I love my local news RSS feeds and you’ll have to pry them from my cold dead feed reader, but they are absolutely flooding me with feed items. I’ve been spending the last week or so trying to build a filtering system that would make the information flow a little more manageable.

Finally this evening I’ve hit a solution that’s reduced my set of RSS feed items (from all of the ~7000 feeds I follow) to review by between 80-90%. I went from having 1920 items in my review queue to 250!

I started by going through my feeds and filtering out as much stuff as I could by keyword. (I made a script to identify common words and word pairs in the feed titles and ran them against several days’ worth of feed items to eliminate a lot of material — celebrity names, brands, companies outside my scope, politicians I did not want to hear from — right off the top.)

That took care of about 20% of the extraneous materials but I still had a lot of useless content in my review queue. This was material I could not filter by keyword, so I decided to try AI.

I generated a few sets of feed data from different sources (university RSS feeds, local news feeds, blogs, etc.) Then I showed the titles to Claude and we talked about the different categories the feed items covered, and what was valuable vs not valuable.

That discussion turned into a text prompt for an Anthropic API call that evaluates each RSS feed item for relevance to my interests and saves it to either a skip or save JSON. I had to tweak the prompt a little but I’m very pleased at the relevance and usefulness of my much smaller queue.

I’m sure Anthropic will occasionally blow it and skip something I would want to see. But my information traps are built with fallbacks in mind. The most essential keywords and phrases that are core to the kind of information I curate — about a dozen of them — I also monitor via Google Alerts and four different search engine / news search APIs that deliver daily reports via email. I also curate small flows from Bluesky and Mastodon. If something’s important enough and Anthropic skips it, I’ll catch it somewhere else. And if I miss something less major, well, I resigned myself to imperfection years ago!

I’m confident a local LLM could do what the Anthropic API call is doing and I want to make that switch at some point, but I need a better computer to run it and that’s $$$. In the meantime evaluating about 2500 feed items means I’m paying less than 30 cents a day. (The API call is only evaluating feed titles, which aren’t much in terms of tokens.) That 30 cents is saving me at least two hours of work; I’m happy to pay it!