🤖 Daily AI Signal

A daily AI-signal digest. It collects high-signal AI posts from Reddit, filters out noise, ranks by practical usefulness, summarizes the best ones with the Gemini API, and publishes a clean Markdown digest to a local output/ folder and to a Daily AI Signal page in Notion.

Built as a sibling to Daily-Youtube-Digest and Daily-News-Digest, reusing the same Gemini + Notion conventions.

What it does

1. Collect — pulls top posts from configured subreddits (Reddit public JSON API, no auth).
2. Filter & rank — matches posts to your AI topics, drops noise, scores usefulness.
3. Summarize — Gemini writes a concise, per-topic Markdown section.
4. Monitor — checks Artificial Analysis for newly published articles and summarizes what changed.
5. Render — assembles a dated Markdown digest, grouped by topic.
6. Publish — saves output/<date>.md and creates a Notion child page (newest on top).

Topics tracked

Local LLMs · Open-source models · AI coding tools · AI agents · AI automation · AI app development · AI productivity & business · AI hardware & edge · Practical AI tools

Quick start

npm install
cp .env.example .env
# Fill in GEMINI_API_KEY and NOTION_TOKEN (reuse from your other digest projects).

Create the Notion page (one time)

Point the setup script at an existing digest page (e.g. your YouTube Digest page). It creates a Daily AI Signal page as a sibling and prints the new page id:

SIBLING_PAGE_ID=<your-existing-page-id> npm run setup:notion
# Copy the printed id into .env as NOTION_PAGE_ID

If your existing page is at the workspace root, just create a page named Daily AI Signal in Notion, share it with your integration, and paste its id into NOTION_PAGE_ID.

Run

npm start          # full pipeline: collect → summarize → check articles → save → publish
npm run collect    # same, but skip Notion (writes local Markdown only)
npm test           # run the unit + integration tests (node:test, no deps)
npm run e2e:x      # end-to-end smoke test of the X source (fixture → Gemini → Markdown)

📊 Artificial Analysis monitor

Watches artificialanalysis.ai/articles — an independent AI benchmarking site — and summarizes anything newly published. The result is its own section at the top of the daily digest.

## 📊 Artificial Analysis — New Articles

> 2 new articles · 2026-07-30 → 2026-07-31

- **[DeepSeek V4 Flash 0731 scores 50 on the Intelligence Index](...)** — 10 points
  above the previous DeepSeek V4 Flash, now matching ...
  - **Why it matters:** open-weights models close the gap on ...

When nothing was published the section says so explicitly rather than disappearing, so a silent section can be told apart from a broken monitor.

How "new" is decided. The site has no RSS feed, so the monitor reads the article index for slugs and each article's OpenGraph tags for title, summary, and article:published_time. Since CI runners keep no state between runs, "new" is a date window: lookbackDays: 1 (the default) means published yesterday or today, which reports each article exactly once for a once-a-day schedule. Widen it with ARTICLES_LOOKBACK_DAYS if a scheduled run is ever missed.

Tunables live in config/articles.json. Disable with --skip-articles, DIGEST_SKIP_ARTICLES=true, or "enabled": false.

Publish dates on the site have day granularity (no clock time), so the window is measured in days rather than hours. A failed monitor never breaks the digest — the step is wrapped and the rest of the pipeline continues.

Reddit credentials are optional. Without REDDIT_CLIENT_ID / REDDIT_CLIENT_SECRET the collector uses anonymous RSS, which Reddit throttles to roughly one request per 30s per IP. The pipeline handles this by pacing requests and waiting the throttle out, so coverage is still complete — it just costs wall time (~5 min for the daily collect). With OAuth (free "script" app at reddit.com/prefs/apps) the same runs take seconds, post scores become available for ranking, and comment sentiment turns on. Subreddits that could not be reached are listed in the report.

