The AI news digest that replaced my morning scroll
# The AI news digest that replaced my morning scroll
My phone used to own the first forty minutes of my day. I would pick it up to check the news, fall into a feed, and surface much later with a head full of angry headlines and almost nothing I could use.
These days the news reaches me once. One short message is waiting when I wake: a handful of items from sources I chose myself, each condensed into a few plain sentences, sorted by how much I am likely to care. I read it, open the two or three stories that earn a closer look, and I am done until tomorrow.
The build behind that message took an afternoon. This is how it works, what it costs to keep running, and the places where it still falls short.
The feed was never on my side
Admitting the real problem took me longer than fixing it. There is no shortage of news, and there never was. The delivery mechanism is the trouble. A feed earns from the time you spend inside it, so it learns to surface whatever keeps your thumb moving, and the things a thumb responds to are mostly outrage and anxiety. Nobody sat in a meeting and decided to wreck my mornings. The incentive managed it on its own.
Once that clicked, I stopped wanting a smarter feed. A better algorithm inside an infinite scroll is still an infinite scroll. What I wanted was an edge: a beginning, an end, and nothing past the end. Every design decision that follows came from that one requirement.
Pull once, then stop
The whole system reduces to four steps. It pulls new items from a source list I picked. It hands each item to a language model for a short summary. It sorts the results against a profile of what I care about. Then it delivers everything as a single message and goes quiet until the next morning.
The load-bearing word is pulls. I fetch the news once, on my schedule, and nothing gets pushed at me forty times a day. That single inversion carries more of the value than the AI does. The model saves real reading time, but the shape of the thing, a finite message that arrives once and ends, is what actually returned my mornings to me. A brilliant summary bolted onto an endless feed would have changed nothing.
RSS never actually died
Step one runs on RSS, which deserves a moment, because plenty of people assume it disappeared years ago.
RSS is a plain, machine-readable list of a site's newest items. Most blogs and news sites still publish one, even when no visible link to it exists anywhere on the page. Before the big platforms took over, this was how much of the web got read. You subscribed to sites you liked, and their posts arrived in order, with nobody in the middle deciding what you were allowed to see.
That property is exactly what I wanted back. My digest subscribes to a couple of dozen feeds I trust. When a source starts wasting my time, I cut it. When I find a good new one, I add it. The list changes only when I change it. Automating that choice away would rebuild the exact problem this thing exists to escape, so source selection stays a manual job forever.
The prompt took more tuning than the plumbing
Each new item goes to the model with one instruction: read this article, then explain in three or four plain sentences what happened and why it might matter. Most articles bury one useful fact under ten paragraphs of padding, and a model is genuinely good at digging the fact out and handing it over. Twenty clean summaries now cost me less time than three half-read articles used to.
The tone was the hard part. My first version came back breathless, because breathless is the register the model has absorbed from a million news pages. So the instruction grew specific: write like a calm friend explaining the story, keep the boring important parts, strip the hype, never guess at anything the article does not say, and admit when a story is small. That one paragraph now does more work than any other component. Getting there meant reading output, wincing, and adjusting, across several rounds, the way you would coach a new assistant on how you like things written.
One caveat I keep in mind: the model condenses, it does not judge. A summary of a wrong article is a faithfully wrong summary.
Ranking is rough, and rough is enough
Twenty summaries still outrun my attention first thing in the morning, so the final step sorts them. The model holds a short standing description of the topics I follow closely and the kinds of stories that are useful to me, and it orders each day's items against that. The three or four things I would most want to know land on top. The rest trails off below, there if I want it.
Some mornings it buries something I would have promoted, or floats something I skip. Fine. A rough ranking delivered once, with a hard bottom, beats a precise engine that never ends. I would rather keep a slightly imperfect digest that respects my time than an accurate one tuned to my outrage.
Reaching the bottom is the feature
Delivery sounds like the boring step. It changed my behavior more than any other part.
Everything runs on a schedule before I wake, so the digest is already sitting there when I pick up the phone. There is no refresh to pull and no next page to load. When I reach the bottom, I have reached the actual bottom, and my brain receives a signal that feeds are specifically engineered to withhold: finished.
I expected to feel less informed. The opposite happened. Absorbing the handful of items that matter beats skimming a hundred that do not, and I notice the difference in how much of yesterday's news I can actually recall.
Where it misses
The digest only knows about the sources I gave it. A big story breaking outside that list simply does not exist for me until one of my feeds picks it up.
The summaries flatten nuance. A careful, complicated piece can lose the very quality that made it worth reading when it gets squashed into three sentences, so anything that really matters gets opened in full.
And the ranking reads the same words I would, which means a sufficiently loud headline can fool it the same way it fools me. Early on everything arrived marked urgent, because everything sounds urgent when read cold. Teaching the system what calm looks like for me specifically took examples, and more than one round of them.
So the digest works as a filter for the daily firehose, never as my only window on the world. The thinking stays manual.
The bubble I could build for myself
One risk deserved harder thought than the rest. Hand-picking every source is a recipe for a comfortable bubble in which every voice already agrees with you. A feed, whatever its sins, occasionally throws something uncomfortable at you, even if it does so for the wrong reasons.
My defense is deliberate. A few feeds on my list belong to writers I regularly disagree with, kept there precisely so the digest cannot become a mirror. Small discipline, real stakes. The scary version of hand-picked news is the one where you trap yourself and enjoy every minute of it. This tool obeys me exactly, which means keeping it honest is entirely my job, and I would rather face that on purpose than pretend an algorithm was the only thing narrowing my view.
Costs, and the jobs that stay mine
Summarizing a couple of dozen short articles a day comes to a few cents, so the model spend rounds to a dollar or two a month. The schedule runs on a machine I already leave on, which means no server bill. Setup was one afternoon. Upkeep is the five seconds it takes to add or drop a feed.
The jobs that stay mine are the ones that should. I choose the sources. I decide when the list has gone stale. And I do the thinking about what any of it means, because a summary is an input to a decision, never the decision itself.
Worth an afternoon
If your feed genuinely serves you, keep it, and no judgment from me. If you have ever looked up from your phone wondering where forty minutes went, this build costs an afternoon and about a dollar a month: a list of feeds you trust, a model to do the reading, a schedule that delivers once.
The AI turned out to be the least interesting part. What fixed my mornings was building something that ends, on purpose, in a world full of things designed never to. I write up small, honest builds like this one at [xavierfok.com](/).
Get new guides and videos first — join the Telegram channel.