@optima-chat/optima-agent 0.9.15 → 0.9.17

package/.claude/skills/scout/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: scout
-description: "Interactive Amazon product research consultant with 1688 supplier sourcing and Keepa historical data. Search products, analyze markets, find suppliers (by keyword or image), track price/BSR history, and generate professional reports. ALWAYS use when user asks about: selling products (卖/开店), product ideas (产品/商品), market research (市场/竞品), suppliers (供应商/厂家), sourcing (采购/进货/货源), image search (图片找货/以图搜货/找同款), evaluating business opportunities (生意/创业), product trends (趋势/走势), price history (价格历史/价格变化), BSR ranking (排名/销量排名), product lifecycle (生命周期/增长期/衰退期). Uses 'scout search', 'scout product', 'scout keepa *', 'scout supplier-search', and 'scout supplier-search-image' commands."
+description: "Interactive Amazon product research consultant with 1688 supplier sourcing and Keepa historical data. Search products, analyze markets, deep-dive reviews for pain points, find suppliers (by keyword or image), track price/BSR history, and generate professional reports. ALWAYS use when user asks about: selling products (卖/开店), product ideas (产品/商品), market research (市场/竞品), reviews/pain points (评论/差评/痛点/用户反馈), suppliers (供应商/厂家), sourcing (采购/进货/货源), image search (图片找货/以图搜货/找同款), evaluating business opportunities (生意/创业), product trends (趋势/走势), price history (价格历史/价格变化), BSR ranking (排名/销量排名), product lifecycle (生命周期/增长期/衰退期). Uses 'scout search', 'scout product', 'scout reviews', 'scout keepa *', 'scout supplier-search', and 'scout supplier-search-image' commands."
 ---
 
 # Optima Scout - AI Product Research
@@ -139,7 +139,43 @@ scout product ASIN
 - **Images count** - 7 is standard, fewer = poor optimization
 - **Specifications** - May include BSR in multiple categories
 
-### 3. Search 1688 Suppliers (Keyword): `scout supplier-search`
+### 3. Get Product Reviews: `scout reviews`
+
+Fetch Amazon product reviews with pagination. Use for deep pain-point analysis beyond the 8-13 top_reviews returned by `scout product`.
+
+**Syntax:**
+```bash
+scout reviews ASIN                              # 默认 3 页 (~30 条)
+scout reviews ASIN --pages 10                   # 最多 10 页 (~100 条)
+scout reviews ASIN --star one_star              # 只看 1 星差评
+scout reviews ASIN --all-stars --pages 5        # 每个星级各 5 页 (~250 条)
+scout reviews ASIN --verified                   # 只看 verified purchase
+scout reviews ASIN --sort helpful               # 按 helpful 排序
+```
+
+**Parameters:**
+- `ASIN`: 10-character Amazon product ID
+- `--pages`: Number of pages per request (1-10, default: 3, ~10 reviews/page)
+- `--sort`: `recent` (default) or `helpful`
+- `--star`: Filter by star rating (`one_star` ~ `five_star`)
+- `--all-stars`: Fetch all 5 star ratings in one batch (max ~500 reviews)
+- `--verified`: Only verified purchase reviews
+- `--domain`: Amazon domain code (default: com)
+- `--format`: `json` (default) or `text`
+
+**Returns:**
+- Review list: id, title, body, rating, date, verified, helpful_votes, images
+- Product summary: title, rating, total_ratings, rating_breakdown (percentages)
+
+**When to use:**
+- `scout product` gives 8-13 top_reviews — enough for quick scan
+- `scout reviews` gives 30-500 reviews — use for **deep pain-point analysis**, sentiment clustering, and improvement opportunity identification
+- **痛点分析**：`scout reviews ASIN --star one_star --pages 5` 拉差评做聚类
+- **竞品评论对比**：分别拉两个竞品的差评，找差异化机会
+
+**Cost awareness:** Each page costs ~$0.0075 Apify credits. `--all-stars --pages 10` (~500 reviews) costs ~$0.38.
+
+### 4. Search 1688 Suppliers (Keyword): `scout supplier-search`
 
 Search for suppliers on 1688.com using keywords.
 
