hf-trending-intelligence

name	hf-trending-intelligence
description	Autonomous daily intelligence pipeline that cross-references HF papers, models, datasets, spaces, leaderboards, and community activity to detect emerging AI trends before they go mainstream. Supports optional topic filtering (LLM, multi-LLM, video-generation, etc.) for focused scans. Produces a scored intelligence report and distributes to Slack + Notion. Use when the user asks to "run daily AI radar", "trending intelligence", "AI research radar", "detect AI trends", "daily paper intelligence", "AI 트렌드 레이더", "일일 AI 인텔리전스", "트렌딩 분석", "연구 레이더", "HF 종합 트렌딩", or wants to know what's trending across the HF ecosystem today. Do NOT use for full paper review with PM analysis (use paper-review). Do NOT use for daily stock analysis (use today). Do NOT use for general web research only (use parallel-web-search). Do NOT use for paper browsing without cross-referencing (use hf-papers). Do NOT use for topic-only scan without full cross-referencing (use hf-topic-radar). Do NOT use for leaderboard-only tracking (use hf-leaderboard-tracker).
metadata	{"author":"thaki","version":"2.1.0","category":"research"}

Trending Intelligence — Daily AI Research Radar

Autonomous daily pipeline that detects emerging AI trends by cross-referencing papers, models, datasets, and community buzz on Hugging Face Hub.

출력 언어

모든 출력물은 한국어로 작성한다. 리포트(markdown), Slack 메시지, Notion 페이지, 인사이트 분석 모두 한글로 작성. 논문 제목, 모델명, 데이터셋명 등 고유명사는 원어 그대로 유지.

Subagent Model Routing

Task Type	Model	Rationale
Exploration / search / file reading	haiku	낮은 비용, read-only 작업
Analysis / implementation / generation	sonnet	균형 잡힌 품질-비용 비율

Agent tool 호출 시 반드시 model 파라미터를 지정한다.

Prerequisites

• hf CLI installed and authenticated (see hf-hub skill)
• Slack MCP configured for posting (channel: #deep-research-trending — C0AN34G4QHK)
• Notion MCP configured for page creation
• jq for JSON processing

Required Skills

• hf-papers — daily paper listing
• hf-models — model search and info
• hf-spaces — trending space search
• hf-collections — curate trending items
• hf-discussions — measure community activity
• hf-dataset-viewer — dataset info
• hf-topic-radar — topic-focused model/space scanning (composable sub-skill)
• hf-leaderboard-tracker — leaderboard cross-reference data
• parallel-web-search — external signal enrichment
• md-to-notion — permanent Notion archive
• kwp-slack-slack-messaging — Slack distribution

Reference Files

• references/scoring-rubric.md — composite scoring weights and thresholds
• references/report-template.md — intelligence report markdown template
• references/collection-naming.md — naming convention for monthly collections

Input

• Date (optional): Defaults to today; can specify a specific date
• Focus area (optional): Filter by domain (e.g., "NLP", "vision", "multimodal")
• Topics (default: LLM, video-generation): Topic filter using HF tags from hf-topic-radar/references/topic-config.md. Phase 1.5 (topic-focused model/space scan) runs by default with these topics. Pass --skip-topics to disable and revert to v1.0 paper-only behavior.
• --leaderboard (optional): 명시적으로 전달해야 Phase 2.5 (leaderboard cross-reference)가 실행된다. 디폴트로는 실행하지 않음.

Pipeline Phases

Phase 1 — Paper Scan

Fetch today's trending papers from HF Hub.

hf papers ls --date today --sort trending --limit 30 --format json

Processing:

1. Parse JSON response
2. Extract paper IDs, titles, upvote counts, authors
3. Generate keyword lists from each paper title for cross-referencing
4. If focus area specified, filter papers by keyword relevance

Output: List of 30 papers with IDs, titles, keywords, and upvote counts

Phase 1.5 — Topic-Focused Model & Space Trending

Activation: Runs by default with topics LLM, video-generation. Skipped only when --skip-topics is explicitly passed.

