RSS as personal context

I’ve been thinking about RSS feeds as a way to give AI another contextual dimension about me. Context not in the technical sense of stuffing more tokens into a model (though I guess it is also that), but more so in the human sense… a better starting point for understanding who I am based on the type of content I subscribe to.

Yesterday I had the thought to ask Notion AI to use my Feedster MCP server to pull the latest articles from the RSS feeds I subscribe to and dump common themes into a Notion database. Today I took it a step further and asked it to review the latest articles in my “Technology blogs” channel and find interesting topics that seem to be bouncing around the collective tech zeitgeist, e.g. things I could write about here.

Feedster gave my agent 30 recent posts from that channel, including articles from The Pragmatic Engineer, SemiAnalysis, MacStories, Pixel Envy, Ftrain, Daring Fireball, Michael Tsai, Asymco, and a handful of other indie tech publications. Notion AI read through them, picked out common threads, and created a database with ten potential blog ideas, each tied back to its sources.

The initial list was actually decent. It included ideas about how AI-generated code is changing code review (this one was actually not very good… way too obvious), physical buttons making a comeback as interfaces for agents, age verification turning into identity infrastructure, and a post about why boring software is often the best software.

The issue here is that all of these themes felt like they could have been written by just about anyone interested in tech/AI. What if I pointed the agent to this site’s RSS feed to give it more context on articles that I had already written, so it got a real sense of who I am as a writer. It read my recent posts and colophon, figured out that I care about personal computing, designing and building developer tools, RSS, and the open web, and then reworked the list.

The first idea became this post.

Personal investment

Giving an AI agent better context usually means handing it a doc, pointing it at a folder of markdown, or writing a verbose prompt. But what does this say about the human doing the prompting? If I ask an agent to find interesting technology writing on the web, it still has to figure out what interesting means to me.

Using an RSS feed subscription has a lot of these decisions are baked into the list itself. When I subscribe to a feed, I’m making a choice about whose work is worth investing in. As I add more feeds to a channel, taste and point of view beings to emerge. When the channel fills up with more feeds around a central topic, this profile gets even more definition.

Unlike tradition web-based personalization that watch what we click, how long we linger, and what makes us come back, then build a profile we can’t really inspect or edit, a feed list is much more literal. I can tell when a channel stops being interesting to me, because reading it feels like work. I can then remove it or double down on my investment and spend time tailoring it further.

I like reading blogs

There is an uncomfortable contradiction in using AI this way. One reason I built Feedster is that search engines and AI answer engines are already giving people fewer reasons to visit individual websites. If an agent reads all of my feeds and returns a tidy summary, it could make that problem worse.

I don’t want to build a feed reader that eliminates reading. The version I find compelling is closer to a reading guide showing me what’s new, connecting ideas across sources, and giving me enough context to decide what to open with a link pointing me back to the original writer/source. With this, I want spend more time with the things most challenge me, while preserving some of the older web’s useful wandering. AI is really good at compressing messy content into a clean answer, but we still desperately need to leave room for questions, contradictions, and thought detours.

Personal context infrastructure

RSS feels a bit like Markdown to me right now. Markdown never went away, but LLMs made everyone remember why plain text is useful (human readability, software parsability, and general portability). Likewise, RSS is open, portable, and structured around links back to the original source.

As it turns out, that’s also pretty useful here. Instead of saying “go look at the entire internet,” I can say “start here.” These are the writers and publications I’ve already decided are worth my time. There are probably more links in this content worth mining for even more contextually relevant things to explore.

I’m not sure what this becomes in Feedster yet. Your own notes in the margins, better briefings, related posts, resurfacing old links — all of that seems possible. But I care more about getting the relationship right. I want the tool to make my feeds easier to explore, not quietly turn them into another inbox that feels like work to read.