@@ -165,7 +201,7 @@ scout supplier-search "KEYWORD" --limit NUMBER
 - Years in business (8+ years = established)
 - Return rate (lower = better, typical: 30-50%)
 
-### 4. Search 1688 Suppliers (Image): `scout supplier-search-image`
+### 5. Search 1688 Suppliers (Image): `scout supplier-search-image`
 
 Search for suppliers on 1688.com using product images - perfect for finding sources for products you see on Amazon, Shopify, or other platforms.
 
@@ -293,37 +329,40 @@ scout keepa tokens
 
 **Phase 2: Deep Dive & Refinement**
 4. **Targeted Analysis** → Based on user feedback, run `scout product` on relevant ASINs
-5. **Historical Trend Analysis** → Run `scout keepa product ASIN1,ASIN2,ASIN3 --stats 180` to get historical data
-6. **Extract Insights** → Combine realtime + historical data:
+5. **Deep Review Analysis** → Run `scout reviews ASIN --star one_star --pages 5` to extract pain points, then `scout reviews ASIN --sort helpful --pages 3` to see most impactful reviews
+6. **Historical Trend Analysis** → Run `scout keepa product ASIN1,ASIN2,ASIN3 --stats 180` to get historical data
+7. **Extract Insights** → Combine realtime + historical + review data:
    - BSR trend → 增长期/成熟期/衰退期？
    - `salesRankDrops30/90` → 估算月销量
    - 价格历史 → 利润空间稳定吗？有价格战吗？
    - 卖家数量变化 → 竞争在加剧吗？
-7. **Iterate & Explore** → User may request different keywords, specific competitor analysis, or alternative product categories
+   - 差评聚类 → 用户核心痛点是什么？能改进吗？
+   - 好评亮点 → 用户最看重什么功能？
+8. **Iterate & Explore** → User may request different keywords, specific competitor analysis, or alternative product categories
 
 **Phase 2.5: Supplier Sourcing (Optional but Recommended)**
-8. **Trigger**: User shows interest in costs/sourcing, or after market validation
+9. **Trigger**: User shows interest in costs/sourcing, or after market validation
    - "成本大概多少？" "批发价怎么样？" "利润空间大吗？"
    - "哪里进货？" "供应商怎么找？"
    - "这个图片能找到货源吗？" "Amazon这个能在1688找到吗？"
 
-9. **Choose search method**:
+10. **Choose search method**:
    - **With image**: `scout supplier-search-image "https://..."` ← More accurate for visual products
    - **With keyword**: `scout supplier-search "产品关键词"`
    - **Best practice**: Try both and compare results
 
-10. **Share Findings**:
+11. **Share Findings**:
     - Wholesale price range (e.g., ¥429-599)
     - MOQ (minimum order quantity)
     - Profit margin calculation: Amazon price ($60-80) - 1688 cost (¥429 ≈ $60) - shipping ($10-15)
     - Supplier quality indicators (years in business, certifications)
 
-11. **Ask**: "要不要我详细分析几个供应商？" or proceed to report
+12. **Ask**: "要不要我详细分析几个供应商？" or proceed to report
 
 **Phase 3: Report Generation (Proactive Suggestion)**
-12. **Gauge Completion** → After sufficient data, suggest: "要不要我整理成一份完整的市场分析报告？"
-13. **Generate Report** → Create comprehensive markdown report with all findings
-14. **Save & Iterate** → Offer to save report and refine based on feedback
+13. **Gauge Completion** → After sufficient data, suggest: "要不要我整理成一份完整的市场分析报告？"
+14. **Generate Report** → Create comprehensive markdown report with all findings
+15. **Save & Iterate** → Offer to save report and refine based on feedback
 
 ---
 
@@ -332,9 +371,10 @@ scout keepa tokens
 When user explicitly requests a report upfront:
 1. Run 3-4 diverse keyword searches
 2. Analyze top 3-5 products with `scout product`
-3. Pull historical trends with `scout keepa product` (batch ASINs)
-4. Search 1688 suppliers (image + keyword)
-5. Generate comprehensive report directly
+3. Deep review analysis: `scout reviews ASIN --star one_star --pages 5` on top competitors
+4. Pull historical trends with `scout keepa product` (batch ASINs)
+5. Search 1688 suppliers (image + keyword)
+6. Generate comprehensive report directly
 
