ideas生成ワークフロー - Antigravityによる自動リサーチ

Phase 1: Trend Discovery（TOP 20選定）
    ↓ step1_trend_list.json (20 tools + YouTube URLs)

Phase 2: Deep Dive Research（事実抽出）×4バッチ
    ↓ step2_knowledge_base_batch_1.json
    ↓ step2_knowledge_base_batch_2.json
    ↓ step2_knowledge_base_batch_3.json
    ↓ step2_knowledge_base_batch_4.json

Phase 3: Content Expansion（コンテンツ生成）×4バッチ
    ↓ instagram_ideas_batch_1.json
    ↓ instagram_ideas_batch_2.json
    ↓ instagram_ideas_batch_3.json
    ↓ instagram_ideas_batch_4.json
    ↓
    → 統合 → final_ideas.json → GASでideasシートにインポート

これよりInstagram自動投稿の環境構築を行います。
以下の内容で、`.agent/workflows/instagram_post_generation.md` というファイルを新規作成してください。

# Instagram Content Generation Workflow

This document defines the end-to-end workflow for generating high-quality Instagram Reels content about AI trends.

## Phase 1: Trend Discovery
**Objective:** Identify the Top 20 AI tools for the target month (e.g., December 2025) and gather video sources.

1.  **Select Tools:** Identify 20 trending AI tools based on social signals (YouTube, X).
2.  **Gather Video URLs:**
    *   Target detailed "Review" or "Tutorial" videos (10min+).
    *   **Fallback Strategy:** If a direct URL is unavailable (or for future simulation), use the format: `SEARCH: [Video Title Query]`.
3.  **Output:** `step1_trend_list.json`
    ```json
    {
      "month": "2025-12",
      "trends": [
        { "rank": 1, "tool_name": "ChatGPT Enterprise v5", "target_video_url": "SEARCH: ChatGPT v5 Full Tutorial" }
      ]
    }
    ```

## Phase 2: Deep Dive Research
**Objective:** Extract factual information, specs, and metrics for each tool.

1.  **Batch Processing:** Process tools in batches of 5 to prevent timeouts.
2.  **Extraction Logic:**
    *   Use Browser Subagent (or Search Web as fallback) to find capabilities, pros/cons, and metrics.
    *   **Validation:** Verify specific numbers (e.g., adoption rates) to avoid hallucinations.
3.  **Output:** `step2_knowledge_base_batch_X.json`
    ```json
    [
      {
        "tool_name": "ChatGPT Enterprise v5",
        "official_name": "...",
        "capabilities": [...],
        "pros": [...],
        "cons": [...],
        "metrics": ["92% of Fortune 500 companies", "1.1% hallucination rate"],
        "quotes": [...]
      }
    ]
    ```

## Phase 3: Content Expansion (Idea Generation)
**Objective:** Generate Instagram post ideas based strictly on Phase 2 data.

### 1. Input Source
- `step2_knowledge_base_batch_X.json`

### 2. Output Schema (Strict 17 Fields)
Each post idea MUST be a JSON object with EXACTLY these fields. No extra fields.

| Field | Type | Description |
|-------|------|-------------|
| `idea_id` | String | Unique ID (e.g., "IDEA-001"). |
| `month` | String | Target month (e.g., "2025-12"). |
| `title` | String | Catchy title for the post. |
| `category` | String | "A" (Versus), "B" (New Tech), "C" (How-to), "D" (Warning), "E" (Ranking). |
| `narration_1` | String | First half of the script/narration (approx 150 chars). |
| `narration_2` | String | Second half of the script/narration (approx 150 chars). |
| `thumb_main` | String | Main text for thumbnail image. |
| `thumb_sub` | String | Sub text/copy for thumbnail image. |
| `set_1` | String | Text for Slide 1 (Intro/Hook). |
| `set_2` | String | Text for Slide 2 (Point 1). |
| `set_3` | String | Text for Slide 3 (Point 2). |
| `set_4` | String | Text for Slide 4 (Point 3). |
| `set_5` | String | Text for Slide 5 (Conclusion). |
| `set_6` | String | Text for Slide 6 (CTA). |
| `main_tool` | String | The primary tool name discussed. |
| `status` | String | Always "draft". |
| `created_at` | String | Current ISO date (e.g. "2025-12-10"). |

### 3. Volume & Strategy
- **Goal:** Generate ~4-5 ideas PER TOOL (Total ~80+ ideas for 20 tools).
- **Strategy:**
    - **Category A (Versus):** Compare with a rival (e.g., ChatGPT vs Claude).
    - **Category B (New Tech):** Highlight a specific killer feature.
    - **Category C (How-to):** Step-by-step usage guide.
    - **Category D (Warning):** Discuss limitations or risks.
    - **Category E (Ranking):** Create a ranking post for the batch.

