agentleFS
asoisoasoiso/trend-radar/trend-deep-dive/SKILL.md

trend-deep-dive

Skill2 starsChanged 7 months agoApache-2.0View on GitHub

Deep-dive analysis of a single topic — heat curve, lifecycle stage, viral anomaly check, near-future prediction, related events, and sentiment evolution via the trendradar MCP. Use when the user asks 深挖XX / 这个话题怎么演变的 / XX 的趋势 / 生命周期 / XX 会不会火 / 异常热度 / 舆情演化 / 把 XX 说清楚 / predict X 的走势 / deep dive on X.

The file

Show its rules and text

We split this skill into the rules it gives an agent

18 rules. About 0 are common, 8 are distinctive, and 10 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 10

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  • could be automatic: a setting could enforce it
  1. Very specific:

    If the user asks about many topics at once, redirect to trend-monitor.

    line 24
  2. Very specific:

    If the user asks two of the above in one breath, run them sequentially in this order: trend → lifecycle → viral → predict.

    line 38
  3. Very specific:

    If unsure, start with trend (cheapest, gives a baseline curve), then escalate.

    line 39
  4. Distinctive:

    Resolve the time window: If the user says "上周", "最近三天", "这个月", "from Apr 10 to Apr 20", call:

    line 47
  5. Very specific:

    Use the returned date_range for every downstream call.

    line 51
  6. Distinctive:

    Surface satellite events: Run:

    line 64
  7. Distinctive:

    (Optional) Period comparison: If the user said "vs 上周", "比上个月怎么样", "compared to last month", call:

    line 85
  8. Distinctive:

    Use these to ground key-node descriptions and sentiment quotes — never paraphrase from titles alone.

    line 98
  9. Very specific:

    Topic too rare / no data: If analyze_topic_trend returns an empty curve, do not fabricate.

    line 145
  10. Very specific:

    English entity name: Run find_related_news with both the English term and the most likely 中文译名 (e.g.

    line 146
  11. Very specific:

    Sentiment-only ask: If the user explicitly only wants 舆情 / 情感, you may skip lifecycle and predict but still run …

    line 148
  12. Very specific:

    Window ≤ 24h: Use granularity="hour". Lifecycle judgement gets noisy on intraday windows.

    line 149
  13. Very specific:

    Don't speculate beyond the data.

    line 153
  14. Distinctive:

    Don't over-claim from one data point.

    line 154
  15. Distinctive:

    Don't skip sentiment when the user asks about "舆情 / 网友怎么看 / 情绪".

    line 155
  16. Very specific:

    Don't merge unrelated topics into one report.

    line 156
  17. Distinctive:

    Don't read article bodies for all related items — only top 3–5.

    line 157
  18. Distinctive:

    Don't output the report before step 4 (sentiment) finishes unless the user explicitly said "skip sentiment".

    line 158

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

trend-deep-dive

Single-topic, multi-angle investigation skill. Given one topic (a person, event, product, hashtag, or phrase), produce a structured report covering heat curve, lifecycle stage, viral anomaly assessment, near-future prediction, related satellite events, and sentiment evolution. Backed by the trendradar MCP.

When to activate

Activate when the user's intent is to understand one topic in depth (not to scan many topics, not to compose a campaign). Trigger phrases include:

  • "深挖 XX" / "把 XX 说清楚" / "XX 是怎么演变的"
  • "XX 的趋势" / "XX 的热度曲线" / "XX 这几天怎么样"
  • "XX 处于什么阶段" / "XX 的生命周期" / "XX 还能火多久"
  • "XX 会不会火" / "XX 接下来怎么走" / "predict X 的走势"
  • "XX 是不是异常热点" / "有没有人为推动" / "是不是被刷的"
  • "XX 的舆情" / "舆情演化" / "网友怎么看 XX"
  • "deep dive on X" / "go deep on X"

If the user asks about many topics at once, redirect to trend-monitor. If they want to make a piece of content off the topic, redirect to viral-forge. If they want a report deliverable, hand off to trend-report after finishing the analysis.

Decision tree for analysis_type

mcp__trendradar__analyze_topic_trend accepts an analysis_type argument. Pick exactly one based on the user's primary goal:

User goal (paraphrased)analysis_typeNotable args
"热度曲线 / 看变化 / 这几天怎么样 / 趋势是啥"trendgranularity="day" (default) or "hour" if window < 3 d
"现在处于哪个阶段 / 还能火多久 / 是萌芽还是衰退"lifecyclenone beyond topic/date_range
"是不是异常热点 / 有没有人为推动 / 是不是被刷的"viralspike_threshold (e.g. 3.0), time_window (e.g. "24h")
"未来几小时/几天会怎样 / 还会不会涨 / predict 走势"predictlookahead_hours (default 24), confidence_threshold (0.6)

Rules of thumb:

  • If the user asks two of the above in one breath, run them sequentially in this order: trend → lifecycle → viral → predict. Do not parallelize different analysis_type calls on the same topic — each call already returns curve data the next one can reference.
  • If unsure, start with trend (cheapest, gives a baseline curve), then escalate.

See references/decision-tree.md for worked examples.

Standard workflow

Execute these steps in order. Skip a step only if its precondition is not met (noted inline).

