An editorial pipeline
A newsroom that runs itself.
Research, writing, fact-checking, charts, licensed photography, translation, scheduling and distribution. One pipeline, under rules you set. Not a prompt in a spreadsheet: a newsroom you configure once and point at a subject.
Counted from our own production, not estimated. Updated hourly, last on 5 November 2026.
What it costs
$0.04 to $1.00 an article
- Typical
- $0.12
- Tokens
- 99k
The cheapest and the dearest of our last 200 articles, priced from the model bills themselves: research, writing, checking, charts and the cover, all of it. The dear end is a long explainer; the cheap end is a short market note.
The pipeline
Seven steps, in this order.
Every one of them is a step you can configure, replace or switch off. The order is the product: a check that runs before the writing is a different thing from one that runs after it.
Research
A worker reads the sources you point it at: feeds, sites, releases, an address you paste in. It proposes what is worth writing about today.
Write
An author with a name, a voice and a brief writes it. Length, structure and tone are settings, not hopes: you say 900 words and get 900 words.
Check
A second model reads the draft against its sources and reports what it cannot stand up. It can correct what it finds, or simply tell you.
Illustrate
Charts drawn from the figures the article cites. Diagrams typeset from the writer's own plan. Photographs fetched under a licence that permits use, credited automatically.
Translate
The same piece, published again in another language, translated by the pipeline with its figures, captions and credits intact.
Hold
A human tick, where you want one. Nothing goes out until somebody has read it, and the article records who did and when.
Publish and push
To a site we build and host, to a site you already run, or into a queue you collect from. Then out to Telegram, Discord and the rest on a cadence per channel.
What gets checked
Nothing is published on a model’s word.
A quotation has to appear on a page we fetched. A chart’s numbers have to appear in the source it cites. A photograph has to carry a licence that permits use, and its credit is built from the provider’s own metadata rather than written by an AI. Anything that cannot be verified is left out rather than approximated.
A fact check that reports before it corrects
It reads the article against its sources and lists what it cannot stand up, with the passage quoted. Correcting is a separate tick, because the first useful thing is knowing.
A human gate you can switch on
Per site: nothing publishes until a person has ticked that they read it. The piece records who, and when, and the site can print it under the article.
Disclosure, by default
Pieces produced with AI assistance say so, in the wording the EU AI Act asks for. It is a setting, not an afterthought.
What the engine draws
It draws the flow, not just the words.
Ask for a sequence in plain words and the writer plans the piece around it: who the parties are, what passes between them, in which order, and where one phase ends and the next begins. A sequence is one of six shapes it draws — flows, stacks, tables, sequences, lifelines and swimlanes — beside bar, line and donut charts built from numbers the article cites.
All of it is typeset into the article rather than pasted in as a picture, as SVG where the drawing needs to be drawn and as HTML where it needs to reflow. Either way it takes the site’s type and colour, stays sharp at any size, and is still readable on a phone.
That figure is lifted from a published piece exactly as the engine drew it, down to the numbering. Open the article and compare it with what the reader sees.
What you control
The interesting settings are the ones about the writing.
Rules that bind every article
House style, what must never be claimed, how sources are cited, how many internal links and where they point. Written in plain English and applied to every piece the site produces, rather than pasted into a prompt and hoped for.
Authors with their own voices
Each author is a persona with a style, a focus and its own rules. A desk can have a sceptic, an explainer and a market writer, and they do not sound alike.
Length, shape and figures
Target length per author or per piece. Ask for four diagrams and a price chart in plain words and the writer plans the article around them.
A schedule, or a trigger
Run at 06:00 on weekdays. Or three times a day between nine and six. Or not on a schedule at all: fire it when you have something to say.
A budget that stops
A daily ceiling in dollars per site, counted against what was actually spent. It is a stop, not a warning.
Your own website, or none
We build and host the newsroom, or we deliver into the site you already run and never touch its design. Both are normal.
What you actually work in
The parts a buyer asks to see.
Every article says how it was made
Not a log an engineer reads, but a record on the piece itself. Which researcher found it, which author wrote it, which model, which checker looked at the cover and what it objected to, which source the quote came from. This is the article behind one of the charts above.
And you can take any of it back by hand
The obvious objection to everything above is what happens when the machine gets it wrong. Every part of a finished piece is yours to change: say what a diagram should show and it is drawn again and swapped in place, have the cover redrawn, rewrite a passage in the author's own voice, move a picture or take it out. Every version is kept, so a redraw you dislike is one click back. The verified blocks survive the edit, because a chart drawn from cited numbers is not something a rewrite should be allowed to retype.
You decide what may stop a publication
Checkers run in the order you set them, with their own round counts. Underneath are the checks that need no model at all: a chart that will not draw, prose too close to its sources. Each one says plainly whether it holds the article or fixes it.
Priced down to the worker
Every call, attributed. Not a monthly invoice to be reconciled afterwards but a running account of which part of the pipeline spent what. That is where the figures on this page come from.
The workers are yours to change
Each one is a prompt, a model and a job. Add a researcher that reads your own feeds, a checker that knows your compliance language, a draughtsman for your house chart style. Nothing here is compiled in.
It runs without being watched
What went out, what is coming, and which task and pipeline each piece belongs to. The point of a newsroom that runs itself is the hours nobody is looking at it.
Every piece carries its own state
Published, read by a person, how many sources it stood on, how many pictures it shows, what it cost. The labels are the questions an editor asks before trusting a piece, answered before they ask.
Worked examples
Three newsrooms you can go and read.
All three run on this platform, with different authors, different rules and different cadences. Open them and check the figures above against what is actually on the page.
Artificial intelligence
SPAISEE
A daily desk on models, funding and releases. The highest-volume site on the platform, and the one that proves the cadence holds.
804 articles published
Crypto analysis
theUnhashed
Longer, argued pieces with charts drawn from cited figures. Where the writing rules and the fact-checking earn their keep.
1,165 articles published
One subject, three languages
NightRiders
The same newsroom publishing in English, Japanese and Czech, translated by the platform rather than re-commissioned.
173 articles published · CS · JA · 92 translated
Where the articles end up
Your site, or no site at all.
A newsroom we build and host
The three above are ours. Yours would be your own design, your own domain, with the reader side already built: search, categories, newsletter, RSS, structured data and consent.
Into the site you already run
You have a site and a design and no intention of changing either. Articles arrive there as drafts for your editor to publish. WordPress delivery is in build. Ask where it is.
A queue you collect from
No site involved: finished pieces wait in a queue and your system takes them when it wants them, with the images, the metadata and the source list attached. The article API is in build. Ask where it is.
And onward, per channel
Telegram and Discord are connected today, with a posting cap per weekday per channel so a quiet community does not get six posts on a Tuesday. 4,263 posts delivered so far across 13 channels.
Contact us
Tell us what you need written.
The useful first conversation is a short one: what the subject is, roughly how much of it per week, where it has to end up, and anything that must be true of every piece. We will tell you plainly whether the pipeline fits, and what it would cost to run.
If you would rather see it than read about it, say so and we will write one piece to your brief.