I Trained an AI Editor With 161 RSS Feeds and Gave Up My Morning Scroll

Every morning, I used to wake up and do the same thing: grab my phone and see what happened overnight. Dozens of unread messages in WeChat, a rotated Weibo trending list, a long string of “important news” push notifications from news apps, and then I'd check a few tech communities for the latest “just released” or “major update.”

After fifteen minutes of scrolling, a strange feeling would settle in. I knew a lot of things, but I couldn't tell you what actually mattered that day. A phone launches, an AI model updates, a company announces a new plan. Each item looks worth tapping, but after reading them all, very little stays with you.

We used to worry about not having enough information. Now the problem is the opposite: there's too much, more than anyone can possibly read. It feels like being swallowed by an ever-expanding information black hole that takes not just time but also your sense of what deserves attention.

So I decided to run an experiment: build my own news curation system using an AI agent, and take back control of what I see every day.

Step One: Ditch the Random Scraper, Plug AI Into a Real Intelligence Network

My first thought 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 was my first briefing assistant.

The output was decent. Before, I had to bounce between news sites, apps, and social feeds to piece together what happened in tech overnight. Now, by the time I made coffee, a briefing was waiting on my phone.

But the novelty wore off in days.

The same model launch could appear three times if three outlets wrote it differently. A story I'd already read yesterday would resurface today with a fresh “latest” tag because another site had rehashed it. I can't blame the agent entirely. A web-wide search returns a bag of loose information: official announcements, media coverage, secondary analyses, reposts, and clickbait, all looking like “new news.” Hand that pile to an AI, and it can summarize fast, but it can't judge what's worth your time.

To cut through the echo, you have to start upstream. Over the years, I'd accumulated 161 RSS subscriptions. But quantity isn't quality. To stay sane in this information black hole, I needed a clear filtering and management system.

Action One: Curate a Tiered Source Pool by Trustworthiness

In an age where information gets repackaged and AI-generated content floods the web, the core principle is source tracing. Every retelling strips away crucial context and can distort the original meaning. The closer you get to the initial node, the truer the picture.

So I sorted my sources by reliability:

• Primary sources: official blogs from OpenAI, Google DeepMind, Anthropic, and other tech companies. These give you unaltered, first-hand announcements.
• Authoritative media: Bloomberg, The Information, Business Insider, WSJ, Reuters, Caixin. They have strong editorial networks and rigorous cross-checking, offering an objective macro view.
• Quality secondary sources and aggregators: The Verge, Techmeme, TechCrunch, MacRumors. They excel at turning raw information into readable, deep content.
• Bloggers and KOLs: tech influencers on Weibo, Bilibili, etc. These serve as supplementary sources for hands-on impressions and unique opinions.

(If you don't mind, you can also subscribe to ifanr and APPSO — they're excellent sources too.)

Action Two: Organize Feeds with a Tree Directory

When your sources hit triple digits, managing them is a chore. Stuff them all in one list and you'll get lost. So I use Folo, an RSS reader I've stuck with. Yes, RSS sounds old-school in the age of algorithmic feeds, but for actively aggregating information and keeping control, nothing beats it.

With Folo, you can organize hundreds of subscriptions into categories, like flipping through a custom magazine. My Folo has six main sections: tech, gaming, culture, AI, autos, and more. Bloomberg, Ars Technica, Wired, TechCrunch, DIGITIMES, MacRumors — each sits in its own list. Opening Folo feels like opening a magazine I built myself.

Action Three: Open the API — Folo CLI as the Agent's “Outer Brain”

More important than all-in-one organization is that Folo is agent-friendly. It offers a CLI tool that turns those 161 feeds from a simple RSS reader into an information library the agent can call directly.

Once configured, the agent can read unread items from my Folo subscriptions. It no longer munches on random titles from the web; it eats a pre-filtered list of sources I trust. Plus, each item has a direct link to the original source, which completely eliminates the problem of LLMs making up links.