Ranking without scores

Anonymous RSS reports no score or comment count. Those posts are not ranked as if they had zero engagement — that would push every Reddit post below every Hacker News post before content is considered. Instead they inherit the median engagement of the posts that do report metrics, so they compete on content signal. See popularityScore in src/filter.js.

X.com source

X is disabled by default to keep the pipeline free. Reddit + Hacker News + the Gemini free API tier cost nothing to run. X requires a paid data source (see Cost & budget below), so enable it only when you're ready.

X is a plugin-style collector. Enable it in config/sources.json (x.enabled = true) and choose a mode:

• mode: "grok" — use your Grok/xAI API key (XAI_API_KEY) with the built-in x_search tool. This is the easiest way to collect X posts and does not require a Twitter developer account. Note: a Grok key is NOT a Twitter/X bearer token, and the xAI API is paid (needs credits).
• mode: "api" — official X recent-search API. Needs a Twitter developer X_BEARER_TOKEN (the free Twitter tier has no search).
• mode: "playwright" — an external scraper command (X_PLAYWRIGHT_COMMAND) that prints a JSON array of posts to stdout.
• mode: "fixture" — load posts from a local JSON file (X_FIXTURE_PATH, default test/fixtures/x-sample.json). No credentials needed — ideal for offline testing, CI, and demos.

The hoursBack, minScore, and lang settings in the x config are applied uniformly across all modes.

Cost & budget

This project is designed to run for free:

• Reddit — free (public RSS, or free OAuth app credentials).
• Hacker News — free (public Firebase API).
• Gemini summarization — free API tier from Google AI Studio (GEMINI_API_KEY). Note: a Gemini Advanced consumer subscription is not the same thing — you need an API key, which has its own free tier.
• Notion publishing — free.

X is the only paid source. There is no usable free tier for X search: xAI's x_search (grok mode) needs account credits, and Twitter's free API has no search. Consumer chat subscriptions (ChatGPT Plus, Claude Pro, Gemini Advanced) do not provide API access, so they can't be used here. If you want X, the cheapest path is funding a small amount of xAI credits (check the data-sharing free-credits program at https://console.x.ai → Billing), then set x.enabled = true.

How to add X cheaply (options people actually use)

If you want X without the full Twitter API price, these are the realistic routes, roughly cheapest/most-reliable first. All of them plug into the existing playwright mode (any command that prints a JSON array of posts to stdout) with no code changes — see the normalized fields in test/fixtures/x-sample.json.

• Third-party scraper APIs (best value) — services that scrape X and return JSON, usually with a small free tier and no Twitter dev account: Apify Twitter/X actors, twitterapi.io, socialdata.tools, scrapecreators.com, or "Twitter" endpoints on RapidAPI. Wrap the call in a tiny script, set X_PLAYWRIGHT_COMMAND to run it, and set x.mode = "playwright".
• xAI/Grok x_search — already implemented as mode: "grok"; needs credits but is the least custom code.
• snscrape (free, flaky) — the open-source snscrape CLI pulls tweets with no key. Great for a manual laptop run; unreliable in an automated GitHub Action because X rate-limits/blocks CI IPs.
• RSS bridges (free, unstable) — RSSHub or a self-hosted Nitter instance can expose X accounts as RSS; most public Nitter instances are now dead, so this needs self-hosting.

Rule of thumb: for a reliable daily automated digest, a paid-but-cheap scraper API or Grok credits is what holds up; free scrapers are fine for manual runs but tend to break in CI.

GitHub Actions (Daily Automation)

This repo includes a scheduled workflow at .github/workflows/daily-ai-signal.yml.

