A Note On Method
How This Paper Is Made
Five front pages, several million pixels, and three colours to show for it.
Reading The Colours
Every morning the front pages of five British national newspapers are collected as images. The top eighth of each page is cut away, because a masthead is a fact about the paper rather than about the day, and without removing it every Tuesday would look like every other Tuesday. What is left is reduced in size and read pixel by pixel.
Anything that is not really a colour is discarded first: the newsprint itself, the black of the text, the white of the margins. Formally, a pixel is kept only if its saturation reaches a quarter and it is neither almost black nor almost white.
What survives is grouped by k-means into fifteen clusters, and the six largest become that paper's colours, each weighted by how much of the page it covered. The five papers' colours are then pooled and compared in the LAB colour space, which measures difference as the eye perceives it rather than as the arithmetic suggests. Colours close enough to be called the same are merged and their weights added together.
Each merged group is then represented by its heaviest member, not by an average of the group. This matters more than it sounds: averaging several unlike colours together reliably produces brown, and an earlier version of this project spent its whole life reporting mud for exactly that reason. A colour printed here is a colour that was actually on a page.
The three heaviest survivors are the colours of the day. They are named by a rule, not by a person: hue decides the family, brightness and saturation supply the adjectives. That is why the names read a little mechanically — Deep Rich Blue, Pale Muted Orange. A machine is doing the looking.
Reading The Headlines
The headlines are read off the same scans by optical character recognition, running locally rather than through any outside service. Recognising the text turns out to be the easy half of the problem: a front page is covered in text, and almost none of it is the headline.
So the splash is found by geometry rather than by meaning. Every line of recognised text is kept with its position and its size, lines of similar size sitting next to one another are grouped into blocks, and the block carrying the most ink wins, weighted towards large type. Nothing is read for sense; the machine has no idea what the words mean, only that these ones are set six times larger than those ones.
The masthead has to be excluded by hand, so to speak, and it resists in three different ways: it sits at the top of the page, it often matches the paper's own name, and The Independent sets its nameplate sideways down the left-hand edge, which to a recogniser looks like a single line of type as tall as the page. All three are tested for.
Headlines are reproduced here as short quotations, each credited to the paper that printed it. The scans themselves are read, measured and discarded; this site never republishes a front page.
A Short History
Three Colour News began in September 2013 as a Twitter bot. It did one thing: each day it read the British front pages, reduced them to three colours, and posted them. It stopped publishing in 2015 and stayed quiet for a decade.
It was very nearly called something else. The account was registered on the fourteenth of September 2013 as Four Colour News, and renamed two days later, six days before it published anything. Four would have been the honest number for process colour, which is printed in four inks; three is the number that reads as a front page rather than a specification.
This is the revival, rebuilt from nothing in 2026 as a newspaper rather than a feed. The method is recognisably the same idea and almost entirely different machinery. The masthead numbers these editions as a new series, because the original run posted its colours and kept no archive of them. The record here begins on Sunday, 22 September 2013.
Following It
There is a feed, at threecolournews.com/feed.xml, carrying the last sixty days. Each entry is a date, the three colours named, and the card showing them, so a reader that displays images will show you the day without your having to open it.
The day's colours also go out on Bluesky, as @threecolournews.com. The handle is this domain, which is the one piece of a social account you can actually own.
The original account, @threecolournews on what was Twitter, is still there, holding the posts from the first run.
The whole archive is also published as JSON — every day, every month, and the colours that recur — with no key and nothing to sign up to. An assistant pointed at threecolournews.com/llms.txt will find a briefing there explaining what the data is, where it is, and what it cannot tell you.
There is no newsletter and nothing to sign up to. The paper this one replaces was a Twitter bot, and a feed is the same idea without depending on anybody's platform.
Where It Is Weak
Three numbers cannot describe a newspaper, and the method has known faults. A day on which only one or two papers could be read is marked as such, because its colours are not comparable with the rest.
Photographs of people make up much of a front page, and their colours count here like any other. That puts a quiet obligation on the filtering: a brightness threshold cuts deep skin tones and a saturation threshold cuts pale ones, so a careless pair would let the day's colours depend on who happened to be photographed. The thresholds used here were chosen by measuring that directly across the archive, and they pass deep and light tones within about one percentage point of each other.
The headline reader is right most of the time and ordinarily wrong in small ways: a dropped word, two stories spliced together, or two words run into one. Where it reads a page as having no headline at all, nothing is printed rather than a guess.
Five papers is not the whole press. The Guardian, the Daily Mirror and the Financial Times are missing because the source these scans come from does not carry them, and four days in July 2026 are missing because that source has a hole in it.