I upgraded my morning briefing prompt:

Use Folo Read https://api.folo.is/skill.md and follow the instructions to use Folo. Read unread items from my Folo subscriptions from the past 24 hours. Apply “editor-in-chief judgment”: merge duplicates, cross-verify details from different sources, prioritize official statements, and label multi-source verified items.

Step Two: A Good Assistant Is Trained by Complaints

Once the information pool was solid, a new problem appeared: industry hot topics aren't the same as my interests.

For a while, open-source models dominated tech headlines. Day one, I click. Day two, another parameter breakdown. Day three, I know it's not affecting my life. Meanwhile, I cared about specific hardware changes — a new laptop's specs leak, a phone's release date.

Humans automatically scroll past what doesn't interest them. Agents don't. They just know the topic is hot and the sources are still writing about it. So I told the agent directly: “Too much AI news today. I want more consumer electronics and hardware. Remember this for future briefings.”

The agent created a MEMORY.md file in the background, writing my preference into long-term rules. And it stuck. In its reasoning traces, I could see it comparing and filtering against my preference. The next briefing was all consumer hardware, no AI models.

That's what I like about agents: they don't read your mind, but they remember what you don't like and what you care about. A good assistant is often the product of being “scolded” into shape.

Step Three: Sew the Fragments Together — Print Yourself a “Cyber Newspaper”

Even with the agent helping, the chat interface felt cramped. Text squeezed together, poor reading experience. Since AI can write code, why not have it turn scattered clues into a personal “cyber newspaper”?

I upgraded the prompt again: take the five final deep briefs and format them as an HTML page with a minimalist UI, card layout, a title, core facts (multi-source), why it matters, and link buttons at the bottom for source tracing.

The result was clean white cards with verified facts and commentary, no more walls of text. Click a button to jump to the original story.

The Same Trick for a Long-Running Story

Foldable iPhone rumors are a perfect case. Leaks, denials, more leaks — each one looks huge, but they're often the same question rephrased. I asked the agent to take on a harder job: weave all the threads into a continuously updated “dynamic encyclopedia.”

The prompt was detailed: make a single-page HTML file tracking foldable iPhone rumors. It had to include a 100-word status summary, a tree diagram of specs (screen ratio, liquid metal hinge, price), a timeline of leaks, a keyword frequency chart based on independent sources, clue cards with credibility and status, filters, and a source list. The rules were strict: use official Apple, supply chain announcements, analyst reports, and original media; merge duplicate coverage; distinguish “confirmed,” “multi-source corroborated,” “single rumor,” and “unverifiable”; never state unconfirmed info as fact. And it had to output only the HTML, with embedded CSS and JS, dark theme, responsive, and native SVG or CSS charts.

The agent delivered. At the top, a 100-word summary of the current state, no fluff. Below, a tree diagram laying out every rumor from screen ratio to price. The timeline showed how a rumor evolved and converged. The keyword frequency chart counted independent sources, not reposts. Each clue card was labeled “confirmed,” “multi-source,” or “single rumor.” Clicking a button took you straight to Ming-Chi Kuo's report or Bloomberg's original piece.

Looking at that tracker, I actually felt less anxious about when the foldable iPhone would launch. Not because I lost interest — a rumor tracker can't replace holding the real thing. But when you can see which clues are supply chain consensus and which are just clickbait, the fear of missing out fades.

This calm comes from filtering out low-quality noise. In 2025, Merriam-Webster picked “slop” as word of the year — a term for AI-generated, low-quality digital content. The more information there is, the harder it is to judge.

In this era of digital slop, building your own reliable sources is the best defense against AI garbage. That's the whole point of this system: AI can collect, dedupe, and organize, but it shouldn't outsource our judgment.

The information flood is still coming. Instead of swimming faster, build a small dam. Subscribe to sources you trust, keep diverse voices, and when you see a conclusion, go back and read the original. Whether it's manually curating sources or teaching an agent your preferences, all we're doing is placing a gate upstream.

What flows to you is already filtered. What deserves a deep read and what deserves a quick skim becomes obvious. Being able to extract what's useful from the noise — that's meaningful enough.