morning-ai（早上艾）

morning-ai: AI News Daily Report Generator

Permissions overview: Collects public data from Reddit, Hacker News, GitHub, HuggingFace, arXiv, and X/Twitter. Requires optional API keys configured in .env or ~/.config/morning-ai/.env. Writes report files to the current working directory. See Configuration for details.

Track 80+ AI entities across 6 sources. Collect updates from the past 24 hours, score and deduplicate them, and generate a structured Markdown daily report. Covers 4 types: Product (feature launches, version releases), Model (new models, open-source weights), Benchmark (leaderboard changes, papers), Funding (rounds, acquisitions, milestones).

Step 0: Configuration Gate (REQUIRED — must complete before any other step)

Run this command FIRST before doing anything else:

if [ -f "$HOME/.config/morning-ai/.env" ] || [ -f ".claude/morning-ai.env" ] || [ -f ".env" ]; then echo "CONFIG_STATUS=READY"; else echo "CONFIG_STATUS=MISSING"; fi

Branch on the output:

• If output is CONFIG_STATUS=READY — read the config file, report which sources are active, then proceed to Step 1.
• If output is CONFIG_STATUS=MISSING — STOP. You MUST complete the First-Time Onboarding below before proceeding to Step 1.

First-Time Onboarding (when MISSING)

CRITICAL: STOP HERE. You MUST complete all onboarding steps below interactively with the user. Do NOT run Step 1 (data collection) until a config file exists and the gate check returns READY. Running data collection without configuration will produce incomplete results.

Walk the user through setup interactively, waiting for their response at each step:

1. Welcome — briefly explain what morning-ai does: tracks 80+ AI entities across 6 sources (Reddit, Hacker News, GitHub, HuggingFace, arXiv, X/Twitter), generates scored daily reports
2. Show what works for free — 6 sources (5 need no API keys, 1 optional):

- Reddit (public JSON), Hacker News (Algolia API), HuggingFace (public API), arXiv (public API), X/Twitter (web search) - GitHub (public API, optional GITHUB_TOKEN for higher rate limits)

1. Ask the user if they want to enable GitHub with higher rate limits:

1. Ask about infographics (optional):

1. Ask about social content distribution (optional):

- Enable social content generation? Set SOCIAL_ENABLED=true - Which platforms? X (Twitter), Xiaohongshu (Little Red Book), or both - For advanced multi-account/multi-style setup, create ~/.config/morning-ai/social_channels.json (see skills/gen-social/SKILL.md for schema). For quick single-channel setup, just set SOCIAL_PLATFORM, SOCIAL_STYLE, and SOCIAL_LANG env vars.

1. Ask about message digest (optional):

- Enable concise message digest for sharing on messaging platforms (WeChat, Telegram, Slack)? - If yes: set MESSAGE_ENABLED=true - Optional settings: MESSAGE_MIN_SCORE (default 5), MESSAGE_MAX_ITEMS (default 10), MESSAGE_LINKS (bottom or inline)

1. Create the config file — collect the keys the user provides and write them to ~/.config/morning-ai/.env in KEY=value format (one per line). Create the directory if needed: mkdir -p ~/.config/morning-ai
2. Confirm — show how many sources are now active (N/9)
3. Verify — re-run the gate check to confirm CONFIG_STATUS=READY:

   if [ -f "$HOME/.config/morning-ai/.env" ] || [ -f ".claude/morning-ai.env" ] || [ -f ".env" ]; then echo "CONFIG_STATUS=READY"; else echo "CONFIG_STATUS=MISSING"; fi

Only proceed to Step 1 if the output is READY.

1. If the user wants to skip API key setup and use only free sources, create a minimal config file first, then proceed to Step 1:

   mkdir -p ~/.config/morning-ai && echo "# morning-ai config — free sources only" > ~/.config/morning-ai/.env

Language

Rules:

• Default is English. Unless --lang is explicitly specified, the report MUST be written entirely in English. All report text — titles, summaries, section headers, table labels, bullet points, "Why It Matters" analysis, and all other human-readable content — must be in English.
• If --lang is specified, use that language for all human-readable content instead.
• Entity names are proper nouns (OpenAI, DeepSeek, Midjourney, etc.) — keep them as-is regardless of language.
• When source data is in a different language than the target (e.g. Chinese source → English report), translate it into the target language during report generation. Do NOT leave untranslated fragments.
• The --lang setting also applies to infographic prompt content (see Step 4).

