The information overload problem

Keeping up with the AI ecosystem means watching at least a dozen sources: Hacker News, TechCrunch, The Verge, MIT Tech Review, the official blogs of OpenAI, Google DeepMind, Anthropic. Not to mention newsletters, Twitter threads, papers on arXiv.

I tried Feedly, Pocket, Readwise. They all require active attention: you have to go there, open it, scroll. Every morning turned into a mini-job before the real job. I wanted something passive: it arrives, I read it, I'm done.

The architecture in 4 steps

  1. RSS fetch: n8n reads the latest articles every morning at 7:00 from 8 preconfigured feeds (Hacker News top stories, TechCrunch AI, The Verge Tech, MIT Technology Review, OpenAI blog, Anthropic news, Google AI Blog, Ars Technica)
  2. Deduplication: Google Sheets filters out articles already included in digests from the last 7 days — so I never read the same story twice
  3. Gemini 2.5 Flash: summarizes each article in 2-3 lines in Italian, with title and original link. All the new articles in a single API call
  4. Telegram: sends the formatted digest every morning at 7:30, 30 minutes after the fetch

Why Gemini 2.5 Flash and not another model

Two practical reasons. The first is the context window: Gemini 2.5 Flash handles up to 1 million tokens. I can pass it 8-10 full articles in a single API call and get all the summaries back at once. No loop, no pagination handling.

The second is cost. With Gemini's free API tier, the cost per digest is essentially zero — less than a thousandth of a euro a day. The model has enough quality to produce readable, accurate summaries, without wasting a production-grade model on a task that doesn't need it.

The digest format

I standardized the format to make reading as fast as possible. Each entry has: title, two lines of summary in Italian, original link.

🤖 AI Digest — Monday, August 3, 2026
1. OpenAI launches o1 mini for all plans
GPT-4o level reasoning at 80% lower cost. Available today on ChatGPT Free, Plus and Team. Optimized for math and coding.
→ openai.com/blog/o1-mini
2. Apple Intelligence now speaks Italian
Siri rewritten with an internal LLM, developer beta available. Supports text rewriting, notification summaries and integration with third-party apps.
→ developer.apple.com/apple-intelligence

The fixed structure matters: the brain learns to process it quickly. After a few days you read the digest the way you read the weather — with no cognitive effort.

Personalization

This is the detail that turns the digest from generic to genuinely useful. In the prompt to Gemini I added an explicit instruction:

Prioritize news about: local and open-source AI, n8n and workflow automation, Claude and Claude Code, tools for independent developers, EdTech and AI applied to education. In case of equal relevance, choose the most recent story.

The result: the digest reflects my interests, not TechCrunch's editorial algorithm. A story about a new local open-source model outweighs a $100 million funding round for yet another LLM startup.

Real cost

  • Self-hosted n8n on a Raspberry Pi: €0/month
  • Gemini 2.5 Flash API: €0/month (generous free tier)
  • Google Sheets API: €0/month
  • Telegram bot: €0/month

The real cost is the initial setup time: about 90 minutes between configuring RSS feeds, Google credentials, and testing the Gemini prompt. After that, the system is fully autonomous.

How to replicate it

The workflow JSON is available in the public GitHub repository. The prerequisites are: self-hosted n8n (a Raspberry Pi works fine too), a free Gemini API key, access to the Google Sheets API. The workflow can be imported directly into n8n — the only manual configuration is the RSS feeds you're interested in and the personalization keywords in the prompt.

I don't read less. I read better. The difference between scrolling a feed for 20 minutes and getting 5 stories selected for me is worth, in attention, what a good spam filter is worth in money.