  1. Resolve the time window. If the user says "上周", "最近三天", "这个月", "from Apr 10 to Apr 20", call:
    mcp__trendradar__resolve_date_range(text=<user phrase>)
    Use the returned date_range for every downstream call. If the user gave no time hint, default to the last 7 days and state that assumption in the final report.
  2. Run the primary analysis. Pick analysis_type via the decision tree above, then:
    mcp__trendradar__analyze_topic_trend(
      topic=<topic>,
      analysis_type=<chosen>,
      date_range=<from step 1>,
      granularity="day"   # use "hour" if window ≤ 72h or user wants intraday detail
    )
    Capture the returned data points, peak timestamps, and any lifecycle/anomaly/prediction fields.
  3. Surface satellite events. Run:
    mcp__trendradar__find_related_news(
      reference_title=<topic>,
      date_range=<same as step 1>,
      threshold=0.5,
      limit=30
    )
    These are the related events orbiting the topic — sub-stories, reactions, parallel angles. Sort by rank/heat.
  4. Analyze sentiment. Over the related-news pool from step 3 (or, if it returned <5 items, over a broader search_news pull on the topic), call:
    mcp__trendradar__analyze_sentiment(
      topic=<topic>,
      date_range=<same>,
      # pass article ids/urls from step 3 if the tool accepts them; otherwise let it re-fetch
    )
    Extract: positive %, neutral %, negative %, and one representative quote per bucket.
  5. (Optional) Period comparison. If the user said "vs 上周", "比上个月怎么样", "compared to last month", call:
    mcp__trendradar__compare_periods(
      topic=<topic>,
      current_range=<step 1>,
      previous_range=<auto-shifted by the same span>
    )
  6. Read the substantive top items. Take the top 3–5 highest-ranked items from step 3 and pull their bodies:
    mcp__trendradar__read_articles_batch(article_ids=[...])
    Use these to ground key-node descriptions and sentiment quotes — never paraphrase from titles alone.

Output schema

Render the final answer as Markdown with all of the following sections, in this order. Omit a section only if explicitly N/A and say so.

## 一句话结论
<one sentence: what stage + what direction + one notable signal>

## 热度曲线
| 日期/小时 | 热度 | 备注 |
|---|---|---|
| ... | ... | ... |

**趋势判定**: 上升 / 平台 / 衰退 (+ 短理由)

## 生命周期阶段
**当前阶段**: 萌芽 / 爆发 / 平台 / 衰退
**判定依据**: <2-3 句,引用曲线斜率、峰值距今时间、相关事件密度>

## 关键节点
- YYYY-MM-DD · 平台 · <事件标题/触发动作>
- ...

## 情感分布
- 正面 X% — 代表观点: "..."
- 中性 Y% — 代表观点: "..."
- 负面 Z% — 代表观点: "..."

## 相关事件 / 卫星话题
1. <标题> (相似度 0.xx) — 一句话定位
2. ...

## 预测 (仅在调用了 predict 时)
- 未来 N 小时方向: 上行 / 持平 / 下行
- 置信度: 0.xx
- 触发再涨/熄火的关键变量: ...

## 数据来源与置信度
- MCP 调用: analyze_topic_trend(<type>), find_related_news, analyze_sentiment, [compare_periods], read_articles_batch
- 数据点数量: N 篇文章 / M 个时间桶
- 置信度评估: 高 / 中 / 低 — <理由,例如样本量、平台覆盖度、是否跨多源验证>

Edge cases

  • Topic too rare / no data. If analyze_topic_trend returns an empty curve, do not fabricate. Tell the user: "过去 <window> 内 trendradar 索引中未发现 <topic>" and offer two follow-ups: (a) widen the date range, (b) try a related broader keyword.
  • English entity name. Run find_related_news with both the English term and the most likely 中文译名 (e.g. "OpenAI" + "OpenAI 公司" / "DeepSeek" + "深度求索"). Merge the two pools, dedupe by article id.
  • Ambiguous / polysemous topic (e.g. "苹果" = company OR fruit, "Mercury" = planet OR element OR celebrity). First run find_related_news with limit=10, glance at returned titles. If you see ≥2 distinct senses, stop and ask the user which sense they mean before continuing — do not silently pick.
  • Sentiment-only ask. If the user explicitly only wants 舆情 / 情感, you may skip lifecycle and predict but still run analyze_topic_trend(analysis_type="trend") to provide the curve context — sentiment without volume context is misleading.
  • Window ≤ 24h. Use granularity="hour". Lifecycle judgement gets noisy on intraday windows; flag confidence as 中 or 低.

Don'ts

  • Don't speculate beyond the data. If only analyze_topic_trend(trend) was called, do not issue a prediction — say "未运行 predict, 不做未来走势判断".
  • Don't over-claim from one data point. A single article does not establish a curve, a stage, or a sentiment split.
  • Don't skip sentiment when the user asks about "舆情 / 网友怎么看 / 情绪". Sentiment is the deliverable in that case, not a nice-to-have.
  • Don't merge unrelated topics into one report. If the user listed three topics, run this skill three times, or hand off to trend-monitor.
  • Don't read article bodies for all related items — only top 3–5. Body reads are expensive and rarely change the report shape past the first few.
  • Don't output the report before step 4 (sentiment) finishes unless the user explicitly said "skip sentiment".

See also

  • Shared MCP tool reference: ../docs/mcp-tools.md
  • Decision tree worked examples: references/decision-tree.md
  • Multi-topic scanning: trend-monitor skill
  • Content generation off the analysis: viral-forge skill
  • Packaging the analysis as a deliverable: trend-report skill

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.