SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsx-ai-radar

x-ai-radar

General & other

View repository

Updated 10-15 20:12

About this skill

X AI Radar — Antigravity exclusive X hot topic radar

Automatically browse the X AI timeline → only take the latest hot posts within 10 hours→ by velocity/impressions/engagement + big-V weightcalculate a heat score → output a Chinese Top 8-10 hot ranking. Purely read-only, no writing, no posting.

When to use

  • Want a list of AI-related, genuinely high-heat posts within 10hrather than obscure posts (see description for trigger phrases)

Core workflow (Antigravity read-only loop)

Phase 1: Browse the timeline (must go through the browser sub-agent)

  1. Sole entry point: after receiving the trigger, must directly invoke_subagent invoke the browser sub-agent（TypeName: "browser"), and the main Agent is strictly prohibited from handwriting temporary CDP scripts to bypass it.
  2. Enter the AI timeline: have the sub-agent open x.com/home(the user's daily AI-tuning feed; if the user has an AI List, prioritize the List), and wait for the first screen to finish loading.
  3. Scroll to capture candidates: scroll down the timeline and capture 15-20 candidates, recording for each:
    • url / 作者 DisplayName + @handle / 是否认证/大V / 粉丝数
    • 发布时间 / 距今小时数 h
    • Views / Likes / Reposts / Replies / Bookmarks
    • 是否有媒体/链接/引用
  4. 10-hour hard window：now - 发布时间 > 10h are all discarded, not included as candidates (better to have fewer than to make do with inferior ones).
  5. Calculate velocity and engagement(calculated by the main Agent after the sub-agent returns the data):
    • viewsPerHour = Views / max(h, 0.5)
    • likesPerHour = Likes / max(h, 0.5)
    • engagementRate = (Likes + Reposts*2 + Replies*2) / Views

Phase 2: Filtering and heat score (big-V/high-velocity first)

  1. Hard threshold + speed compensation (passing either one is enough)：
    • English/general：Views ≥ 30,000 且 Likes ≥ 200
    • Chinese/AI circle：Views ≥ 5,000 且 Likes ≥ 50
    • Potential viral compensation(especially new posts):h ≤ 2 且 (viewsPerHour ≥ 5,000 或 likesPerHour ≥ 100) 且 Replies ≥ 15 → considered rising, pass directly
    • Anything that does not pass the threshold is all removed, and the sub-agent needs to keep scrolling down until enough candidates are gathered
  2. Account and velocity weighting：
    • Big-V (verified / followers ≥ 10k) directly weighted
    • Non-big-V but with viewsPerHour extremely high is also retained (viral spread), not judging solely by follower count
    • Engagement-farming/ad accounts (engagementRate < 0.001 且 Replies < 5) are downweighted and removed
  3. AI relevance(keyword pre-screening + LLM re-judgment):
    • Read references/ai-keywords.json for pre-screening (if hit, maybe）
    • then have the LLM judge AI-relevance: high/medium/low, only retaining high/medium（vibe coding / cursor / agent / gemma / deepseek and other implicit AI items that the LLM recovers)
  4. Heat score(used for sorting, scale 0-100): 
    heatScore = (log10(Views+1)*10 + Likes*0.4 + Reposts*0.8 + Replies*1.0 + viewsPerHour*0.02) / 10
# 示例：80k Views / 600 Likes / 120 Reposts / 90 Replies / 20k views/h → 49+240+96+90+400=875 → /10 ≈ 88 分
    • Big-V +5 points,h ≤ 2 rising +8 points, containing original post link/paper +3 points (all added after /10)
    • Sort by heatScore in descending order and take Top 8-10
  5. Deduplication: deduplicate by tweet id/url , and for the same author, if there are multiple viral posts within 10h, keep only the highest-scoring one

Phase 3: Summary output (within the session only)

  • Do not write files, do not post, do not like, do not follow, output Markdown only in the current session
  • Summarize each in 1 sentence of Chinese Core facts(not slogans), keep necessary English product/model names, do not add #标签, do not make a watered-down summary
  • By heatScore sort output, at the end give 1 sentence on today's AI circle trend pulse

Output template (strictly follow this):

## 𝕏 AI 雷达 · 10h 热榜（YYYY-MM-DD HH:mm CST）

> 窗口：近 10 小时 | 候选 18 → 入榜 9 | 按热度分排序

| # | 热度分 | 作者 | 曝光/互动 | 流速 | 一句话摘要 | AI标签 |
|---|--------|------|-----------|------|------------|--------|
| 1 | 88 | 花叔 @huashu | 128k / 1.2k♡ 180↻ 95💬 | 18k/h | Grok 4.1 泄露的上下文窗口实测细节，工程侧关注 KV 缓存改动 | LLM |
| 2 | 79 | ... | ... | ... | ... | Agent |

**今日脉搏**：一句话总结今日最强的 1-2 个信号（例：Agent 框架的本地化与低成本推理成为主线）

**候选透视**：共抓 18，剔除 低热 6 / 非AI 3 / 超 10h 2

Field descriptions:曝光/互动 = Views / Likes Reposts Replies；流速 = viewsPerHour

Boundaries and prohibitions

  • 10-hour hard constraint: posts older than 10h do not enter the list, even if highly popular
  • Read-only: strictly prohibited type_text / 点赞 / 转推 / 关注, this skill does not produce any writes
  • No writing: do not generate Quote/Reply drafts; if writing is needed, call a writing-type skill separately (such as x-browser-automation, if installed)
  • Not logged in: if the sub-agent finds it is not logged in, immediately error out and give https://x.com/login, do not fabricate data
  • Insufficient candidates: if fewer than 8, truthfully output N items + note the reason, do not pad the numbers

References

  • references/ai-keywords.json — AI keyword list (about 60 words, preliminary screening before LLM re-judgment)