Content Creation

How Jacob Ben David Runs an AI News Channel with OpenClaw

Built by jacob-bd · · Source: github

Quick answer

Jacob Ben David built an open-source OpenClaw pipeline that scans RSS, Reddit, X, GitHub and web search for AI news, scores and dedupes it, and has a cheap LLM pick up to 7 stories. They say it powers Gen AI Spotlight, an automated Telegram news channel.

Result: According to the README, the pipeline powers Gen AI Spotlight, a fully automated AI news channel on Telegram, for about $5 a month without the optional twitterapi.io add-on (about $10 more).

What they built

Jacob Ben David (jacob-bd on GitHub) built openclaw-newsroom, an open-source news scanning pipeline for OpenClaw. It pulls AI news from five sources, scores and dedupes it, and has an LLM pick the stories worth posting. According to the README, it powers Gen AI Spotlight, which they describe as a fully automated AI news channel on Telegram. They credit the tech-news-digest ClawHub skill by dinstein as the inspiration.

How it works

The pipeline runs as an OpenClaw cron job. The example schedule fires every two hours from 9am to 9pm, seven runs a day. The scripts are plain Python with no pip packages and also work standalone. Each run does this:

  • Collect from five sources: 25 RSS feeds through blogwatcher, 13 AI subreddits through Reddit's public JSON API, tiered X accounts through the bird CLI, GitHub (new repos, fast-growing repos and releases from 16 key AI repos), and 5 Tavily web searches. If one source fails, the run keeps going.
  • Filter and score. An AI keyword filter trims the list, then scoring and a persistent SQLite database remove duplicates within the batch and against earlier scans. The top 50 survive.
  • Enrich the top 8 articles with full text, capped at 1,200 characters each.
  • Curate. An editor script reads an editorial profile and recent post history, then asks an LLM for up to 7 ranked picks with categories and one-sentence summaries. It tries Gemini 3.1 Flash Lite first, falls back to Grok 4.1 Fast through OpenRouter, then Gemini 3 Flash Preview.
  • Deliver through OpenClaw channels such as Telegram or Slack.

The README says a typical run goes from about 100 raw items to 3 to 7 curated picks. An optional nightly job updates the editorial profile based on approvals and rejections.

Jacob also shares a few lessons. Cap Telegram output at 5 stories with short summaries, because going over Telegram's 4,096 character limit makes delivery fail silently. Use a cron model that supports tool execution, or it may make up stories instead of running the script. They suggest GPT-5.3 Codex or Claude Sonnet 4.6 for the orchestrating agent and warn against Kimi K2.5.

The result

Jacob puts the running cost at about $5 a month. By their breakdown, Gemini Flash Lite costs $1 to $2, the OpenRouter failover costs $0 to $1, and Tavily and the GitHub API are free. Optional X keyword search through twitterapi.io adds about $10. The README gives no audience numbers for the Telegram channel.

How to build something similar

  • Start with one or two sources you already read, like a handful of RSS feeds, and get a scheduled scan posting to a private chat before adding more.
  • Do the cheap work in code first. Keyword filters and a dedupe database cut the pile down so the LLM only sees a short list.
  • Write down your editorial taste in a profile file the model reads on every run, and update it when picks miss.
  • Set a fallback model and keep messages well under your chat platform's length limit so a single failure doesn't drop a whole post.

The original post

Jacob Ben David (jacob-bd) · View on github

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases