subtitles + logo
---
# UGC Campaign Pipeline
Complete workflow for creating UGC-style promotional videos from a product.
## Pipeline Overview
```
┌─────────────────────────────────────────────────────────────────────┐
│ UGC CAMPAIGN PIPELINE │
├─────────────────────────────────────────────────────────────────────┤
│ │
│ STEP 1: HERO IMAGE │
│ ├─ Input: Product image + model selection │
│ ├─ Tool: morpheus-fashion-design │
│ └─ Output: ~/clawd/outputs/{project}/morpheus/hero.png │
│ │
│ STEP 2: VARIATIONS │
│ ├─ Input: Hero image │
│ ├─ Tool: multishot-ugc │
│ └─ Output: ~/clawd/outputs/{project}/multishot/*.png (10 images) │
│ │
│ STEP 3: SELECTION │
│ ├─ Analyze all 11 images │
│ ├─ Criteria: variety, no errors, lip-sync friendly │
│ └─ Output: 4 best images selected │
│ │
│ STEP 4: SCRIPT │
│ ├─ Write 4-scene dialogue script │
│ ├─ Format: PURE DIALOGUE (no annotations) │
│ └─ Output: 4 lines of dialogue │
│ │
│ STEP 5: UGC VIDEOS │
│ ├─ Input: 4 images + 4 script lines │
│ ├─ Tool: veed-ugc (run 4 times) │
│ └─ Output: ~/clawd/outputs/{project}/ugc/*.mp4 (4 videos) │
│ │
│ STEP 6: FINAL EDIT │
│ ├─ Input: 4 videos + logo │
│ ├─ Tool: Remotion │
│ ├─ Add: subtitles, transitions, logo ending │
│ └─ Output: ~/clawd/outputs/{project}/final/video.mp4 │
│ │
└─────────────────────────────────────────────────────────────────────┘
```
## Execution Checklist
### Before Starting
- [ ] Product image received
- [ ] Brand/product name known
- [ ] Target audience understood
- [ ] Tone defined (casual, professional, energetic)
### Step 1: Hero Image (Morpheus)
```bash
# Select model from catalog
cat ~/clawd/models-catalog/catalog/catalog.json | jq '[.talents[] | select(.gender == "male/female") | {id, name, ethnicity}]'
# Generate hero image
COMFY_DEPLOY_API_KEY="..." uv run ~/.clawdbot/skills/morpheus-fashion-design/scripts/generate.py \
--product "product.jpg" \
--model "models-catalog/catalog/images/model_XX.jpg" \
--brief "..." \
--target "..." \
--aspect-ratio "9:16" \
--output "~/clawd/outputs/{project}/morpheus/hero.png"
```
### Step 2: Variations (Multishot)
```bash
COMFY_DEPLOY_API_KEY="..." uv run ~/.clawdbot/skills/multishot-ugc/scripts/generate.py \
--image "~/clawd/outputs/{project}/morpheus/hero.png" \
--output-dir "~/clawd/outputs/{project}/multishot" \
--resolution "2K" \
--aspect-ratio "9:16"
```
### Step 3: Selection Criteria
Analyze all 11 images for:
- ✅ Face clearly visible (frontal or 3/4)
- ✅ Mouth not obscured
- ✅ No distorted hands/fingers
- ✅ Product visible
- ✅ Different from other selected images
Select 4 images that are:
1. Different angles/perspectives
2. Suitable for lip-sync
3. Error-free
### Step 4: Script Writing
Write 4 lines of pure dialogue:
```
ESCENA 1: [Opening hook - grab attention]
ESCENA 2: [Problem/benefit - relate to audience]
ESCENA 3: [Feature highlight - specific value]
ESCENA 4: [CTA/brand mention - close strong]
```
**Rules:**
- Pure dialogue only
- NO annotations: [excited], (pause), *smiles*
- NO stage directions
- NO tone indicators
- Just what the person says
### Step 5: Generate Videos (VEED)
```bash
for i in 1 2 3 4; do
COMFY_DEPLOY_API_KEY="..." uv run ~/.clawdbot/skills/veed-ugc/scripts/generate.py \
--image "selected_image_$i.png" \
--script "Script line $i" \
--output "~/clawd/outputs/{project}/ugc/escena$i.mp4"
done
```
### Step 6: Final Edit (Remotion)
Create Remotion project with:
- All 4 video clips sequenced
- Animated subtitles for each scene
- Logo animation at end
- Render to final MP4
---
## Script Templates by Industry
### Snacks/Food
```
Escena 1: Che, tenés que probar esto.
Escena 2: [Sabor/textura highlight]. Te pega de una.
Escena 3: Y mirá, no es solo [feature], tiene ese gustito que te deja queriendo más.
Escena 4: [Brand + Product]. Una vez que arrancás, no parás.
```
### Tech/Gadgets
```
Escena 1: Mirá lo que me llegó.
Escena 2: Esto cambia todo. [Key feature].
Escena 3: Y lo mejor? [Secondary benefit].
Escena 4: [Brand]. Ya no vuelvo atrás.
```
### Beauty/Skincare
```
Escena 1: Ok, necesito contarte algo.
Escena 2: Este [producto] es increíble. [Result].
Escena 3: Lo uso hace [tiempo] y mirá la diferencia.
Escena 4: [Brand]. Tu piel te lo va a agradecer.
```
### Fashion
```
Escena 1: Encontré mi nueva obsesión.
Escena 2: Mirá este [prenda]. [Quality/style].
Escena 3: Combina con todo y es súper [comfort/style].
Escena 4: [Brand]. Estilo sin esfuerzo.
```
---
## Output Structure
```
~/clawd/outputs/{project}/
├── morpheus/
│ └── hero.png # Original hero image
├── multishot/
│ ├── 1_00001_.png # 10 variations
│ ├── 2_00001_.png
│ └── ...
├── ugc/
│ ├── escena1.mp4 # Individual scene videos
│ ├── escena2.mp4
│ ├── escena3.mp4
│ └── escena4.mp4
├── assets/
│ └── logo.png # Brand logo
└── final/
└── video.mp4 # Final edited video
```
---
## Quality Checklist (Before Delivery)
- [ ] Hero image: product visible, model looks natural
- [ ] Variations: 4 selected are distinct and lip-sync ready
- [ ] Script: matches brand tone, pure dialogue format
- [ ] Videos: lip-sync quality is good, no artifacts
- [ ] Final: subtitles readable, transitions smooth, logo appears
- [ ] Audio: voice quality clear, timing natural
识别并去掉文本里的 AI 写作痕迹,让文字读起来更像人写。
把文本里 AI 写作的典型痕迹挑出来并改写得更自然,同时保留原意的写作编辑技能。
识别并去除 AI 生成文本痕迹的中文文字编辑技能。
wechat-search 检索公众号热门文章,并按相关性、热度和时效整理结果。
从公开短视频分享链接提取标题、简介和口播文案,并给出改写版本。
humanizer 为中英文文本提供更自然的表达和风格编辑指导。
写网文章节并按文件落盘,聊天里只回摘要与指针。
把一组营销方法模块收进同一个技能,按请求路由到对应模块取清单与可交付稿。
数据月份:2026-09 · 下载/星标/安装数聚合自公开技能注册库(ClawHub、SkillHub),按月更新。口径说明见数据方法。