/last30days
◉ LIVE SIGNAL AGGREGATOR

/last30days 訊號搜尋引擎Signal Search Engine

Google 聚合編輯，/last30days 搜尋人群。 Google aggregates editors. /last30days searches people. 以真實社群互動分數（上票數、按讚、Polymarket 賠率）排序。 Scored by real engagement — upvotes, likes, and Polymarket odds.

v3.3.2 14 Sources Parallel Fetch Engagement Scoring Expert Mode npx skills add mvanhorn/last30days-skill
02 · 設計哲學Philosophy

人群智慧，不是編輯觀點Crowd Signal, Not Editorial Opinion

傳統搜尋引擎以反向連結和 SEO 決定排序；/last30days 以實際人類互動行為決定排序。 Traditional search ranks by backlinks and SEO. /last30days ranks by actual human interaction behaviour.

「Google 聚合編輯，/last30days 搜尋人群。」 "Google aggregates editors. /last30days searches people."
— last30days README

傳統搜尋引擎Traditional Search

  • 反向連結決定排序Backlinks determine rank
  • SEO 優化可操縱SEO manipulation possible
  • 文章發布時間衡量新鮮度Publish date = freshness
  • 編輯/機構聲音主導Editorial voices dominate
  • 無真實參與度資料No real engagement data

/last30days 訊號搜尋/last30days Signal Search

  • 真實上票/按讚決定排序Real upvotes/likes determine rank
  • Polymarket 真實金錢賠率Polymarket real-money odds
  • 過去 30 天滾動時間窗Rolling 30-day time window
  • 社群聲音（Reddit、X、HN）Community voices (Reddit, X, HN)
  • 轉錄稿、評論深度擷取Transcript and thread depth fetch
03 · 訊號源Signal Sources

14 個訊號來源平台14 Signal Source Platforms

並行搜尋全部平台，各自擷取深度資料（轉錄稿、留言串、投票數）。 All platforms searched in parallel. Each fetches depth data (transcripts, comment threads, vote counts).

Reddit
上票數 · 留言數Upvotes · Comments
完整留言串Full threads
YouTube
觀看數 · 按讚 · 留言Views · Likes · Comments
完整轉錄稿Full transcripts
TikTok
觀看數 · 按讚 · 分享Views · Likes · Shares
部分字幕Partial captions
Instagram Reels
按讚 · 留言 · 播放Likes · Comments · Plays
互動摘要Engagement summary
Hacker News
積分 · 留言數Points · Comments
完整討論串Full discussion
Polymarket
真實金錢賠率Real-money odds
機率 + 成交量Probability + volume
GitHub
星標 · Fork · IssueStars · Forks · Issues
最近活動Recent activity
Digg
社群投票Community votes
基本摘要Basic snippet
Threads
按讚 · 回覆Likes · Replies
部分深度Partial depth
Pinterest
儲存數 · 瀏覽Saves · Views
基本摘要Basic snippet
Bluesky
按讚 · 轉發Likes · Reposts
完整串Full threads
Perplexity
AI 聚合網頁結果AI-aggregated web results
引用來源Cited sources
Web
Google / Bing 基準線Google / Bing baseline
摘要Snippet

完整深度Full depth · 部分深度Partial · 基本摘要Basic

04 · v3 管線流程v3 Pipeline

7 步驟搜尋管線7-Step Search Pipeline

v3 引入全並行擷取，速度提升約 3 倍；深度擷取在並行搜尋同時進行。 v3 introduces fully parallel fetching (~3x speedup); depth fetching happens concurrently with source search.

1

解析查詢Resolve Query

分析查詢意圖，決定最佳來源組合與擷取策略。選擇深度 vs. 廣度模式。Analyse query intent, determine optimal source mix and fetch strategy. Choose depth vs. breadth mode.

intent-detection source-routing
2

並行搜尋全部來源Parallel Search All Sources

同時對所有 14 個平台發出搜尋請求。v3 移除序列等待，大幅縮短延遲。Simultaneously query all 14 platforms. v3 removes sequential waits, drastically reducing latency.

parallel · async 14 sources
3

擷取深度資料Get Depth

對高訊號結果擷取轉錄稿（YouTube）、完整留言串（Reddit/HN）、Polymarket 成交量與機率。For high-signal results, fetch transcripts (YouTube), full comment threads (Reddit/HN), Polymarket volume and probability.

transcripts comment-threads market-odds
4

合併去重叢集Merge & Deduplicate Clusters

將跨平台重複討論同一主題的結果合併為叢集，避免重複計算互動分數。Merge cross-platform results discussing the same topic into clusters, avoiding double-counting engagement.

clustering dedup
5

依互動分數合成Synthesize by Engagement

以真實社群分數（上票數、按讚數、Polymarket 賠率）加權排序，而非由 AI 主觀判斷重要性。Rank by real social scores — not AI's subjective importance judgment.

engagement-weighted evidence-first
6

專家後續追問（可選）Expert Follow-up Mode (Optional)

自動生成深入追問問題，供使用者或 AI 進行下一輪搜尋。由 --expert 旗標啟用。Auto-generate intelligent follow-up questions for the next search round. Enabled via --expert flag.

