How to make your AI watch the sources you follow

To make your AI watch the sources you follow, give it a list of places to read, have it read them every hour, and have it weigh every new item against what you are pursuing. Only the items that move a pursuit reach you, as a decision, and everything else is set aside with one sentence about why.

We built this in Trillion as "Watch my sources", and we built it because it was missing. On 2026-10-08 I had to send my AI a video myself, and nothing in my system had noticed the video existed.

The video I had to send myself

OpenAI had just posted a video called "Build Plugins for ChatGPT". In the video, ChatGPT plugins get a sidebar home and a side panel beside the conversation, and the plugins are built on MCP apps. MCP is an open standard that lets an AI app connect to outside tools and data, and an MCP app lets a tool show its own screen inside the AI conversation.

That is a change in the world that touches the work I do. I only knew about the change because I saw the video, and then I had to hand the video to my AI by hand.

I was not happy about that, and I said so: "the fact I have to be the one to share this YouTube video ... and this was not caught by our system."

Look at what that sentence says. My AI knows my goals, my files and my decisions. The AI could not tell me that OpenAI had just changed where plugins live, because the AI never saw the announcement. So why did the whole system miss it?

Why nothing caught the video

Trillion reads what reaches the person: their notes, their voice memos, their email and their chats. If something happens in your life, Trillion can see it. If something happens in the world and never lands in one of those places, Trillion has nothing to read.

Nothing in Trillion read public sources at all. A change in the world never arrived to be weighed, so there was nothing to weigh.

We did have one feed reader. That reader only understood JSON, and most of the web's feeds are RSS or Atom. A reader that understands one format of feed is a reader that misses most feeds, so even the single reader we had missed most of what it was pointed at.

The fix was to give Trillion a place where the sources you follow live, and a tool that reads all of them. That is where "Watch my sources" comes in.

What Watch my sources does

The places you follow are rows of a list called Sources I watch. A row can be any blog or news feed, a YouTube channel, a changelog, a docs page, Hacker News or Product Hunt. X accounts can be rows too, and they need a feed address, because X has no public feed.

One tool reads every row in the list every hour. Every new item it finds becomes a row in a list called Signals, so a new blog post, a new video or a new changelog entry each get a row of their own.

The first read of a new source is quiet. The tool reads the source and remembers what was already there, and it does not hand you the whole back catalog as if the back catalog were news. Only items that show up after that first read count as new.

Say you add a changelog today. Today it fills in quietly. Next week, when the changelog gets a new entry, that entry becomes a row in Signals. So now the changelog is being watched, and the next question is who decides whether the new entry matters.

How each new item is weighed, and what it costs

Your AI decides. Each new item is weighed against two things: what you are pursuing, and the content formats you make.

The weighing also looks for what I call unrelated dots. An unrelated dot is an idea from a different field that applies to one of your pursuits. Here is an example. Say a baker follows a blog about shipping software. A post about releasing small changes often has nothing to do with bread, but the baker can use the same idea to launch one new flavor at a time. A filter that only matches keywords would never connect the post to the bakery. An AI that knows the baker's pursuit can.

After the weighing, each item goes one of two ways. If the item moves a pursuit, it is handed to your AIs as a decision. Everything else is set aside, with one sentence that says why it was set aside. You hear the outcome of the weighing. You do not hear the feed.

The cost stays small for two reasons. A source that has not changed is not read again, because the tool only asks the source whether anything changed. And each run weighs all of the new items in one question to your AI, instead of one question for every item.

So the reading is cheap and the weighing is cheap. What is left is the choice of how to build the whole thing.

Build the mechanism, not the instance

I could have solved the missing video by writing a watcher for YouTube and a watcher for OpenAI. That would have caught that video and nothing else, and every new source would have meant new code.

We built the generic mechanism instead. The sources are data, and no list of sites is written into the product. When you want to watch a new place, you add a row to Sources I watch, and the product stays exactly the same.

The same idea runs through everything we make. A tool that solves one instance of a problem gets rebuilt for the next instance. A tool that solves the mechanism gets a new row.

Watch my sources is in the Trillion Marketplace as a workflow that anyone can run for their own sources and pursuits. The workflow starts by reading what you already follow, which means your saved links, your likes and your subscriptions, so you do not type a list from scratch.

The first thing it caught by itself was a Hacker News post that put the new "decision" models to a test, including the decisions endpoint OpenAI had just launched. The post reached the front page after Trillion started watching, my AI weighed it as a change in how buyers will hear the word decision, and it was handed to my AIs as work without me sharing anything.

How to watch your own sources

You can set this up for your own pursuits with the steps below.

1. Connect your AI to Trillion, then find the workflow "Watch the places you follow for what moves your pursuits" in the Trillion Marketplace and run it.
2. Let the workflow read what you already follow: your saved links, your likes and your subscriptions. The workflow turns what it finds into rows in Sources I watch.
3. Open Sources I watch and read the rows. Remove any source you no longer care about, and add the places you know are missing, such as a changelog or a docs page.
4. For an X account, add the account's feed address, because X has no public feed.
5. Let the first read finish quietly. The first read only remembers what is already there, so nothing arrives yet.
6. Tell your AI what you are pursuing and which content formats you make. The weighing depends on both.
7. Wait for the next hourly read, then check Signals to see the new items and the decisions your AIs received.

After the first week, look at what reached you. If a decision showed up that you would have missed, the watch is working. If an item you wanted was set aside, read the one sentence that explains why, then adjust your pursuits or your sources.

The short version

• Trillion read what reaches the person, and nothing read public sources, so a change in the world never arrived to be weighed.
• Sources I watch is a list of rows: blogs, news feeds, YouTube channels, changelogs, docs pages, Hacker News, Product Hunt, and X accounts with a feed address.
• One tool reads the list every hour, and every new item becomes a row in Signals.
• Your AI weighs each new item against your pursuits and your content formats, including unrelated dots, and only what moves a pursuit reaches you as a decision.
• The sources are data, so adding a place to watch never changes the product.