How I Built a Custom AI News Curator That Actually Gets My Hobbies

The Morning Scroll Trap

Every morning, I used to wake up and grab my phone. There were dozens of unread messages, a fresh row of trending hashtags, and a string of 'breaking news' alerts from every app I’d ever installed. I’d swipe through them for fifteen minutes, and then the weirdest feeling would hit: I knew a lot of stuff, but I couldn’t name one thing that truly mattered.

A phone launches, an AI model updates, a company makes a new announcement. Each one looks clickable. But after reading them all, almost nothing sticks. We used to worry about not having enough information. Now the problem is the opposite—there’s so much that it swallows not just your time, but your ability to tell what’s worth your attention.

So I decided to run an experiment: build a personal news filtering system with an AI agent, and take back control of what I see every day.

Step One: Ditch the Scraps, Plug Into a Professional Wire

My first idea was simple: set a scheduled task in an agent to collect five tech news items and send them to me at 8 a.m. That became my first briefing assistant. It worked okay—I used to jump between news sites, apps, and social media to piece together what happened overnight. Now, while making coffee, a summary was already waiting on my phone.

But the novelty wore off in days. The same product launch, written three different ways, could eat up three slots in the briefing. A story I’d already read would come back as 'latest' just because another outlet rehashed it. I couldn’t blame the agent—it was only doing what I asked. The problem was that web search hands you a bag of loose, unverified bits: official announcements, media reports, second-hand takes, reposts, and clickbait, all looking equally fresh.

To cut the echo, I had to start with the sources. Over the years, I’d collected 161 RSS subscriptions. But quantity doesn’t mean quality. If you want to stay sane in an information black hole, you need a clear filtering system.

Build a Trust Ladder

The core principle is simple: go to the source. Information gets distorted in the retelling. So I sorted my feeds by trust level:

• Primary sources: official blogs from OpenAI, Google DeepMind, Anthropic, and other companies—first-hand, untouched.
• Serious media: Bloomberg, The Information, WSJ, Reuters, Caixin—they have real reporting and cross-checking.
• Curated aggregators: The Verge, Techmeme, TechCrunch, MacRumors—they turn raw info into readable, deep content.
• Bloggers and KOLs: tech YouTubers and microbloggers—useful for hands-on impressions and unique takes.

If you don’t mind a plug, I also subscribe to ifanr and APPSO—they’re excellent sources.

Organize With a Tree Structure

When you have a hundred-plus feeds, a single list is a mess. You need a tool that lets you sort them into categories. I use Folo, which is essentially an RSS reader. RSS sounds old-school in the age of algorithmic feeds, but nothing beats it for active aggregation and control.

In Folo, I’ve organized my six core sections: tech, gaming, culture, AI, and autos. Opening it feels like flipping through a custom magazine I’ve built over time.

Open the Door for Your Agent

What really matters is that Folo is agent-friendly. It has a CLI tool that turns those 161 feeds into an information library an agent can call directly. After setup, the agent reads the unread items from my Folo subscriptions. It’s no longer grabbing random headlines from the whole web; it’s working with a pre-screened set of sources. And every item comes with the original link, so the AI never hallucinates a URL.

My upgraded morning briefing prompt looks like this: 'Use Folo Read skill.md and follow the instructions. Read the unread items from the past 24 hours. Act as an editor: merge duplicates, cross-verify details, and prioritize official statements. Mark multi-source verified items.'

Step Two: A Good Assistant Is Trained, Not Born

Once the information pool was solid, a new issue appeared: industry trends aren’t the same as my interests. For a while, open-source model news flooded the tech pages. Day one, I’d click. Day two, I’d see parameter breakdowns of the same model. Day three, I knew it didn’t affect my day. What I actually cared about were concrete hardware changes—a new laptop spec leak, a phone launch date.

A human would scroll past, but the agent didn’t. It only knew the topic was hot and sources were writing about it. So I typed directly: 'Too much AI news today. I want more consumer electronics and hardware. Remember this for future digests.'

The agent created a MEMORY.md file and saved that preference as a long-term rule. It actually worked—the next digest was full of hardware news, not AI model updates. That’s what I love about agents: they don’t read your mind, but they remember what you don’t like. A good assistant is often shaped by your complaints.

Step Three: Stitch the Fragments Into a Cyber Newspaper

The agent was summarizing well, but the chat interface was cramped. Since AI can write code, why not have it turn the scattered clues into a personal HTML newspaper?

I asked it to output five deep news items in a minimal, card-based layout: a title, core facts with multi-source verification, why it matters, and link buttons at the bottom. The result was a clean white page where verified facts and commentary sat neatly, and I could jump to the original source with one click.

The same approach works for long-running topics. Take the foldable iPhone—rumors, denials, more rumors. Each one looks huge, but together they’re just the same question rephrased. I challenged the agent to track that thread into a dynamic encyclopedia.

My prompt was detailed: a single-page HTML file with a 100-word status summary at the top, a tree diagram of specs (size, screen, hinge, price, etc.), a timeline of key leaks, a keyword frequency chart based on independent sources, and clue cards with credibility tags like 'confirmed,' 'multi-source,' or 'single rumor.' I also required a source list and filtering by category, credibility, and status. And I banned unverified claims from being stated as fact.

The agent delivered. It broke down two years of rumors into structured data. The top summary said what we know and what’s still up in the air. The tree diagram showed everything from screen ratio to liquid metal hinge to price. The timeline revealed how a rumor evolved and reached consensus. Each clue card was tagged, and the keyword chart counted independent sources, not reposts. Clicking a button took me straight to the original report from Ming-Chi Kuo or Bloomberg.

After seeing that page, I actually felt less anxious about when the foldable iPhone would launch. It’s not that I lost interest—a detailed rumor map can’t replace holding the real device. But when you can tell which clues are supply-chain consensus and which are just clickbait, the fear of missing something big just melts away.

Why This Matters Now

That anxiety disappears because you’ve shut out the low-quality noise. In 2025, Merriam-Webster picked 'slop' as its word of the year. It’s used for AI-generated, low-quality digital content. There’s more info than ever, and it’s packaged to look legit, but judging what’s real is harder than ever.

Building your own reliable source list is one of the best ways to fight back. AI can collect, dedupe, and organize, but it shouldn’t outsource your judgment. The flood of information is still coming. Instead of trying to swim faster, build a small dam upstream: subscribe to sources you trust, keep different voices, and when you see a conclusion, go back to the original.

Whether it’s curating feeds or teaching an agent your preferences, what you’re really doing is installing a filter at the top of your information stream. What reaches you has already been sifted. You can see at a glance what deserves a slow read and what’s just passing noise. Being able to pull something useful out of the noise—that’s a hobby worth having.