<aside>

Before

I spend 30+ mins per day reading Transformer, SO MUCH Zvi, news.smol.ai, Peter Wildeford, LessWrong, X, Anthropic/OpenAI/Deepmind releases, and half a dozen other substacks. And still get sent things by my coworkers that I somehow missed.

</aside>

<aside>

After

Every morning, a Claude Opus 4.5-based agent, prompted with my preferred sources and specifics of what I want to hear about, does the human equivalent of hours of research to write and send a single, concise, customized newsletter.

</aside>

github.com/jesenator/ai-newsletter

<aside> ⚙

How it works

1. Initiates at 7 am every morning - using macOS LaunchAgent
2. Builds the initial prompt
   1. Fetches recent posts from configured RSS feeds (all Substacks have RSS feeds!)
   2. Scrapes non-RSS sources (Hacker News, Anthropic News, etc.)
   3. Loads recent newsletters to avoid duplication
   4. Adds detailed input from the user about what they want to know about. Mine includes things like prominent AI model releases, new AI safety research, updates to productivity tools I use (Notion, Raycast, Cursor, etc.), important AI policy updates, etc.
3. Passes everything to an AI agent
   • Uses OpenAI Agent framework with Claude Opus 4.5 (costs ~$3-4 per day, which I happily spend)
   • Has various tools (Google search, web scraping, asking Perplexity)
4. The agent does more research as needed and generates a concise and styled HTML-based newsletter with highlighting, lots of linking, and dense information.
5. Saves locally and sends emails to me and other people who want it (although its optimized for me) </aside>

Claude Opus 4.5 is more than capable enough to do this - It is a perfect task for AI, and I was quite surprised no one else had done it. This forms a new AI-powered “ingestion” layer for my life.

Example newsletters

• Example 1
• Example 2

Eventually

• Make available more broadly
  • I think this would be a great service to offer people. It would have a frontend website/app where people can configure their own newsletters (sources, what you must know about, what is good for you to know about, frequency, timing, AI model, etc.).
  • I would pay for this! Let me know if you would, or even if you wouldn’t, if you would find it useful.
• Image understanding
  • Lastly, I want to add image understanding so it can decide which images are worth including. I would have wanted it to include the METR image if it had run the day after that paper was released.

Go home | Give me feedback

© Jesse Gilbert 2026

Untitled