Ai digest

AI Digest

Weekly intelligence briefing covering AI, development, and Agent ecosystems. Produces a Chinese-language digest that answers:

1. What happened this week? — Trending topics, frameworks, tools, and breakthroughs across all major sources.
2. What's worth paying attention to? — Curated picks with analyst-style commentary (pros, cons, who benefits).
3. What are people arguing about? — Industry viewpoints, debates, and contrarian takes.

The output is designed to be directly shareable — read in terminal, save as Markdown, or paste into team channels.

Scope: This skill is for weekly intelligence gathering and analysis. It discovers unknown unknowns through curated discovery, not just keyword search.

---

Prerequisites

OpenCLI (Required)

This skill uses OpenCLI (14.8K+ stars) as its primary data acquisition engine. OpenCLI provides CLI access to 79+ websites including Twitter/X, Hacker News, Reddit, and more.

Install OpenCLI:

# Install globally
npm install -g @jackwener/opencli

# Install Browser Bridge extension (required for Twitter/X)
# 1. Download opencli-extension.zip from https://github.com/jackwener/opencli/releases
# 2. Unzip and load in chrome://extensions (enable Developer mode first)
# 3. Ensure Chrome is running and logged into Twitter/X

# Verify setup
opencli doctor

For OpenCLI Skill integration (recommended):

npx skills add joeseesun/qiaomu-opencli-skills --skill qiaomu-opencli-usage

Headless / Server Environments:

# Install Xvfb (Debian/Ubuntu)
sudo apt-get install -y xvfb

# Option A: xvfb-run prefix
xvfb-run opencli hackernews top --limit 10 -f json

# Option B: Persistent virtual display
Xvfb :99 -screen 0 1280x1024x24 &
export DISPLAY=:99
opencli twitter trending -f json   # No prefix needed

Note: Public API commands (hackernews, devto, lobsters, stackoverflow) work without browser setup. Headless mode with Xvfb covers reddit and most browser-backed commands. Twitter/X requires a logged-in Chrome session.

---

Workflow

Step 0: Pre-flight Check & Memory

0a. OpenCLI Health Check (MUST pass before proceeding):

# Auto-detect headless mode
if [ -z "$DISPLAY" ]; then
  OPENCLI_CMD="xvfb-run opencli"
  which xvfb-run || { echo "WARNING: No display and xvfb-run not found."; OPENCLI_CMD="opencli"; }
else
  OPENCLI_CMD="opencli"
fi

# Verify OpenCLI
which opencli || { echo "ERROR: OpenCLI not installed."; exit 1; }
$OPENCLI_CMD doctor

# Test public API (should always work)
$OPENCLI_CMD hackernews top --limit 1 -f json

# Test browser-backed commands (requires Chrome + extension + login)
$OPENCLI_CMD twitter trending -f json    # Requires Chrome + X login
$OPENCLI_CMD reddit hot --limit 5 -f json  # Works with Xvfb

If browser-backed commands fail (exit code 69 or 77):

• Public API commands still work without browser
• Reddit works with Xvfb
• Twitter/X requires a logged-in Chrome session — proceed without X data if unavailable

0b. Read Previous Digests (Memory):

1. Scan current working directory for ai-digest-*.md files
2. Read all found digest files to extract previously covered topics
3. Build a Coverage Memory — list of topics, projects, and URLs already reported
4. During Steps 1-3, flag overlapping topics:
   • Identical topic → SKIP unless significant new development this week
   • Similar topic → Mark as 🔄 Update, write as supplement referencing prior digest date
   • New topic → Proceed normally

Update format (see references/output-template.md for full template):

🔄 Update from Week 2026-04-07:
{Title} 有了新进展——{1-2 句说明新变化}。
上期点评认为 {回顾上次判断}，本期 {确认/推翻/延伸}：

Memory Rules:

• Read ALL previous digest files, not just the most recent
• Match by project name, key phrase, or URL domain — not just exact title match
• If no previous digests exist, proceed normally

Step 1: Layer 1 — Curated Discovery

Purpose: Discover topics you didn't know to search for, by leveraging human curators.

Fetch from curated newsletters and aggregators using mcp__exa__web_search_exa. See references/search-queries.md for all query templates and the curated source table.

Key behavior:

• Extract all mentioned topics, projects, and debates as seed signals
• If a topic appears in 2+ curated sources → flag as high-confidence signal
• Use freshness:week parameter to focus on recent content

Step 2: Layer 2 — Platform Hotness (via OpenCLI)

Purpose: Validate seed signals with community-driven popularity data, and catch topics that curators missed.

Primary tool: OpenCLI. Use it for ALL platform data where a built-in adapter exists.

Fetch from community platforms in parallel. See references/search-queries.md for all OpenCLI commands and fallback methods, and references/source-tiers.md for platform → command mapping.

Key platforms:

• Hacker News — $OPENCLI_CMD hackernews top/best/new
• X / Twitter — $OPENCLI_CMD twitter trending/search/timeline
• Reddit — $OPENCLI_CMD reddit hot/subreddit
• GitHub Trending — $OPENCLI_CMD github trending (if plugin installed)
• Product Hunt — exa search only (no OpenCLI adapter)

X/Twitter content processing rules (see references/source-tiers.md):

• Default quality: Tier 2
• Verified KOL announcements: Tier 1
• Unverified speculation: Tier 3
• Always find original source link rather than X post alone

Nitter RSS fallback (only when OpenCLI browser commands fail): See references/source-tiers.md for instance priorities.

Step 3: Layer 3 — Domain Deep Dive

Purpose: Ensure key domains are never missed, even if they don't trend.

Use semantically expanded keyword searches with mcp__exa__web_search_exa. See references/search-queries.md for all domain-specific query templates and semantic expansion rules.

Three domains to deep-dive:

1. Agent & Automation — MCP, multi-agent, orchestration, agentic workflow
2. Developer Tools & Open Source — AI coding assistants, new releases, SDK launches
3. Industry Viewpoints & Analysis — debates, safety discussions, expert commentary

Step 4: Synthesize & Rank

Deduplication:

• Merge the same topic across sources (e.g., GitHub trending + HN front page + Reddit = one entry)
• Keep the highest-quality source link as primary, reference others as cross-validation

Scoring: See references/scoring-algorithm.md for the full formula.

Score = (frequency × 0.25) + (source_quality × 0.30) + (recency × 0.20) + (cross_validation × 0.25)

Topic Selection:

Section	Count	Criteria
🔥 Top 5	5	Highest scores across all categories
🤖 Agent & Automation	5-8	Top from Agent domain
🛠 Dev Tools & Open Source	5-8	Top from Dev Tool domain
💡 Important Viewpoints	3-5	Top opinion/analysis pieces
⚡ Quick News	3-8	One-liner updates
📖 Extended Reading	2-3	Long-read recommendations

Step 5: Output & Save

Commentary Style: Every notable item must include 5-section analyst-style commentary. See references/commentary-style.md for the full template and rules.

The 5 mandatory sections:

1. 背景与意义 — Why this matters now, what shift it signals
2. 谁该关注 — Who should care and in what scenario
3. 对比与定位 — How it compares to alternatives, what gap it fills
4. 硬伤与风险 — Specific limitations with data/version numbers
5. 趋势判断 — Likely trajectory and recommended action

Every item must also include 👍 (值得点) and 👎 (需注意) with verifiable specifics.

Output Format: See references/output-template.md for the complete Markdown template.

File Saving:

1. Save the full digest as Markdown to the current working directory
2. Filename: ai-digest-{YYYY-MM-DD}.md (Monday's date of current week)
3. Confirm the file path to the user

Follow-up Options (offer after saving):

1. 深挖 — Choose a specific news item for deeper analysis
2. 对比 — Compare 2-3 projects/frameworks
3. 搜索 — Deep-search a specific topic
4. 分享 — Save to Feishu docs or send to Feishu group

---

Rules

Search Rules

• OpenCLI is the primary data acquisition tool — use it first for all supported platforms
• Use mcp__exa__web_search_exa for: curated newsletter discovery, Product Hunt, GitHub Trending, anything OpenCLI doesn't cover
• Use fetch_fetch/webfetch only for Nitter RSS fallback
• Use HN Firebase API as fallback when OpenCLI is unavailable
• Search from sources in parallel for speed
• Every item must include a source link — no exceptions
• Never fabricate content — note gaps explicitly if a source is unreachable
• Time range: past 7 days from execution date
• Deduplicate aggressively — same topic across sources = one entry
• Scoring formula: frequency × 0.25 + quality × 0.30 + recency × 0.20 + cross_validation × 0.25

Memory Rules

• ALWAYS read all existing ai-digest-*.md files before starting
• NEVER repeat a topic from a previous digest unless there is significant new development
• For recurring topics, use 🔄 Update format referencing the prior digest date
• Match topics by project name, key phrase, or URL domain — not just exact title match

Commentary Rules

• Have a stance — NEVER write "既有机遇也有挑战"
• Each commentary must include ALL 5 sections (背景与意义, 谁该关注, 对比与定位, 硬伤与风险, 趋势判断)
• ALWAYS include both 👍 and 👎 with specific, verifiable points
• Comparison is mandatory — position every item against its alternatives
• Include concrete details: version numbers, star counts, test coverage %, competitor names
• If overhyped, say so directly. If undervalued, argue for why.
• See references/commentary-style.md for full rules and template

Output Rules

• Titles stay in English (original language) for traceability
• Summaries and commentary in Chinese
• Date range uses Monday-Sunday of the current week
• 🔄 Update items reference the prior digest date: "🔄 Update from Week {date}"

---

Reference Files

Detailed specifications are split into reference files under references/:

File	Contents
references/search-queries.md	All search query templates for Layers 1-3, OpenCLI commands, semantic expansion rules
references/source-tiers.md	Source quality tiers (1-3), KOL tracking lists, Nitter fallback config, X content processing rules
references/scoring-algorithm.md	Composite score formula, dimension weights, topic selection thresholds, deduplication rules
references/commentary-style.md	5-section commentary template, update format, mandatory commentary rules
references/output-template.md	Complete Markdown output template, update format, file saving rules, follow-up options