 ### Multi-Keyword Research Strategy
 
@@ -420,10 +460,10 @@ scout product B0788F3R8X
   - Product description completeness
   - A+ content presence
 - ✅ **User Feedback Deep-Dive**
-  - Top 3 positive mentions (from top_reviews)
-  - Top 3 complaints/pain points (from top_reviews)
-  - Improvement opportunities
-  - Common feature requests
+  - Quick scan: `scout product` top_reviews (8-13 条精选)
+  - Deep dive: `scout reviews ASIN --star one_star --pages 5` (差评痛点聚类)
+  - Top 3 positive mentions and top 3 complaints
+  - Improvement opportunities and common feature requests
 - ✅ **Competitive Advantages**
   - Unique selling points in feature_bullets
   - Price positioning vs. competitors

package/.claude/skills/video-edit/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: video-edit
-description: "剪辑用户【已有的】口播视频——去除卡顿停顿让视频更流畅紧凑。
+description: "剪辑用户【已有的】口播视频——去除卡顿停顿、NG 重拍、口头禅让视频更流畅紧凑。
 
   必备前提：用户已经有视频文件（拍好的、上传的、或给出文件路径）。
 
@@ -26,30 +26,50 @@ description: "剪辑用户【已有的】口播视频——去除卡顿停顿让
 只有当用户提到平台但**没说目标时长**时（如"剪成 TikTok"无时长），才追问"目标多长？"。
 其他情况**直接动手**，不要废话。
 
-## 核心原则：一键出片，不满意再调
+## 核心原则：你是语义剪辑器
 
-用户说"剪一下"，你**直接出成片**，不要中途停下来问确认。像开拍 app 一样：上传 → 自动出片 → 不满意再调。
+cut 工具**只能删静音**（>=0.25s 的停顿），它**永远删不了**这些：
+- NG 重拍（重新说一遍同一句）
+- 重复字（"我我我"，发声不停顿）
+- 口头禅（"嗯"/"啊"/"那个"/"就是"）
+- 跑题片段（语法完整但内容废）
 
-**不满意时的调整**：用户看完成片说"第X段不该删"/"XX那句话留着"，你编辑 `cut_proposal.md`（去掉/加上 `#` 注释），重跑 `cut` + `subtitle`，几秒出新版。
+**所以你必须自己审一遍 proposal**——你就是这个流程的语义剪辑器。如果跳过审片直接出片，成片会有大量 NG/口头禅/重复，用户会觉得"非常多卡顿"。
 
 ## 你的内部流程
 
 ### 默认剪辑（剪 + 字幕，最常见）
 
-一步到位，不要分步等确认：
+四步走：
+
+```bash
+video-edit analyze <video>
+```
+
+读 `<video>.work/cut_proposal.md`。每行是一个短语，看每行决定保留还是删除：
+
+- **删（行首加 `#`）**：
+  - 单独的口头禅 phrase："嗯"/"啊"/"呃"/"那个"/"就是"/"对吧"/"然后那个"
+  - NG 重拍：相邻两行说同一件事但措辞稍变（"我们要做的就是——我们要做的是"），**留最后一遍**（通常最干净）
+  - 字面重复："我我我"/"那那那"
+  - 跑题/废话：与主题无关的旁白、自语、笑场
+  - 太短的连接词单独成行（如只有"那"/"嗯啊"）
+- **留**：所有有信息量的句子
+
+按上述规则编辑 proposal，保存。然后：
+
 ```bash
-video-edit both <video>
+video-edit cut <video>
 video-edit subtitle <video>
 ```
 
-`both` = analyze + cut，一口气完成转写、静音检测、生成 proposal、执行剪辑。
-`subtitle` 从原始转写推算字幕时间轴（不会重新调 whisper），烧录输出 `<原片>_subbed.mp4`。
+`cut` 按你标注的 proposal 剪辑，`subtitle` 对成片重新转写并烧字幕。
 