--expert optional
7

格式化輸出Formatted Output

依輸出契約（8 LAWs）格式化輸出：徽章行、行內引用連結、引擎頁腳，不含「Sources:」區塊或章節標題。Format output per the Output Contract (8 LAWs): badge line, inline citation links, engine footer — no Sources block, no section headers.

8 LAWs enforced GateGuard hook
05 · v3 新功能v3 New Features

v3 重大更新v3 Major Updates

v3 是架構性重寫，不只是新增功能。v3 is an architectural rewrite, not just a feature addition.

v3 NEW
⚡
全並行擷取Full Parallel Fetch
14 個來源同時搜尋，深度資料並行擷取。總延遲從序列加總降為最慢單一來源。All 14 sources queried simultaneously. Total latency drops from sum-of-all to slowest-single.
v3 NEW
🔬
深度資料擷取Depth Data Fetching
YouTube 完整轉錄稿、Reddit 完整留言串、Polymarket 成交量與機率，而非只有摘要。Full YouTube transcripts, complete Reddit comment threads, Polymarket volume and probability — not just snippets.
v3 NEW
📊
互動分數排序Engagement Score Ranking
以真實上票數、按讚數、Polymarket 賠率加權排序，而非 AI 主觀判斷。Weighted by real upvotes, likes, and Polymarket odds — not AI's subjective relevance.
v3 NEW
🧠
專家追問模式Expert Follow-up Mode
自動生成智慧後續問題，幫助使用者深入探索主題。由 --expert 旗標啟用。Auto-generates intelligent follow-up questions. Enabled via --expert flag.
🛡️
GateGuard Hook
PreToolUse Hook，阻止在宣告事實前執行 Bash，強制先驗後行，防止臆測輸出。PreToolUse hook that blocks Bash until facts are declared. Enforces evidence-first output.
🔄
過期克隆自檢Stale-Clone Self-Check
啟動時自動偵測本地技能版本是否落後於上游，提示更新。At startup, auto-detects if local skill version is behind upstream and prompts to update.
06 · 指令與旗標Commands & Flags

指令參考Command Reference

自然語言旗標作為查詢的修飾詞，改變搜尋範圍、深度或輸出格式。Natural-language flags act as modifiers on your query, changing scope, depth, or output format.

/last30days <query>
主指令ENTRY

主要搜尋指令。在過去 30 天內聚合 14 個平台的社群訊號，依互動分數排序輸出。Primary search command. Aggregates social signals from 14 platforms over the past 30 days, ranked by engagement score.

範例用法Example Usage
  • /last30days react 19 opinions
  • /last30days bun vs node speed
  • /last30days openai vs anthropic
輸出包含Output Includes
  • 徽章行（來源 + 分數）Badge line (sources + score)
  • 行內引用連結Inline citation links
  • 引擎頁腳Engine footer
/last30days <q> --days N
MODIFIER

自訂時間回溯窗口。預設 30 天；可縮短（7 天追蹤趨勢）或延長（90 天看大趨勢）。Customise the lookback window. Default 30 days; shorten for trend tracking or extend for macro-trends.

典型值Typical Values
  • --days 7 — 即時熱點Hot right now
  • --days 30 — 預設Default
  • --days 90 — 長期趨勢Long-term trend
注意Notes
  • 窗口越長，訊號越稀Longer = sparser signals
  • Polymarket 不受天數影響Polymarket ignores --days
/last30days <q> --sources reddit,hn
FILTER

限定搜尋特定平台子集。適合只關心技術社群或主流輿情的場景。Restrict search to a specific platform subset. Useful for tech-only or mainstream-only queries.

可用名稱：Valid names: reddit · twitter · youtube · tiktok · instagram · hn · polymarket · github · digg · threads · pinterest · bluesky · perplexity · web
/last30days <q> --expert
MODE

啟用專家追問模式。搜尋完成後自動生成 3–5 個深入追問問題。Enable expert follow-up mode. After search, auto-generates 3–5 intelligent follow-up questions.

適合場景Best For
  • 深度市場調查Deep market research
  • 技術決策比較Tech stack decisions
  • 競品分析Competitor analysis
追加輸出Appends
  • 3–5 個追問問題3–5 follow-up questions
  • 建議子查詢Suggested sub-queries
/last30days <q> --min-score N
FILTER

設定最低互動分數門檻，過濾低互動冷門帖，只保留社群真正共鳴的內容。Set minimum engagement score threshold, filtering out low-engagement posts.

