Claude · Automation

The machine that reads AI news for me and throws away seven out of eight

Every morning at six, a server reads the AI news, rejects almost all of it, and hands me a framed screenshot, a teleprompter script and a caption. Here is how it is built.

QQuentin Megevand
August 3, 2026 · 7 min read

The real cost of watching the news is not the reading

Publishing about AI news means doing three things before seven in the morning: finding what shipped, checking it is true, and deciding whether it is worth talking about. The third one is what kills you.

Because most AI news is unusable. A fifty billion dollar data center, a funding round, a research paper. Impressive, and nothing you can do with it on Monday.

A news machine is judged on what it refuses to send you, not on what it finds.

Mine has been running since 14 July. Twenty one mornings delivered, one hundred and fifty six stories rejected, roughly seven refusals for every keeper. Here is how it is built, stage by stage.

Claude AI Lab

The Claude AI Lab is my Skool community where I share my Claude systems and the more advanced modules. Access is $67/month.

Join the Lab →
What you need
1
A machine that runs without you. A small server for a few euros a month. Your laptop is closed at six in the morning, that is the whole point.
2
A model you can call from the command line. It does the orchestrating: it searches, it reads, it decides, it writes.
3
A search that filters by date. Without a freshness filter you get last week's news presented as new.
4
A delivery channel. Telegram does the job perfectly, and you read it when you wake up.
5
A history file. One line per day sent. That is what stops you covering the same subject twice.
1

The scout

🔎 date filtered search

The machine wakes at six and asks one thing: what shipped in AI in the last twenty four hours.

Then it does the single most expensive thing to automate: it opens the original article.

📰
Aggregators recycle
A site reposts last week's story with today's date. Your search tool surfaces it as fresh.
🔗
Hence the rule
The machine traces back to the primary source, reads the real publication date, and drops the story if it cannot confirm it.

This is the only stage where it is allowed to be slow. Two minutes of checking beat a video announcing as new something that shipped ten days ago.

The rule
No confirmed date, no story. A source quoting another source does not count as a source.
2

The filter

🧪 one test, two questions

This is the core piece, and the one I rewrote the most.

Early on, the machine handed me the biggest story of the day. One morning it sent me a fifty billion dollar data center expansion. Technically the biggest AI story that day. And I had absolutely nothing useful to say about it.

So every candidate goes through two questions.

⚙️The test
6
Can I say what this changes for you? Not comment on it, not summarize it. Say what is different now.
7
Can you act on it this week? Open a tool, change a setting, rethink a way of working.

Two noes and the story dies. And it dies in writing: the machine records what it dropped and why.

🎯
It keeps
Model releases, price changes, new limits, product features, tools you can put to work.
🚫
It kills
Data centers, infrastructure, chips, funding rounds with no product, earnings, geopolitics, research papers.

Across twenty one mornings it has dropped one hundred and fifty six stories. Roughly seven refusals for every keeper. That ratio is not a side effect, it is the product.

Worth remembering
A machine that keeps everything it finds does not save you time, it builds you one more feed.
3

The shot

📸 headless browser

The video shows the article headline in the background. If the headline is not readable, the whole format is pointless.

It took three versions.

⚙️What got fixed, in order
8
Version one: capture the top of the page. The headline ended up at the bottom of the image, hidden behind my head in the edit.
9
Version two: push the headline to the top. The editing app's crop then ate it, leaving a single line visible.
10
Version three: park it at a fixed height, around one fifth down. Readable everywhere, whatever the site.

The machine also handles cookie banners, sticky navigation bars, and images that load late and shift the layout. And when a site resists, it notices, and goes and shoots somewhere else.

The principle
An automated screenshot is never right the first time. It is a setting you correct against real sites, not in theory.
4

The script

🎙️ write, then self review

The script goes straight to the teleprompter, so it has to be written for a mouth, not for a page.

⚙️The constraints it holds itself to
11
A hook in the first four seconds. That is all you get.
12
Between one hundred and one hundred and thirty words. Enough to say something, short enough to be watched to the end.
13
Numbers spelled out. Otherwise I stumble on them reading aloud.
14
No dashes. They break the reading rhythm on a teleprompter.
15
The same closing line every time. Consistency builds recognition.

And one stage I added along the way: after writing its draft, the machine rereads it and asks what sounds artificial in there. Then it rewrites.

A first draft always sounds written. It is the second pass that makes it spoken.

It then delivers three separate messages, and that separation matters: the screenshot with its context, the script alone and copyable, the caption alone and copyable. Nothing to chop up by hand at seven in the morning.

5

The guardrails

🛡️ what it refuses
📅
Never a date it did not read
It opens the original article or it drops the story.
🔁
Never the same subject twice
The history is checked before anything else, not after.
🧲
Never a story with no action
Big news and nothing to do with it means no.
🤐
Never an invented call to action
No made up keyword, no fake urgency.
💻
Never my laptop
It runs on a server. I sleep, the work gets done.

And the guardrail that costs the most: some mornings it finds nothing that passes the test, and it says so instead of filling the slot with something lukewarm.

The real luxury
A machine allowed to send nothing. That is what lets you trust it on all the other days.

What breaks, because things break

Twenty one runs finished cleanly, two failed.

The most recent failure was this morning. The story of the day touched on cybersecurity, and the model refused to process the request. The run stopped dead and the delivery did not go out on schedule.

That is the nature of these systems: they depend on outside services with their own rules, and those rules change without warning. What matters is not avoiding the failure, it is making it visible. A failed run leaves a timestamped line in a log with the exact reason. I know what happened in ten seconds.

The build lesson
Do not try to make your machine bulletproof. Make it legible when it falls over.

Where to start

Do not build all five stages at once.

Run the search and the date check by hand first, once. You will immediately see how many articles carry a date that is not the real one.

Then add the filter, with your own two questions. That is the only part that genuinely belongs to you: it encodes what your audience can actually use. Nobody can write it for you.

Put the whole thing on a clock, on a machine that does not sleep. The rest, the screenshot and the writing, comes later, once you already know the selection holds.

And let it stay silent on the days when there is nothing. That is the hardest feature to allow yourself, and it is the one that makes everything else credible.

Want to go further?

And day-to-day, I post one reel a day on Instagram: @quentin_iamarketing