Invoke hf-topic-radar as a composable sub-skill, running Phases 1-5 only (model scan, space scan, paper filter, dedup, scoring). Skip Phase 6-7 (report/Slack) — those are handled by this skill's own distribution phase.

topic_results = hf-topic-radar(topics=user_topics, date=date, distribute=false)

Processing:

1. Collect topic radar's trend_items[] output
2. Cross-reference with Phase 1 papers: link topic radar items to papers by author/org or keyword overlap
3. Add topic-tagged items to the main pipeline for scoring in Phase 4

Output: Additional scored items tagged with topics, merged into the main pipeline

Phase 2 — Cross-Reference (Parallel)

For each of the top 10 papers, run 3 parallel searches to measure ecosystem impact.

Subagent 1 — Model Association:

hf models ls --search "PAPER_KEYWORDS" --sort downloads -q --limit 10

Count how many models reference or implement this paper's technique.

Subagent 2 — Dataset Association:

# Use hf-dataset-viewer to check if related datasets exist
# Search by paper keywords
hf datasets ls --search "PAPER_KEYWORDS" --sort downloads -q --limit 10

Count associated datasets.

Subagent 3 — Community Buzz: For repos associated with top papers:

hf discussions list REPO_ID --status open --format json

Count open discussions as a proxy for community engagement.

Constraint: Max 4 concurrent subagents.

Output per paper: {model_count, dataset_count, discussion_count}

Phase 2.5 — Leaderboard Cross-Reference (Optional)

Activation: --leaderboard 플래그를 명시적으로 전달해야 실행된다. 디폴트로는 비활성화.

Check if hf-leaderboard-tracker has recent snapshot data in output/hf-leaderboard/. If snapshots exist and are less than 7 days old, cross-reference top papers and models with leaderboard positions.

Processing:

1. Load latest snapshots from output/hf-leaderboard/*-latest.json
2. For each top-scoring paper/model from Phase 2, check if it appears on any leaderboard
3. If found: add leaderboard rank and score to the item's metadata
4. Flag items that are both trending AND top-ranked on leaderboards as "leaderboard-validated" — these get a scoring bonus in Phase 4

Scoring bonus: Items appearing on a leaderboard get +0.10 to their final trend score (capped at 1.0).

Output: Enriched items with optional leaderboard_rank and leaderboard_score fields

Phase 3 — Web Enrichment

For the top 5 papers by upvotes, search the wider web for external signals.

Use parallel-web-search skill with queries:

• "PAPER_TITLE" site:twitter.com OR site:x.com
• "PAPER_TITLE" blog post 2026
• "PAPER_TITLE" github implementation

Processing:

1. Count external mentions per paper
2. Identify notable reactions (industry leaders, major labs)
3. Flag papers with implementations on GitHub

Output per paper: {web_mentions, github_repos, notable_reactions[]}

Phase 4 — Trend Scoring

Compute a composite score for each paper:

trend_score = (
  0.30 * normalize(paper_upvotes) +
  0.20 * normalize(model_downloads) +
  0.20 * normalize(dataset_activity) +
  0.15 * normalize(discussion_count) +
  0.15 * normalize(web_mentions)
)

See references/scoring-rubric.md for normalization rules and thresholds.

Classification:

• Score >= 0.7: HOT — likely to become mainstream within 2 weeks
• Score 0.4-0.7: WARM — emerging, worth monitoring
• Score < 0.4: COOL — early stage, low ecosystem impact

Output: Ranked list of papers with scores and classifications

Phase 5 — Intelligence Report

한국어로 구조화된 마크다운 리포트를 생성한다:

# AI 리서치 레이더 — YYYY-MM-DD

## 요약
(오늘의 트렌드 2-3문장 요약)

## 트렌딩 TOP 5

### 1. [PAPER_TITLE] — HOT 🔥
- **점수:** 0.85
- **논문:** PAPER_ID (↑ UPVOTES 추천)
- **모델:** N개, 대표: MODEL_ID (X 다운로드)
- **데이터셋:** N개
- **커뮤니티:** N개 토론 진행 중
- **외부 반응:** N개 웹 언급, M개 GitHub 구현체
- **핵심 인사이트:** (이것이 왜 중요한지 1문장 분석)

## 모델 지형 변화
(부상 중인 모델과 하락 중인 모델)

## 데이터셋 공백
(논문은 있지만 학습 데이터가 부족한 연구 영역)

## 커뮤니티 시그널
(가장 활발한 토론과 떠오르는 논쟁)

## 실행 제언
(팀을 위한 권장 사항)

Output: outputs/hf-trending/YYYY-MM-DD-radar.md (한국어)

Phase 6 — Curate

Maintain a running HF collection for monthly trends.