分數跨平台標準化（Reddit 上票數與 YouTube 觀看數不可直接比較）Scores are normalised cross-platform (Reddit upvotes ≠ direct comparison with YouTube views)
/last30days <q> --no-depth
MODIFIER

跳過深度資料擷取（轉錄稿、完整留言串）。速度最快，適合快速概覽。Skip depth data fetching (transcripts, full threads). Fastest mode; best for a quick overview.

/last30days <q> --beta
BETA

啟用 Beta 頻道實驗功能。功能可能不穩定，不建議用於正式研究。Enable Beta channel experimental features. May be unstable; not recommended for serious research.

07 · 輸出契約Output Contract

8 條 LAW — 強制輸出格式規則8 LAWs — Mandatory Output Format Rules

SKILL.md 定義的嚴格輸出契約。GateGuard Hook 在執行前強制驗證。Strict output contract defined in SKILL.md. GateGuard hook validates before execution.

LAW 1
徽章行必須存在Badge Line Required
輸出首行必須包含來源清單與互動分數徽章。First line must include source list and engagement score badge.
LAW 2
禁止「Sources:」區塊No Sources Block
不得在輸出末尾加「Sources:」清單區塊。Must not append a "Sources:" list block at the end.
LAW 3
禁止發明標題No Invented Titles
使用引擎返回的原始標題，不得自行改寫。Use the exact title returned by the engine — never rewrite it.
LAW 4
禁止破折號（em-dash）No Em-Dashes
輸出中禁用 — （em-dash），改用其他標點符號。Prohibited: — (em-dash). Use other punctuation instead.
LAW 5
禁止章節標題No Section Headers
輸出為流暢散文，不得插入 ## 或粗體章節標題。Output is flowing prose — no ## or bold section headers.
LAW 6
引擎頁腳直接傳遞Engine Footer Pass-Through
搜尋引擎的來源歸因頁腳必須原樣保留。The search engine's attribution footer must be preserved verbatim.
LAW 7
禁止原始證據叢集No Raw Evidence Clusters
不得直接傾倒原始搜尋結果；必須先合成再輸出。Do not dump raw search result clusters — synthesise first.
LAW 8
行內引用連結Inline Citation Links
所有引用必須以 Markdown 連結形式嵌入散文中，而非收集在末尾。All citations embedded as Markdown links in prose — not collected at the end.
PLANNER REQUIREMENT · 在執行實際搜尋前，必須先產生 Planner（搜尋計劃）。GateGuard Hook 會阻止未宣告事實即執行 Bash。Before executing actual searches, a Planner (search plan) must be generated. GateGuard Hook blocks Bash unless facts are declared first.
08 · 設定配置Configuration

API 金鑰與環境設定API Keys & Environment Configuration

設定層次：per-run 旗標 > 環境變數 > 預設值。大部分來源無需 API 金鑰即可使用。Config layers: per-run flags > env vars > defaults. Most sources work without API keys (lower rate limits).

API Keys Matrix
平台Platform 需求Required 環境變數Env Var 無金鑰限制Without Key
Reddit OPTIONAL REDDIT_CLIENT_ID 可用，速率較低Works, lower rate
X / Twitter REQUIRED TWITTER_BEARER_TOKEN 不可用Unavailable
YouTube OPTIONAL YOUTUBE_API_KEY 有配額限制Quota limited
Polymarket NONE Public API 完整可用Full access
GitHub OPTIONAL GITHUB_TOKEN 60 req/hr
Perplexity REQUIRED PERPLEXITY_API_KEY 跳過此來源Source skipped
其他 8 源Other 8 sources NONE 公開搜尋Public search 完整可用Full access
推理提供者優先順序Reasoning Provider Priority
優先Priority 提供者Provider 環境變數Env Var
1st本機 Harness 模型Local harness model自動偵測Auto-detected
2ndOpenRouterOPENROUTER_API_KEY
3rdAnthropic DirectANTHROPIC_API_KEY
4thOpenAI DirectOPENAI_API_KEY
09 · 詞彙表Vocabulary

核心概念定義Core Concept Definitions

CONCEPTS.md 定義的術語，了解生態系各層次的職責分工。Terms defined in CONCEPTS.md — understand the responsibility of each ecosystem layer.

