Turning a Public Data Feed Into Original, Editor-Grade Articles at Scale
A two-stage system that polls a public filings data source through a strict quality gate, then writes original, angle-led articles from the raw records, using story templates and a keyword strategy that keep the output reading like journalism rather than data dumps.
Where they started.
There is a huge amount of publishable news sitting inside public data feeds, if you can find the stories worth telling and write them well. Raw filings are dry, structured, and endless. The client wanted a system that could watch a public filings data source, spot the records that actually make a story (a notable appointment, a resignation pattern, an ownership change), and turn each one into an original, angle led article good enough to publish, without a journalist writing it and without it reading like an obvious machine.
What we did.
The build separates finding from writing. A poller continuously scans the feed and pushes each record through a ten filter quality gate, so only records with genuine story potential ever reach the writing stage. That keeps the pipeline from drowning the site in noise. The writing stage then routes each record by filing type down one of several research paths, and writes to a purpose built story template for that type, so an appointment story and a resignation story are structured and angled differently, the way a real desk would handle them.
The hardest part was voice. Early output was accurate but betrayed itself through forced keyword insertion, the SEO seams showing through the prose. The fix was architectural: instead of writing clean copy and then jamming a keyword in afterwards, the writer now chooses a natural, human sounding keyword up front and writes with it already woven in, so the optimisation is invisible. The prompt was also taught to include the context a human editor would expect, board composition, prior company names, other directorships, so the articles carry real depth rather than restating the filing.
Two workflows do the work. The first is the poller: it scans the data source and applies the ten filter gate to decide what is worth covering. The second is the generator: a routing layer sends each qualifying record across 26 filing types down one of seven research paths, then hands it to the writer. The writer prompt went through many iterations to reach a reliable editorial standard, carrying story templates per filing type, an explicit keyword naturalness rule (if you would not say it out loud, pick a different keyword), a set of anti detection rules that ban hollow AI phrasing, and a kill check that stops a piece that does not clear the bar. A downstream step verifies the keyword is present and adds outbound links to legitimate sources, acting as a safety net rather than the primary writer.
What changed.
The client can turn a stream of dry public records into a steady supply of original, angle led articles that read like they came off an editor's desk, complete with proper context, natural keyword placement, and credible sourcing. The quality gate ensures only real stories get written, and the template routing means each story type is handled in its own correct shape. It is a repeatable way to mine a public data feed for genuine, publishable journalism at a volume no newsroom could match by hand.
Built with n8n, Public filings data source, LLM writing with template routing, Anti-AI-detection editorial rules, SEO scoring integration.
AI Automation
This is AI Automation applied to data: raw filings and records turned into accurate, publishable articles on repeat. We can point the same pipeline at the data that matters to your audience.