trend-monitor

trendradar

When to activate

• 今日热点 / 今天有什么热点 / 今天热搜
• 各平台热搜 / 看一下当前热搜 / 全网热点 / 全网在聊什么
• 现在在火什么 / 最近在聊什么
• what's trending / what's trending now / multi-platform sweep / hotspot snapshot
• 给我看一下 zhihu / weibo / douyin / bilibili / toutiao / baidu / thepaper / ifeng / tieba / cls-hot / wallstreetcn-hot 的热搜

Workflow

1. Resolve the time window. If the user mentions any date that isn't "now / 现在 / 今天 / today", call

   mcp__trendradar__resolve_date_range

   first and reuse that range in the next steps. For "今天 / now / latest", skip this step.
2. Pull the latest news batch. Call

   mcp__trendradar__get_latest_news

   with

   limit=80

   and

   include_url=False

   . Set

   include_url=True

   only if the user explicitly asked for links. If the user named specific platforms, pass

   platforms=[...]

   (e.g.

   ["weibo", "zhihu"]

   ); otherwise let the tool return the full set.
3. Run keyword extraction in parallel. Issue both calls in the same response:

   • mcp__trendradar__get_trending_topics

     with

     extract_mode="keywords"

     ,

     top_n=20

     — preset focus list (curated themes the project tracks).

   • mcp__trendradar__get_trending_topics

     with

     extract_mode="auto_extract"

     ,

     top_n=20

     — emergent terms surfaced from this batch only.
4. Cross-reference. Build two buckets before drafting the output:
   • 全网共振: keywords that appear on ≥3 platforms in this batch (pulled from either extraction mode).
   • 新冒头: terms present in

     auto_extract

     results but absent from the

     keywords

     preset list — these are the emergent signals worth flagging.
5. Pick the output template based on user intent (see "Output templates" below). Default to 平台分组视图 unless the user asked for themes / topics / "what's everyone talking about across platforms".

Output templates

平台分组视图 (default)

# 今日全网热点 · <YYYY-MM-DD HH:mm>

## 微博 (weibo)
1. <标题> · 热度 <metric>
2. ...
(top 5)

## 知乎 (zhihu)
1. ...
(top 5)

## 抖音 (douyin)
...

## B站 (bilibili)
...

(只列出本批次实际有数据的平台，每个平台 top 5)

---

### 全网共振
- **<关键词>** — 出现在 weibo / zhihu / toutiao（3 个平台）
- **<关键词>** — 出现在 weibo / douyin / bilibili / baidu（4 个平台）

### 新冒头 (auto-extract only)
- **<新词>** — 首次在本批次浮现，主要源自 <平台>
- **<新词>** — ...

主题分组视图 (when user asks for cross-platform themes)

# 今日热点主题 · <YYYY-MM-DD HH:mm>

## <主题 / 关键词>
- 微博 #2: <标题片段>
- 知乎 #5: <标题片段>
- 抖音 #1: <标题片段>

## <主题 / 关键词>
- 头条 #3: ...
- 百度 #7: ...

(每个主题列出实际证据：哪个平台、第几名、标题片段)

---

### 全网共振
- **<关键词>** — 出现在 X / Y / Z（N 个平台）

### 新冒头 (auto-extract only)
- **<新词>** — 首次在本批次浮现，主要源自 <平台>

Edge cases

• get_latest_news

  returns empty / very stale. Call

  mcp__trendradar__get_system_status

  to check crawl freshness. If the latest crawl is clearly stale, fall back to

  mcp__trendradar__get_latest_rss

  and tell the user the snapshot is from RSS feeds, not the platform crawl. Don't silently substitute.
• User asks for a platform not in the supported list. Supported set:

  zhihu

  ,

  weibo

  ,

  douyin

  ,

  bilibili

  ,

  toutiao

  ,

  baidu

  ,

  thepaper

  ,

  ifeng

  ,

  tieba

  ,

  cls-hot

  ,

  wallstreetcn-hot

  . If they ask for something else (e.g. xiaohongshu, twitter), say so explicitly, list what is available, and ask whether to proceed with the closest covered platforms.
• One of the two extraction calls fails. Render the output with whatever succeeded and note which bucket (preset / emergent) is missing. Don't block the whole response on a single failure.
• User asks for a date in the past. Resolve via

  mcp__trendradar__resolve_date_range

  , then use

  mcp__trendradar__get_news_by_date

  instead of

  get_latest_news

  . Same extraction + cross-reference flow afterwards.
• Ambiguous "今天" near midnight. Trust

  get_latest_news

  — it returns the most recent crawl batch regardless of wall-clock date.

Don'ts

• Don't paginate huge lists. Cap at top 5 per platform / per theme. The user wants a snapshot, not a dump.
• Don't include URLs unless the user explicitly asked for them.

  include_url=False

  is the default for a reason — links blow up the output.
• Don't fabricate trend data when MCP fails. If

  get_latest_news

  errors out and the RSS fallback also fails, say so plainly and stop — never invent platform rankings.
• Don't merge preset-keyword and auto-extract output into one undifferentiated list. The 全网共振 / 新冒头 split is the whole point of running both extraction modes.
• Don't issue the two

  get_trending_topics

  calls sequentially. They are independent — fire them in parallel.

See also

• ../docs/mcp-tools.md — full

  mcp__trendradar__*

  tool catalog (parameters, return shapes, rate-limit notes).