Prerequisite: Step 0 must have returned CONFIG_STATUS=READY. If you have not completed Step 0, go back and run it now.

Step 1: Data Collection

Run the Python collector to gather data from automated sources:

cd {SKILL_DIR} && python3 skills/tracking-list/scripts/collect.py --date {YYYY-MM-DD} --depth default -o {CWD}/data_{YYYY-MM-DD}.json

Parameters:

• --date: Target date, default today (YYYY-MM-DD)
• --depth: Collection depth — quick (fast, fewer results), default, or deep (comprehensive)
• --sources: Specific sources only, e.g. --sources reddit hackernews github
• -o: Output JSON file path

What it does:

• Runs 5 collectors concurrently (Reddit, HN, GitHub, HuggingFace, arXiv)
• Time window: [Yesterday 08:00, Today 08:00) UTC+8
• Pipeline: collect → score (1-10) → deduplicate → cross-source link → verification bonus
• Returns structured JSON with all items, stats, and collection metadata

Timeout: Allow up to 3 minutes for default depth, 5 minutes for deep.

If the user provides --exclude types (e.g. --exclude Funding), note which types to filter out in Step 3.

X/Twitter Search

After the automated collection completes, use web search to discover recent X/Twitter updates from tracked entities. The tracked X handles are listed in {SKILL_DIR}/lib/entities.py under X_HANDLES.

Search Strategy: Multi-Layer Account Checking

Search X/Twitter in three layers, in priority order:

Layer 1 — Official Accounts (highest priority): Search for recent posts from official company/product accounts. Handles are listed in entity files under {SKILL_DIR}/entities/.

• Focus: Model releases, product launches, API updates, pricing changes

Layer 2 — CEO / Core Personnel Accounts: Check key people's accounts for announcements, previews, and context that official accounts may not cover. Listed as "Key People" in each entity file.

• Focus: Early previews, strategic context, technical details, competitive commentary

Layer 3 — KOLs & Benchmark Institutions: Check AI opinion leaders and evaluation accounts for independent analysis, benchmark results, and trending discoveries. See {SKILL_DIR}/entities/kol.md and {SKILL_DIR}/entities/benchmarks-academic.md.

• Focus: Paper highlights, benchmark rankings, community trends, independent testing

Search Execution

For each search depth:

Use web search queries like:

• site:x.com @{handle} since:{yesterday} — for specific account posts
• site:x.com "{entity name}" AI announcement — for broader discovery
• site:x.com AI model release OR benchmark OR open-source {date} — for trending AI news

RT/Quote Tweet Handling

When a discovered post is a retweet (RT) or quote tweet:

1. Trace to the original post — the event timestamp is the original post time, not the RT/quote time
2. Validate the original post time falls within the 24-hour collection window [Yesterday 08:00, Today 08:00) UTC+8
3. Use the original post URL as the source_url, not the RT/quote URL
4. Credit the original author in source_label (e.g., "@AnthropicAI on X (via @karpathy RT)")

Timeliness Validation

• All X post times must be converted to UTC+8 for window validation
• The collection window is [Yesterday 08:00, Today 08:00) UTC+8
• Posts outside this window are rejected — do NOT include them regardless of relevance
• For RT/quote tweets, validate the original post time, not the retweet time

Source Priority for X-Sourced Items

Items sourced only from KOL accounts (Priority 7) should be scored conservatively and flagged for cross-verification with an official source.

Incorporating X Findings

For each verified X/Twitter update:

1. Create a TrackerItem with source: "x", source_url pointing to the original tweet, and source_label as "@{handle} on X"
2. If the same event was already found by an automated collector (Reddit, HN, GitHub, etc.), merge it as a cross_ref rather than creating a duplicate — this strengthens the verification score
3. If it's a genuinely new finding not in the automated data, add it as a new item with appropriate scoring

Step 2: Read Specifications

After data collection completes, read the tracking specification to understand scoring criteria, record format, and timeliness rules:

Read {SKILL_DIR}/skills/tracking-list/SKILL.md

This specification defines:

• 4 tracking types (Product / Model / Benchmark / Funding) with include/exclude criteria
• Source priority rankings
• Scoring criteria (1-10 scale with 5 dimensions)
• Timeliness validation rules (event date ≠ page date)
• Cross-verification requirements (7+ scores need 2+ independent sources)
• Record format for the report

Internalize the specification before writing the report. Pay special attention to the scoring reference tables and type classification guide.