 ### 用户要求调整时
 
 用户看完成片说某段不对：
-1. 读 `<video>.work/cut_proposal.md`
-2. 编辑对应行（`#` 注释 = 删除，去掉 `#` = 保留）
+1. 重新读 `<video>.work/cut_proposal.md`
+2. 调整对应行（`#` 注释 = 删除，去掉 `#` = 保留）
 3. 删除 `<video>.work/subs.ass`（让 subtitle 重新生成）
 4. 重跑：
 ```bash
@@ -61,22 +81,10 @@ video-edit subtitle <video>
 
 ### 限时长精剪（用户给了目标时长）
 
-```bash
-video-edit analyze <video>
-```
-
-读 `<video>.work/cut_proposal.md`，**你自己**做这些判断：
+`analyze` → 审 proposal 时**额外做这些**：
 - 找钩子句放最前
-- 删跑题、删 NG、删填充词
-- 压到目标时长
-- 把要删的行首加 `#`，保存
-- 删除 `<video>.work/subs.ass`
-
-然后：
-```bash
-video-edit cut <video>
-video-edit subtitle <video>
-```
+- 压到目标时长（更激进地删）
+- 其余流程同上
 
 ### 仅字幕（用户明确说不剪）
 
@@ -89,11 +97,24 @@ video-edit subtitle <video>
 ### 无字幕版（用户明确说不要字幕）
 
 ```bash
-video-edit both <video>
+video-edit analyze <video>
+# 审 proposal 标 #
+video-edit cut <video>
 ```
 
 交付 `<原片>_edited.mp4`，不跑 subtitle。
 
+## 审片心法
+
+读 proposal 时**不要逐行纠结**——每行 1 秒判断，按这个决策树：
+
+1. 这行只有口头禅/单字 → `#`
+2. 这行跟前一行/后一行说同一件事 → 留信息量最高的那遍
+3. 这行内容跟主题无关 → `#`
+4. 其他 → 留
+
+**追求"看着不卡"，不追求 100% 完美**。删多了用户会说"这句要留"，你再调；删少了成片很卡，用户体验更差。**宁可激进一点**。
+
 ## 交付怎么说
 
 简洁，**只说结果**：
@@ -101,13 +122,13 @@ video-edit both <video>
 - ✅ "剪好了，xxx_subbed.mp4，时长 23 秒"
 - ✅ "无字幕版剪好了，xxx_edited.mp4，时长 23 秒"
 - ❌ "我用了 video-edit analyze，然后编辑了 cut_proposal.md..."
-- ❌ "review 显示 3 个 WARN，但都不影响..."
+- ❌ "审片删了 8 行 NG..."
 
 ## 不要做的事
 
-- ❌ 用户说"剪一下"就追问平台/钩子/时长——直接 `both` + `subtitle`
+- ❌ 跳过审 proposal 直接 `both` 出片——必有大量卡顿
+- ❌ 用户说"剪一下"就追问平台/钩子/时长——按默认流程动手
 - ❌ 用户说"剪一下"只交付 `_edited.mp4` 不烧字幕——**默认带字幕**
-- ❌ 剪之前停下来展示方案等确认——**直接出片**，不满意再调
 - ❌ 把命令名、cut_proposal.md 路径暴露给用户——内部实现细节
 - ❌ 编造命令（仅 5 个：both / analyze / cut / review / subtitle）
 - ❌ 用户说"剪掉前 X 秒"/"剪掉中间一段"——定点裁剪不是去卡顿，告诉用户"我擅长去卡顿停顿，定点裁剪请用别的方式"
@@ -116,8 +137,8 @@ video-edit both <video>
 
 | 命令 | 用途 |
 |---|---|
-| `video-edit both <video>` | **默认用这个**：analyze + cut 一步到位 |
-| `video-edit analyze <video>` | 仅转写 + 生成 proposal（限时长精剪时用） |
+| `video-edit analyze <video>` | 转写 + 生成 proposal（**默认入口**，审完后跑 cut） |
 | `video-edit cut <video>` | 按 proposal 剪辑，生成 `_edited.mp4` |
+| `video-edit subtitle <video>` | 对成片重新转写 + 烧字幕，输出 `<原片>_subbed.mp4` |
+| `video-edit both <video>` | analyze + cut 一步到位（**已弃用**：跳过审片质量差） |
 | `video-edit review <video>` | 诊断用：检查 `_edited.mp4`，输出 review_report.md |
-| `video-edit subtitle <video>` | 烧字幕。传**原片路径**，自动找 `_edited.mp4` 做输入，输出 `<原片>_subbed.mp4` |

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@optima-chat/optima-agent",
-  "version": "0.9.15",
+  "version": "0.9.17",
   "description": "基于 Claude Agent SDK 的电商运营 AI 助手",
   "type": "module",
   "main": "dist/src/index.js",