• Schedule: daily at 03:30 America/New_York (DST-safe via 07:30 + 08:30 UTC triggers)
• Also supports manual run via Actions → Daily AI Signal → Run workflow

Required GitHub repository secrets

• GEMINI_API_KEY
• NOTION_TOKEN
• NOTION_PAGE_ID

Optional GitHub repository secrets

• REDDIT_CLIENT_ID
• REDDIT_CLIENT_SECRET
• X_BEARER_TOKEN (only for X api mode)
• XAI_API_KEY (only for X grok mode — your Grok/xAI key)

Optional GitHub repository variables

• DIGEST_TIMEZONE (default: America/New_York)
• DIGEST_LANGUAGE (default: en)
• DIGEST_SKIP_NOTION (default: false)
• GEMINI_MODEL
• GEMINI_MODEL_FALLBACKS
• GEMINI_REQUEST_TIMEOUT_MS
• GEMINI_DELAY_MS
• NOTION_ROOT_TITLE
• NOTION_VERSION
• REDDIT_USER_AGENT
• REDDIT_DELAY_MS (default: 30000 without OAuth, 600 with)
• REDDIT_RSS_RETRIES (default: 3)
• REDDIT_RSS_MAX_429_SUBS (default: 0 = never give up early)
• X_API_BASE
• X_PLAYWRIGHT_COMMAND
• X_PLAYWRIGHT_TIMEOUT_MS
• XAI_API_BASE
• XAI_MODEL
• XAI_REQUEST_TIMEOUT_MS
• ARTICLES_LOOKBACK_DAYS (default: 1)
• DIGEST_SKIP_ARTICLES (default: false)

If Reddit OAuth secrets are not provided, the workflow still runs — it just paces the RSS requests, which makes the collect step take a few minutes instead of seconds.

X is scaffolded in plugin form. To collect X posts, set config/sources.json -> x.enabled = true, then choose a mode:

• mode: "grok" + XAI_API_KEY (Grok/xAI key — recommended, no Twitter account)
• mode: "api" + X_BEARER_TOKEN (Twitter developer bearer token)
• mode: "playwright" + X_PLAYWRIGHT_COMMAND
• mode: "fixture" + X_FIXTURE_PATH (offline/CI, no credentials)

Configuration

• config/sources.json — subreddits, sort (top/hot/new), time window, per-sub limit, min score.
• config/sources.json — source settings for Reddit, Hacker News, and future X collection.
• config/topics.json — topics, their matching keywords, and max posts per topic.
• config/articles.json — Artificial Analysis monitor: base URL, lookback window, and how many article pages to check per run.

All secrets and tunables live in .env (see .env.example).

Project structure

config/
  sources.json          # source configuration (Reddit today)
  topics.json           # topics + keywords for filtering/grouping
  articles.json         # Artificial Analysis monitor config
src/
  collectors/
    reddit.js           # Reddit collector (implements the collect() contract)
    hackernews.js       # Hacker News collector (official Firebase API)
    x.js                # X.com collector (api | playwright plugin scaffold)
  monitors/
    artificial-analysis.js  # watches artificialanalysis.ai/articles for new posts
  filter.js             # relevance filter, ranking, topic grouping
  summarize.js          # Gemini summarization (raw HTTP + model fallback)
  render-markdown.js    # builds and saves the Markdown digest
  notion.js             # Notion REST helpers (markdown → blocks, publish)
  index.js              # pipeline orchestrator
scripts/
  setup-notion.js       # one-time: create the Daily AI Signal parent page
output/                 # generated digests: <date>.md

Adding a new source (X, GitHub, RSS, ...)

1. Create src/collectors/<name>.js that exports async collect(config) returning the normalized post shape documented in reddit.js.
2. Add a <name> section to config/sources.json.
3. Register it in the COLLECTORS map in src/index.js.

No other code needs to change — filtering, summarizing, rendering, and publishing all operate on the normalized post shape.

Roadmap

• Reddit collection (MVP)
• Gemini summarization + Notion publishing
• Artificial Analysis article monitor
• LLM-based relevance scoring (replace keyword matching)
• More monitored publications (blogs, changelogs, release notes)
• X.com / GitHub trending / RSS / newsletter sources
• Scheduled daily run (GitHub Actions / cron)
• Email / webhook / Slack delivery