Curated, Not Random: Building an AI News Feed for Digital Signage Screens
Most signage content problems get solved on a whiteboard in about four minutes. Someone says the screens should show news, someone else finds an RSS widget, and the item is closed. It stays closed until the morning a customer sits down in a waiting room and reads a headline about a fatal crash, set in sixty-point type, on a screen with your logo in the corner.
That is not an edge case. It is the default behaviour of a news feed. A wire service is built for people who chose to read the news and can scroll past anything they do not want. A waiting-room screen serves people who chose nothing and cannot look away. Those two audiences want opposite things, and no amount of platform configuration reconciles them, because the platform was never the problem.
What an RSS widget actually does to your screen
It passes everything through. Clickbait, tragedy, party politics, a headline 140 characters long that wraps into a grey wall nobody reads from across the room. It pulls whatever image the publisher attached, at whatever crop the publisher chose. It has no concept of whether an item suits a dental clinic, a car dealership or a bank branch, because that judgement was never encoded anywhere.
And the reputational cost of a bad item does not land on the publisher. It lands on whoever owns the screen. That asymmetry is the whole argument for treating signage news as an editorial product rather than a data feed.
The decision that shaped everything: AI drafts, a human approves
Full automation was the obvious build and the wrong one. A language model can be instructed to avoid tragedy, and it will still be wrong occasionally. Occasionally is a perfectly acceptable error rate for a recommendation engine and an unacceptable one when the failure is public, silent and sitting in your lobby for eight hours.
So the split is deliberate. The AI does the volume work: reading every source, writing the summaries, proposing a category. A person does the judgement, which is the single question of what actually airs. In practice that is a few minutes a day on a queue of candidates. Nothing reaches a screen without a human having said yes to it.
What happens before anyone looks at the queue
Each morning the selected sources are fetched. A per-customer keyword filter strips out the subject matter that customer never wants on its screens. It is a plain list, editable, and it belongs to the customer rather than to us. What survives goes to Claude Haiku, which writes a two-to-three sentence summary in modern Greek and proposes one of six categories.
The results land in a curation queue, newest first. The curator selects, adjusts the category if the model guessed wrong, and publishes. Publishing regenerates one feed per category, and the screens pick up the change on their next refresh. Every item carries the name and logo of the source it came from. We summarise and attribute; we do not republish articles and we do not present anyone else’s reporting as our own.
Writing for three metres, not thirty centimetres
The least obvious lesson of the build had nothing to do with AI. A summary that reads well on a phone is too long on a screen. Someone glancing up from a waiting-room chair gives the display a few seconds, from several metres away, usually at an angle. That constraint rewrites the brief: short sentences, no subordinate clauses, no numbers that need re-reading, and a headline that wraps into three lines at most at the font size the room actually requires.
It also rules out a lot of otherwise good journalism. An article that needs context to make sense is a bad fit for a screen, no matter how well written. The summaries that work are the ones that are complete in themselves.
Two ways to run it
The managed version runs in our cloud, on the customer’s own subdomain, with the customer’s branding on the widget and unlimited editors. Updates, backups and support are ours to worry about.
The on-premises version is a prepared appliance installed on the customer’s own site. It works fully locally, uses the customer’s own AI account so the running cost is visible and controlled, and never sends content outside the building. That version exists because banks, clinics and public-sector organisations kept asking the same question, and for them the answer has to be architectural rather than contractual.
Where it fits
This is not a signage platform and it does not try to become one. It produces content and hands it over through a single URL, which any reasonable platform can display: Yodeck, Xibo, Scala or whatever is already installed. Nothing about the existing deployment changes; a media item is added, and the screens have something worth showing.
It has been running in daily production since spring 2026 across a national screen network in six categories, which is long enough to have found the failure modes that do not show up in a demo: holidays with no news, sources that change their feed format without warning, and the difference between a summary that is accurate and one that is also readable.
How Amazing Projects can help
If you already run screens and they are showing a loop somebody built months ago, the gap is content, not hardware. The product page, ScreenFeed, Curated News for Digital Signage, has a live look at what the screens actually display, and our digital signage services cover the network itself if you are starting further back than that.
Yodeck, Xibo and Scala are trademarks of their respective owners. ScreenFeed is an independent product and is not affiliated with those companies.
Editorial note — This article was researched and drafted with the assistance of Claude (Anthropic), and reviewed and approved by Amazing Projects before publication.