# Create monthly collection if it doesn't exist
COLLECTION_SLUG=$(hf collections ls --owner hyojunguy --search "trending-$(date +%Y-%m)" -q | head -1)
if [ -z "$COLLECTION_SLUG" ]; then
  hf collections create "AI Trends $(date +%Y-%m)" \
    --description "Auto-curated trending papers and models" \
    --namespace hyojunguy
fi

# Add HOT papers to the collection
hf collections add-item "$COLLECTION_SLUG" PAPER_ID paper --note "Score: 0.85 — HOT"

Output: Updated monthly collection with new HOT items

Phase 7 — Distribute

Post the report to Slack and Notion. Slack 메시지는 반드시 한글로 작성한다.

Slack Posting Identity: ALL text messages MUST be posted via scripts/slack_post_message.py (uses SLACK_USER_TOKEN) so they appear from the user, not the RandomGame Slack app. Do NOT use slack_send_message MCP tool for text. Media uploads (images, videos) use scripts/slack_upload_file.py. See .cursor/rules/slack-posting-identity.mdc.

Path B (fallback): scripts/slack_post_message.py가 없거나 SLACK_USER_TOKEN 미설정 시에만 Slack MCP(slack_send_message)로 전환한다. 이 경우 메시지가 앱 아이덴티티로 게시됨을 리포트에 명시한다. (정상 경로는 항상 python 스크립트)

# Main message (capture message_ts from stdout)
python3 scripts/slack_post_message.py --channel C0AN34G4QHK --message "{text}"

# Thread replies
python3 scripts/slack_post_message.py --channel C0AN34G4QHK --message "{text}" --thread-ts "{message_ts}"

Slack (#deep-research-trending):

1. Main message — 논문 중심 Executive Summary:

🔬 HuggingFace 트렌딩 인텔리전스 리포트 | YYYY-MM-DD
수집 논문: N편 | 분석 범위: Top 10 크로스-레퍼런스, Top 5 웹 리서치

*Executive Summary*
YYYY-MM-DD HuggingFace 트렌딩 분석 결과, N가지 핵심 트렌드 부상:
🔥 *HOT* — {HOT 카테고리 요약}
🌡️ *WARM* — {WARM 카테고리 요약}
❄️ *COOL* — {COOL 카테고리 요약}

특히 {Top1_논문} ({upvotes} upvotes, score {score})은 {1문장 인사이트}, {Top2_논문} ({upvotes} upvotes, score {score})은 {1문장 인사이트}.

2. Thread reply 1 — HOT 논문 상세: 각 HOT 논문별 점수, upvotes, 관련 모델/데이터셋 수, 웹 반응 요약

3. Thread reply 2 — 토픽 레이더 하이라이트 (Phase 1.5 결과): 토픽별 트렌딩 모델/스페이스 상위 항목

4. Thread reply 3 (optional, --leaderboard 시에만) — 리더보드 TOP 10

Notion: Use md-to-notion to publish the full report (한국어) as a Notion page under the research parent.

Output: Slack thread + Notion page URL

Output Summary

• Intelligence Report (markdown) — outputs/hf-trending/YYYY-MM-DD-radar.md
• HF Collection — monthly curated collection of HOT papers/models
• Slack Thread — in #deep-research-trending
• Notion Page — permanent archive

Domain Memory

Before running the pipeline, check outputs/hf-intel/ for today's date directory — if a report already exists, skip re-execution unless --force is specified. Cross-reference with knowledge-bases/intelligence/raw/ for recently ingested HF papers to avoid duplicate KB ingestion. Check hf-leaderboard-tracker snapshot dates to determine if leaderboard data is stale (>24h).