Step 3: Generate Report

1. Read the JSON output from Step 1
2. Read the report template: Read {SKILL_DIR}/templates/report.md
3. Generate report_{YYYY-MM-DD}.md in the working directory

Report generation rules:

• Language: Default is English. Write ALL content in English unless --lang is explicitly specified. If source data is in a different language, translate it. Entity names (proper nouns) stay as-is.
• Source links: Every item MUST include a clickable source link [Source Name](URL) pointing to the original content. This applies to all sections: TLDR, detailed entries, and compact table rows.
• Detail quality: Summary bullet points must include specific details — version numbers, percentage improvements, parameter counts, pricing, availability dates, benchmark scores. Avoid vague descriptions without concrete numbers.
• Factual verification: See skills/tracking-list/SKILL.md → "Factual Detail Verification" for the full protocol. Never write a number from memory or inference — omit unverifiable details.
• Filter out any excluded types (if --exclude was specified)
• Sort items by score within each type section
• TLDR section: Only items with score 7+ (across all types), sorted high to low. Each item includes a one-line summary with specifics, plus an _Impact_ sentence explaining why it matters. Must include a source link [[Source](URL)] at the end.
• Type sections: Group by score range (9-10 / 7-8 / 5-6 / 3-4)
• For items with score 7+, include multi-source verification if available
• Item format: Follow the record format defined in the tracking specification (read in Step 2), including detail level requirements, "Why It Matters", and "Key Data" sections. For mid-score (5-6) and lower-score (3-4) items, use the compact formats defined there.
• Fill in the statistics summary table

Step 4: Generate Infographics (Optional)

This step is optional. Skip if no image generation capability is available or configured.

1. Read the infographic specification:

   Read {SKILL_DIR}/skills/gen-infographic/SKILL.md

1. Generate cover + per-type sections + stitch (see Image Strategy in skills/gen-infographic/SKILL.md):

- Always generate cover (9:16 portrait) + per-type section images (9:16 portrait), then stitch into one long image

1. Cover image: Sort by score and select the top 4-5 updates (across all types). Build prompt using the Cover Prompt Template (9:16 portrait).

1. Per-type section images: For each type (Model/Product/Benchmark/Funding) with 7+ score items, build a prompt using the Per-Type Prompt Template (9:16 portrait).

- Default (IMAGE_GEN_TYPES=auto): only types with 7+ score items - Set IMAGE_GEN_TYPES=all for all types, none for cover only

1. Generate images and stitch:

Option A — Native tool (Claude Code or other tools with built-in image generation): Use your tool's built-in image generation capability, one call per image. Then stitch sections together.

Option B — Python script batch mode (any environment, requires IMAGE_GEN_PROVIDER configured): Build a manifest JSON with all prompts and outputs, then run:

   cd {SKILL_DIR} && python3 skills/gen-infographic/scripts/gen_infographic.py --batch {CWD}/manifest.json --stitch

Supported providers: gemini, minimax. See Configuration for API keys. Requires pip install Pillow.

The final output is news_infographic_YYYY-MM-DD_combined.png — a single long image containing cover + all section images.

1. Insert images into the report:

- Combined long image at the beginning - Individual per-type images optionally at the top of each type section

Step 5: Generate Social Content (Optional)

Skip this step if SOCIAL_ENABLED is not true or no social channels are configured.

Generate platform-optimized copy and images for social media distribution (X, Xiaohongshu, etc.).

1. Read the social content specification:

   Read {SKILL_DIR}/skills/gen-social/SKILL.md

1. Load channel configuration:

- If SOCIAL_CHANNELS_FILE exists → read the JSON channel list - Else if SOCIAL_PLATFORM env var is set → build a single channel from SOCIAL_PLATFORM + SOCIAL_STYLE + SOCIAL_LANG - Else → skip this step

1. For each channel:

a. Read the channel's template: {SKILL_DIR}/skills/gen-social/templates/{platform}/{style}.md b. Select top items from the report data (filter by min_score, limit by items, translate if lang differs from source) c. Generate copy following the template's format rules, tone, and character limits d. Validate character counts — each tweet ≤ 280 chars, Xiaohongshu title ≤ 20 chars, body ≤ 1000 chars e. Write copy to {CWD}/social/social_{YYYY-MM-DD}_{channel_id}.md f. If channel has image: true — generate platform-adapted images using the same providers as Step 4 - X: 16:9 or 1:1 aspect ratio - Xiaohongshu: 3:4 portrait, carousel multi-image supported g. Write images to {CWD}/social/social_{YYYY-MM-DD}_{channel_id}_{N}.png

1. Write manifest to {CWD}/social/social_{YYYY-MM-DD}_manifest.json listing all generated files

Channel config examples: See skills/gen-social/SKILL.md for the full JSON schema and quick-setup env vars.

