X Collect

Collect trending topics and research materials for X content creation using a systematic 4-round web search strategy.

Prerequisites

WebSearch tool available
Internet connection

Optional State (ContextStore)

If available, leverage user preference state to improve search quality:

State dir: ~/.claude/skills/x-create/state/
Positive samples: liked_topics.json (preferred topics/entities/sources)
Negative samples: rejected_topics.json (avoid low-value/repetitive topics)

If state files don't exist, proceed normally.

Workflow

Input

User provides:

Topic (required): The subject to research (e.g., "AI Agent最新进展", "Claude 4 发布")
Language (optional): Output language, defaults to Chinese (zh-CN)

4-Round Search Strategy

Simulate human research thinking process with progressive depth.

Query expansion rules (lightweight):

Start with user topic
Add 2-5 query variants based on:
- User profile (domains/persona if known)
- liked_topics.json high-frequency keywords/entities (if available)
Keep ~20% queries as exploration (cross-domain / adjacent keywords) to avoid information bubble

Dedup & clustering (at the end of each round):

Merge near-duplicate URLs/titles
Group materials into topic clusters (rough tags are fine)

Round 1: Official Sources (权威信息)

Search: "{topic} 官方文档"
Search: "{topic} GitHub"
Search: "{topic} official announcement"

Goal: Get authoritative first-hand information

Round 2: Technical Analysis (技术解析)

Search: "{topic} 详细介绍"
Search: "{topic} 教程 tutorial"
Search: "{topic} how it works"

Goal: Understand technical details and mechanisms

Round 3: Comparison & Reviews (对比评测)

Search: "{topic} vs {competitor}"
Search: "{topic} 评测 review"
Search: "{topic} pros cons"

Goal: Get different perspectives and comparisons

Round 4: Supplementary Verification (补充验证)

# Analyze gaps from previous rounds
missing_info = analyze_gaps(previous_results)
Search: "{missing_info}"
Search: "{topic} 最新 latest 2024 2025"

Goal: Fill information gaps and get latest updates

Output Format

Generate structured material document:

# {Topic} 素材收集报告

## 收集时间
{timestamp}

## 核心信息
- **官方定义**: ...
- **关键特性**: ...
- **最新动态**: ...

## 素材列表

### 素材 1
- **标题**: ...
- **来源**: {url}
- **摘要**: 2-3句话概括
- **关键点**:
  - 要点1
  - 要点2
- **潜在选题角度**: ...
- **推荐推文类型**: [高价值干货/犀利观点/热点评论/故事洞察/技术解析]
- **QualityScore (0-10)**: ...
- **TopicCluster**: ...
- **HookCandidates**:
  - ...

### 素材 2
...

## 热度分析
- **当前热度**: 高/中/低
- **趋势**: 上升/稳定/下降
- **讨论焦点**: ...

## 争议点
- 争议1: ...
- 争议2: ...

## 下一步建议
使用 `/x-filter` 对素材进行打分筛选

Additionally, append a machine-readable block for hooks/state ingestion:

MATERIALS_JSON
{
  "schema_version": "x_skills.materials.v1",
  "topic": "{Topic}",
  "timestamp": "{timestamp}",
  "items": [
    {
      "title": "...",
      "url": "...",
      "summary": "...",
      "key_points": ["..."],
      "quality_score_0_10": 0,
      "topic_cluster": "...",
      "hook_candidates": ["..."]
    }
  ]
}

Execution Steps

Receive topic from user
Execute Round 1 searches (official sources)
Execute Round 2 searches (technical analysis)
Execute Round 3 searches (comparisons)
Analyze gaps from rounds 1-3
Execute Round 4 searches (fill gaps)
Synthesize results into structured format
Output MATERIALS_JSON (for hooks/state ingestion)
(Optional) Append event for feedback loop:
- python ~/.claude/skills/x-create/scripts/x_state.py event --event collect.completed --payload-json '{"topic":"...","items_count":12,"explore_ratio":0.2}'
Report summary to user

Example

User: /x-collect Claude MCP协议

Expected behavior:

Search "Claude MCP协议官方文档"
Search "MCP Model Context Protocol GitHub"
Search "MCP协议详细介绍"
Search "MCP协议教程"
Search "MCP vs function calling"
Search "MCP协议评测"
Identify gaps: need more about security, adoption rate
Search "MCP协议安全性"
Search "MCP协议最新 2025"
Generate structured material report

Integration

After collection, suggest:

素材收集完成！共找到 X 条相关素材。

下一步：运行 /x-filter 对素材进行打分筛选，≥7分的选题将进入创作池。

Tips

For trending topics, prioritize recency (2024-2025)
For technical topics, prioritize official docs and GitHub
For controversial topics, collect multiple perspectives
Always note the source URL for credibility