I Trained an AI Editor on 161 News Feeds—It Beat Me at Finding Real News

Morning Scroll, Empty Mind

Every morning, I wake up and do the same thing: grab my phone and see what happened overnight. Dozens of unread messages in WeChat. Weibo trending topics already cycled through several rounds. News apps push a long list of “important updates.” Then I open a few tech communities and see more “just released” and “major update” posts.

After fifteen minutes of scrolling, a strange feeling creeps in: I know a lot of things, but I can’t say what actually matters today. A phone launched. An AI model updated. A company announced a new plan. Each item looks worth tapping, but after reading them all, little sticks. We used to worry about not enough information. Now the problem is too much—so much that it swallows your time and your ability to judge what deserves attention.

So I ran an experiment: build an AI agent that filters news for me, a system that puts me back in control of “what to read today.”

Step 1: Ditch the Firehose, Plug Into a Curated Network

My first idea was simple: set a scheduled task in the agent to run at 8 a.m., gather five tech news items, and send them to me. That gave birth to my first briefing assistant. It worked, sort of. Before, I had to hop between news sites, apps, and social feeds to piece together what happened in tech overnight. Now, while I made coffee, a briefing was already waiting on my phone.

But the novelty wore off within days. The same model launch, written in three different ways, would occupy three of the five slots. A story I read yesterday would reappear as “latest” because another outlet had rewritten it. The agent wasn’t slacking—it was fed a messy bag of “loose information.” Official announcements, media reports, secondary analysis, reposts, clickbait—they all look like “new news.” The AI can summarize fast, but it can’t easily tell you which ones are worth your time.

To cut the echo, I had to start with the sources. Over the years, I’d collected 161 RSS feeds. But quantity isn’t quality. To stay sane in this information black hole, I needed a clear filtering and management system.

Action 1: Build a Tiered Source Pool by Credibility

In an age of recycled content and AI-generated fluff, the core principle for getting useful information is “trace back to the source.” Every retelling and aggregation loses context and can even distort the original meaning. Only by getting as close to the original node as possible can you see the true, accurate picture.

So I sorted my sources by trust level:

• Primary sources (official announcements and earnings): OpenAI News, Google DeepMind News, Anthropic News—these ensure first access to unaltered original info.
• Credible media (serious reporting): Bloomberg, The Information, Business Insider, WSJ, Reuters, Caixin—they have strong editorial networks and cross-verification.
• Quality secondary sources and aggregators: The Verge, Techmeme, TechCrunch, MacRumors—they turn raw info into readable deep dives.
• Bloggers and KOLs: Digital big names on Weibo, Bilibili tech YouTubers—useful for hands-on experience and unique opinions.

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

Action 2: Organize Feeds with a Tree Structure

When you have over a hundred sources, managing them is a chore. If you stuff them all into one list, you’ll get dizzy. So I use a tool called Folo. It’s essentially an RSS reader. RSS pushes the latest content from websites to you. In an era of algorithmic feeds, it feels old-school, but for actively aggregating info and staying in control, nothing beats it.

With Folo, I can flip through my subscriptions like a custom magazine. I’ve organized mine into six main sections—from Bloomberg, Ars Technica, Wired, TechCrunch to DIGITIMES, MacRumors—each sorted into categories like tech, gaming, culture, AI, and autos. Opening Folo feels like reading a tech magazine I’ve built myself over time.

Action 3: Give Your Agent an “Outer Brain” with Folo CLI

What matters more than organization is that Folo is agent-friendly. It offers a command-line tool that turns those 161 feeds into an information library my agent can call directly. After setup, the agent reads unread items from my Folo subscriptions. It’s no longer scraping random headlines from the web—it’s working with a batch of sources I’ve already vetted. Plus, each item includes a direct link, so the AI can’t hallucinate URLs.

My upgraded morning briefing prompt: “Use Folo Read https://api.folo.is/skill.md and follow the instructions. Read unread items from the past 24 hours. Then act as an editor-in-chief: merge duplicates, cross-verify details from different sources, and prioritize primary sources. Mark items as ‘multi-source verified’ when applicable.”

