A long while back NetNewsWire had an engagement score available to the scripting library. I used it to help me figure out which feeds I actually read and which I didn’t, and I’d unsubscribe from the noise. The newest incarnation of NNW doesn’t have that, and I use Unread anyhow, but I wanted to update the concept with AI, so I made a web-based RSS triage tool.

micro.blog’s recent-ish Inkwell provides an API. The app hits the API for unread articles, and I can love/hate them (along with a way to train for/against source sites, authors, keywords, and topics) or star them (in which case it hits the Inkwell API and stars the article so it’ll pop up in Unread). In addition to my direct feedback, I’ve also got a topical list that adds weight.

Borrowing from my old project, it’s also tracking the read rates on each site and using that alongside my article ratings to provide an overall score for websites that it takes into account.

So when an article comes in, a cheap Haiku agent takes a look, factors in my previous feedback and my interest list. The fun part is that the app also talks to aSystems notes, ideas, and tasks, so Haiku knows to look for relevance there. Likewise, as AI development articles come in, especially ones about “human-in-the-loop” and ML, it’s looking at my aSystem feature backlog and noting the relevance.

I added a “no, that’s not quite right” feedback loop, as well: Since the agent records its impressions on a given post, I can nudge it. For instance, it downgraded an Oregonian article about a housing policy story because it wasn’t relevant to my interests list, which includes homelessness. I added a “that’s not quite right” feedback, and when an article came in later today, the agent’s take reflected that new nudge:

Relates to user’s Portland interest and touches on homelessness/housing policy context (ICE enforcement affects vulnerable populations), but is primarily a crime/legal story which the user has rejected. Maxine Bernstein is a crime reporter whose work the user may want to filter.

It’s true: I’ve added crime stories as negative signal, and it has picked up on the reporter being associated with those, suggesting I just add her to the filter. The article came in at a score of 0.59, though, which represents a strong-ish score, so my countervailing guidance to be a bit more inclusive worked. It has likewise done a good job with downscoring roundup articles, advice columnists, and headlines that don’t reveal key details. In other words, it is doing a great job getting me to the three Oregonian articles I care to read on any given day.