Skill
SKILL.md 文件 + 工具鏈。定義搜尋行為的「配方」。儲存於 ~/.agents/skills/，相容 50+ Harness。The SKILL.md file + tool chain. The "recipe" defining search behaviour. Stored in ~/.agents/skills/, compatible with 50+ harnesses.
Engine
底層搜尋服務（Perplexity、Web Search API 等），執行實際查詢並返回原始結果。The underlying search service (Perplexity, Web Search API, etc.) that executes actual queries and returns raw results.
Harness
AI 程式碼助理（Claude Code、Cursor、Copilot 等），載入並執行 Skill 的環境。The AI coding assistant (Claude Code, Cursor, Copilot, etc.) that loads and executes the Skill.
Beta Channel
預發布功能頻道。透過 --beta 旗標啟用，提前體驗下一版功能，可能不穩定。Pre-release feature channel. Enabled via --beta flag. Access next-version features early; may be unstable.
GateGuard
PreToolUse Hook，在執行 Bash 前強制要求先宣告事實。防止未驗證的臆測進入搜尋查詢。PreToolUse hook requiring facts to be declared before running Bash. Prevents speculation from entering search queries.
Engagement Score
跨平台標準化互動分數。聚合上票數、按讚數、觀看數、Polymarket 賠率作為排序依據。Cross-platform normalised score aggregating upvotes, likes, views, and Polymarket odds as the ranking signal.
10 · 使用流程Workflows

典型使用場景Typical Usage Scenarios

逐步流程圖，涵蓋最常見的使用模式。Step-by-step flows covering the most common usage patterns.

場景 AScenario A · QUICK POLL
快速輿情調查Quick Opinion Poll
「大家對 X 怎麼看？」"What do people actually think about X?"
1
YOU
/last30days [topic] --no-depth
--no-depth 跳過轉錄稿，速度最快--no-depth skips transcripts for speed
2
SKILL
GateGuard 驗證 → 並行搜尋 14 源 → 按互動分數排序GateGuard validates → parallel search 14 sources → rank by engagement
3
OUTPUT
徽章行 + 散文合成 + 行內引用連結 + 引擎頁腳Badge line + prose synthesis + inline links + engine footer
場景 BScenario B · TECH DECISION
技術選型調查Tech Stack Decision
「該用 X 還是 Y？」"Should we use X or Y?"
1
YOU
/last30days X vs Y --sources reddit,hn,github --expert
限定技術社群來源，啟用專家追問Focus on tech communities, enable expert follow-ups
2
SKILL
Reddit + HN + GitHub 並行搜尋，擷取完整留言串與 star 趨勢Parallel Reddit + HN + GitHub, fetch full threads and star trends
3
OUTPUT
社群比較分析 + 3–5 個追問問題Community comparison + 3–5 follow-up questions
4
YOU
選一個追問問題，繼續深挖Pick one follow-up question to drill deeper
場景 CScenario C · MARKET INTEL
市場情報調查Market Intelligence
競品分析、品牌輿情監控Competitor analysis, brand sentiment
1
YOU
/last30days [company] --days 7 --min-score 50
7 天短窗口，過濾低互動雜訊7-day window, filter low-engagement noise
2
SKILL
全源並行搜尋，Polymarket 檢查市場預測，Twitter 追蹤即時情緒Full parallel search, Polymarket checks market predictions, Twitter tracks real-time sentiment
3
OUTPUT
高互動帖子合成，含 Polymarket 賠率（若有相關市場）High-engagement synthesis, includes Polymarket odds if relevant markets exist
4
YOU
/last30days [company] negative --days 30
針對負面聲音進行深入調查Follow up on negative sentiment specifically
場景 DScenario D · DEEP RESEARCH
深度專家研究Deep Expert Research
需要完整轉錄稿與深度分析Needs full transcripts and deep analysis
1
YOU
/last30days [topic] --expert --days 90
全深度擷取 + 90 天寬窗口 + 專家追問Full depth + 90-day window + expert follow-ups
2
SKILL
擷取 YouTube 完整轉錄稿、Reddit 深層留言串、HN 技術討論全文Fetch full YouTube transcripts, deep Reddit threads, complete HN discussions
3
SKILL
合成深度資料，生成涵蓋反方觀點的全面分析Synthesise depth data into comprehensive analysis including counterarguments
4
OUTPUT
完整散文分析 + 5 個追問問題 + 引擎頁腳Full prose analysis + 5 follow-up questions + engine footer
5
YOU
迭代使用追問問題，逐步建立完整知識圖景Iterate using follow-up questions to build a complete knowledge picture
安裝方式INSTALL
npx skills add mvanhorn/last30days-skill
安裝至 ~/.agents/skills/，相容 Claude Code、Cursor、Copilot、Codex 等 50+ Harness。Installs to ~/.agents/skills/ — compatible with Claude Code, Cursor, Copilot, Codex, and 50+ harnesses.
/last30days v3.3.2 · mvanhorn/last30days-skill · 以社群真實訊號搜尋Search by real social signal