Step 2: A Good Assistant Is Shaped by Criticism

Once the info pool was solid, a new problem emerged: industry hot topics aren’t necessarily your topics. For a while, open-source models dominated tech headlines. Day one, I clicked. Day two, another breakdown of the same model’s parameters. Day three, I realized it had zero impact on my day. Meanwhile, specific hardware news—like a new laptop’s specs or a phone launch date—was what I actually cared about.

Humans skim past what doesn’t interest them. Agents don’t. They only know the topic is hot and sources keep writing about it. So I told my agent directly: “Too much AI news today. I want more consumer electronics and hardware coverage. Remember this for future briefings.” The agent generated a MEMORY.md file and stored that preference as a long-term rule. And it worked—the next briefing was full of hardware news, not AI model announcements.

That’s what I love about agents: they don’t magically understand you, but they remember what you dislike and what you care about. A good assistant is often shaped by your feedback.

Step 3: Stitch Fragments Into a “Cyber Newspaper”

Even with a smart assistant, the chat interface felt cramped. Text squeezed together, not great to read. Since the AI can write code, why not have it turn scattered clues into a personal digital newspaper?

I upgraded the prompt: “Take the final five deep briefs and make an HTML page with a minimalist UI, card-based layout. Include a headline, core facts (multi-source), why it matters, and source buttons at the bottom.” The result was a clean page with white cards, each showing verified facts and commentary. No more dense text blocks. Click a button to jump to the original article.

The same approach works for tracking a long-running story. Take the foldable iPhone: leaks, denials, more leaks—each one looks like a big deal, but together they’re just the same question rephrased. I challenged the agent to build a “dynamic encyclopedia” of foldable iPhone rumors. The prompt asked for a self-contained HTML file with:

• A 100-word status summary at the top
• A tree diagram showing categories like release date, form factor, screen, hinge, specs, price, and unconfirmed items
• A timeline of key leaks with first appearance, latest change, and current status
• A keyword frequency chart counting independent sources, not reposts
• Clue cards with specific claims, confidence level, status, earliest date, and original source link
• Filters by category, confidence, and status
• A full source list

The rules were strict: prioritize official Apple statements, supply chain announcements, analyst reports, and original media; merge duplicate reposts; clearly label items as confirmed, corroborated, single rumor, or unverifiable; never present unconfirmed info as fact; every confidence rating must trace back to a source. The output was a dark-themed, self-contained HTML file with embedded CSS and JavaScript—no external frameworks or images.

The result was impressive. The agent dissected two years of scattered rumors into structured data. A 100-word status summary at the top cut to the chase: production timeline and core uncertainties. Below, a parameter tree showed everything from screen ratio to liquid metal hinge to final price. More valuable was the timeline and keyword frequency chart, which revealed how a rumor evolved and reached consensus. Each clue card was tagged “confirmed,” “corroborated,” or “single rumor.” For hot topics like “no crease,” you could see exactly how many independent sources backed it, with buttons linking to original reports from Ming-Chi Kuo or Bloomberg.

This multi-dimensional cross-referencing saved me from piecing together truth across dozens of web pages. It handed me a logical “investigation report.” After reading it, I felt less anxious about when the foldable iPhone would actually launch. Not because I lost interest—rumors, no matter how detailed, can’t replace hands-on experience. But when you can see which clues are supply-chain consensus and which are just clickbait, the fear of missing out fades.

Building a Dam Against the Slop

That anxiety disappears when you filter out low-quality information. In 2025, Merriam-Webster chose “slop” as its word of the year. Originally meaning “swill,” it now describes AI-generated, low-quality digital content. Information is abundant, packaging looks better, but judgment gets harder.

In this flood of digital slop, building your own reliable source pool is the best defense against AI garbage. That’s the whole point of this system: AI can collect, dedupe, and organize, but you can’t outsource your judgment.

The information will keep pouring in like a flood. Instead of trying to swim faster, build a small dam upstream: subscribe to sources you trust, keep diverse voices, and go back to the original when you see a conclusion. Whether it’s manually filtering feeds or teaching an agent your preferences, it’s all about installing a gate upstream. What flows to you is already settled. What deserves a slow read and what deserves a quick skim becomes clear.

Being able to extract what’s truly useful from the noise—that, in itself, is meaningful enough.