Step 6: Generate Message Digest (Optional)

Skip this step if MESSAGE_ENABLED is not true.

Generate a concise, share-friendly message digest suitable for messaging platforms (WeChat, Telegram, Slack, etc.). The digest provides bold titles with one-line summaries and reference links — optimized for copy-paste sharing.

1. Read the message specification:

   Read {SKILL_DIR}/skills/gen-message/SKILL.md

1. Read the digest template:

   Read {SKILL_DIR}/skills/gen-message/templates/digest.md

1. Select items from the report data (data_{YYYY-MM-DD}.json):

- Filter by score >= MESSAGE_MIN_SCORE (default: 5) - Sort by importance score descending - Limit to MESSAGE_MAX_ITEMS (default: 10) - Use MESSAGE_LANG for language (default: from --lang)

1. Generate text digest following the template format:

- Write to {CWD}/message_{YYYY-MM-DD}.md - Each item: emoji marker + bold title + one-line summary + source link - Source link (🔗 URL) after each item by default (or grouped at bottom if MESSAGE_LINKS=bottom) - Language-specific header and footer

1. If image generation is available (IMAGE_GEN_PROVIDER is configured):

- Build a 9:16 portrait image prompt from the specification's image prompt template - Generate {CWD}/message_{YYYY-MM-DD}.png using the same image generation method as Step 4 (native tool or Python script)

Output files:

• message_{YYYY-MM-DD}.md — copy-paste text for messaging
• message_{YYYY-MM-DD}.png — accompanying image (only if image generation is configured)

Entity Reference

The entities/ directory contains detailed entity registries organized by tracking group:

Each file lists X/Twitter accounts, key people, official blogs, changelogs, GitHub repos, and other source URLs for every tracked entity. Read these files when you need to verify or supplement the automated collection.

Custom Entities

Users can add their own tracked entities by placing markdown files in entities/custom/ (or ~/.config/morning-ai/entities/, or a path set via CUSTOM_ENTITIES_DIR). Custom entity files use a simplified format — see entities/custom-example.md for the template. Custom entities are automatically merged into the built-in registries at runtime and collected alongside the default 80+ entities.

Scheduling

Morning-AI is designed for daily automated execution. Each run produces date-stamped files (report_YYYY-MM-DD.md, data_YYYY-MM-DD.json), making it safe to run on a recurring schedule.

Schedule Configuration

Use --schedule to set a custom cron expression (default: 0 8 * * *):

The schedule is passed to the agent's native scheduler (CronCreate, /loop, system cron, etc.). Morning-AI itself does not run a scheduler — it relies on the host agent or system to trigger runs.

Unattended Behavior

• Idempotent: Re-running on the same date overwrites previous output — no duplicate accumulation
• No interactive prompts: All steps run without user input when API keys are configured
• Partial success: If some sources fail, the report generates with available data and logs warnings
• Timeout: Allow 3 min (default depth) or 5 min (deep)

Agent Integration Examples

Claude Code (CronCreate / loop):

/loop 24h /morning-ai

With custom schedule:

/morning-ai --schedule "0 9 * * 1-5"

System cron (manual setup):

0 8 * * * cd /path/to/workspace && claude -p "/morning-ai"

OpenClaw / always-on bot:

schedule: "0 8 * * *"
skill: morning-ai

Configuration

Config File Locations (priority order)

1. Environment variables (highest priority)
2. Project config: .env in skill directory
3. Global config: ~/.config/morning-ai/.env

Config File Format

# ~/.config/morning-ai/.env
GITHUB_TOKEN=ghp_xxx

Free Sources (no API key needed)

Message Digest Configuration

See skills/gen-message/SKILL.md for message digest configuration variables (MESSAGE_ENABLED, MESSAGE_MIN_SCORE, MESSAGE_MAX_ITEMS, etc.).

Security & Permissions

• Data access: Reads public web/platform data only. No private or authenticated content is accessed.
• API keys: Stored locally in .env files. Never transmitted except to their respective APIs.
• File writes: Only writes report files (report_*.md, data_*.json), message digest files (message_*.md, message_*.png), and cache files to the skill/working directory.
• Network: Outbound HTTP/HTTPS requests to public APIs (Reddit, GitHub, etc.). No inbound connections.
• No telemetry: No usage data is collected or sent anywhere.