### 4. Tone Guidelines
- Benefit-oriented.
- Use explicit line breaks `\n` in slide text.
- **Forbidden Words:** "神", "魔法", "革命", "最強" (unless in "Ranking" context), "結論" (use sparsely).

### 5. JSON Example
```json
{
  "idea_id": "IDEA-001",
  "month": "2025-12",
  "title": "ChatGPT v5 vs Claude 3.5: エンジニアが選ぶべきは？",
  "category": "A",
  "narration_1": "2025年12月、AIの頂上決戦。ChatGPT Enterprise v5とClaude 3.5 Sonnet、どっちを使うべき？結論から言うと…",
  "narration_2": "推論と分析ならChatGPT。一方、実務と操作ならClaude。Computer UseでPC操作を代行。あなたの仕事に合わせて選びましょう。",
  "thumb_main": "ChatGPT v5 vs Claude 3.5",
  "thumb_sub": "2025年、どっち選ぶ？",
  "set_1": "【徹底比較】\nChatGPT v5 vs Claude 3.5\nエンジニアの正解は？",
  "set_2": "【ChatGPT v5】\n推論力と分析力が圧倒的。\nセキュリティも万全。",
  "set_3": "【Claude 3.5】\nComputer Use機能搭載。\nPC画面を直接操作可能。",
  "set_4": "【使い分け】\n脳を使う仕事はGPT。\n手を使う仕事はClaude。",
  "set_5": "【結論】\n両方使うのが\n最強のビジネスマン。",
  "set_6": "あなたのメインはどっち？\nコメントで投票して！",
  "main_tool": "ChatGPT Enterprise v5",
  "status": "draft",
  "created_at": "2025-12-10"
}
```

{
  "idea_id": "IDEA-101",
  "month": "2025-12",
  "title": "ChatGPT v5 vs Claude 3.5: 2025年の覇権争い",
  "category": "A",
  "narration_1": "2025年、AIの頂上決戦。ChatGPT Enterprise v5とClaude 3.5 Sonnet、どっちを使うべき？結論から言うと「役割」が違います。",
  "narration_2": "推論と分析ならChatGPT。圧倒的なコンテキストとデータ分析力。一方、実務と操作ならClaude。Computer UseでPC操作を代行。あなたの仕事に合わせて選びましょう。",
  "thumb_main": "ChatGPT v5 vs Claude 3.5",
  "thumb_sub": "2025年、どっち選ぶ？",
  "set_1": "【徹底比較】\nChatGPT v5 vs Claude 3.5\nエンジニアの正解は？",
  "set_2": "【ChatGPT v5】\n推論力と分析力が最強。\n幻覚率1.1%の信頼性。\n分析業務ならこれ。",
  "set_3": "【Claude 3.5】\nComputer Use機能搭載。\nPC画面を直接操作可能。\n自動化ならこれ。",
  "set_4": "【使い分けの鍵】\n脳を使う仕事はGPT。\n手を使う仕事はClaude。",
  "set_5": "【結論】\n両方使うのが\n最強のビジネスマン。",
  "set_6": "あなたのメインはどっち？\nコメントで投票して！",
  "main_tool": "ChatGPT Enterprise v5",
  "status": "draft",
  "created_at": "2025-12-10"
}

[
  {
    "idea_id": "IDEA-001",
    "month": "2025-12",
    "title": "Title with strong hook",
    "category": "A",
    "narration_1": "First half of script (150-175 chars)",
    "narration_2": "Second half of script (150-175 chars)",
    "thumb_main": "Short trigger text (Max 8 chars)",
    "thumb_sub": "Sub text (Max 6 chars x 3 lines)",
    "set_1": "Slide 1 text (Max 14 chars x 3 lines)",
    "set_2": "Slide 2 text",
    "set_3": "Slide 3 text",
    "set_4": "Slide 4 text",
    "set_5": "Slide 5 text",
    "set_6": "Slide 6 text",
    "main_tool": "[Tool Name]",
    "status": "NEW",
    "created_at": "YYYY-MM-DD"
  }
]

The video has no transcript. Instead, read the video description and the top 10 comments to extract feature names and user opinions.

No YouTube video found. Use the official documentation URL instead: [URL]