I never check RSS feeds consistently enough to make them useful. arXiv is the same (I know there’s good stuff in there but hunting for it every day doesn’t happen) so I made an LLM agent do it for me.
Every morning a GitHub Actions cron job fetches feeds from arXiv, Hacker News, LessWrong, and the Alignment Forum, dumps the last 48 hours of items into a prompt, and asks Gemini 2.5 Flash to pick out the 5–8 most relevant things for me specifically and summarize them. The result gets committed to the repo and Cloudflare Pages rebuilds the site.
The prompt does most of the real work. It describes my interests, tells the model to skip tutorials and hype pieces, and asks for a relevance score and short tags per item.
The feed is at /reading.