Anti-Gold-Plating

Fetch only the configured topics (default: LLM, multi-LLM, video-generation). Do not expand to all 50+ HF categories unless explicitly requested. Cap cross-referencing to top-10 items per source (papers, models, spaces). If the user asks for "what's trending today," return the scored report — do not auto-trigger KB ingestion or Notion publishing.

Rationalization Detection

If HF API returns empty trending lists or the cross-reference score is 0 for all items, report "NO SIGNIFICANT TRENDS DETECTED" honestly. Do not fabricate emerging trends from stale data or repackage yesterday's results as today's intelligence.

Honest Reporting

State data freshness explicitly: "Papers: {date}, Models: {N} trending, Spaces: {M} trending, Leaderboard: last snapshot {date}." If any source fails (API timeout, rate limit), report which sources succeeded and which failed rather than presenting partial data as complete.

Session Separation

• Each intelligence run is an independent session — never carry over trending data or scores from prior runs
• Re-fetch all HF API sources at the start of every run, even if the previous run was recent
• Do not assume HF CLI, Slack MCP, or Notion MCP availability from a previous session

Verification Protocol

• After the report is generated, verify the output file exists and contains data from at least 3 sources
• Cross-check the trending item count against actual API response counts
• Confirm Slack posting returned a thread timestamp before declaring distribution complete

Examples

Example 1: Default radar (논문 + 토픽 레이더)

User says: "Run daily AI radar" or /hf-trending

Actions:

1. Run Phase 1 (paper scan — 30 papers)
2. Run Phase 1.5 (topic radar: LLM + video-generation)
3. Run Phase 2 (cross-reference — top 10 papers)
4. Skip Phase 2.5 (leaderboard — 디폴트 비활성)
5. Run Phase 3 (web enrichment — top 5 papers)
6. Run Phases 4-7 (scoring, report, curate, distribute)

Result: HOT/WARM/COOL 분류된 논문 중심 리포트 + 토픽별 트렌딩 모델/스페이스, Slack thread, Notion page

Example 2: Custom topic override

User says: "HF 종합 트렌딩 — multi-LLM 중심으로" or /hf-trending topics=multi-LLM

Actions:

1. Run Phase 1 (paper scan)
2. Run Phase 1.5 (topic radar for multi-LLM only — overrides default)
3. Run Phase 2-3 (cross-reference + web enrichment)
4. Run Phases 4-7

Result: multi-modal LLM 중심 트렌드 리포트

Example 3: Leaderboard 포함 모드

User says: /hf-trending --leaderboard

Actions:

1. Run Phase 1-1.5 (papers + topic radar)
2. Run Phase 2 (cross-reference)
3. Run Phase 2.5 (leaderboard cross-reference — 명시적 활성화)
4. Run Phases 3-7

Result: 논문 + 토픽 + 리더보드 랭킹 포함 풀 리포트

Example 4: Legacy paper-only mode

User says: "AI research radar --skip-topics"

Actions:

1. Run Phase 1 (paper scan)
2. Skip Phase 1.5 and 2.5
3. Run Phases 2-7 normally

Result: 논문 전용 레이더 (v1.0 behavior)

Error Recovery

Phase	Error	Recovery
1	No papers for today	Fall back to yesterday; report "no papers today"
2	Cross-reference search fails	Continue with available data; mark missing dimensions
3	Web search rate-limited	Skip web enrichment; score with 4 dimensions only
4	All scores below threshold	Report "quiet day" with top papers by upvotes only
6	Collection create fails	Continue without curation; log error
7	Slack post fails	Save report locally; retry in next run
7	Notion upload fails	Save report locally; manual upload later

Subagent Contract

When spawning Task tool subagents:

• Always pass absolute file paths — subagent working directories are unpredictable
• Share only load-bearing code snippets — omit boilerplate the subagent can discover itself
• Require subagents to return: { status, file, summary } — not full analysis text
• Include a purpose statement in every subagent prompt: "You are a subagent whose job is to [specific goal]"
• Never say "do everything" — list the 3-5 specific outputs expected

Timeout & Retry Policy:

Subagent	Timeout	Retry	Fallback
Default	120s	1 retry with backoff	Log error + skip non-blocking; abort if blocking
Long-running (research, batch)	300s	0	Partial result if available; else mark PARTIAL