MioLink.it — AI Press Desk

MioLink.it was built to test how far an editorial workflow can be assisted by AI without losing sight of quality. Instead of handing all the work to a single model with one generic prompt, I built something that looks more like a real newsroom: every stage of the journalistic process has its own dedicated AI model, with its own role, its own "profile" and a specific task — much like a real newsroom, where the reporter doesn't do the fact-checker's job and the editor-in-chief doesn't write the articles.

The newsroom, role by role

The flow is orchestrated with n8n and moves through eight steps, each handled by a model chosen for the task — all open models run locally with Ollama, not paid services:

• News gathering — around twenty-five tech sources are read via RSS by a fast, lightweight model: this needs sorting speed, not reasoning depth.
• Editor-in-chief — a much more technical model evaluates the clusters of gathered news and decides angle, priority and framing, the way an editor would in a newsroom meeting.
• Desk editor — routes the work between the two writing lines depending on the type of story.
• Tech reporter & feature reporter — the actual writing is handled by two different models, both specialized in writing Italian text, but with a different style: one more technical, the other more narrative.
• Senior byline — reviews and polishes the piece, giving it a consistent editorial tone before it moves to review.
• Fact-checker & SEO — another technical LLM, separate from the writers, compares the data in the article against the cited sources, checking that numbers, names and facts match what was actually gathered — not a grammar check, a truth check.
• Headline writer & copy editor — writes the headline and does a final editing pass.
• Art director & publishing — picks the cover image and publishes the finished piece to WordPress.

Splitting the roles this way has a precise reason: a model optimized to write fluent Italian isn't necessarily reliable at verifying facts, which is why fact-checking is handed to a different, more "technical" model whose only job is to compare the article against the sources — not rewrite it, not judge its style, just check its truthfulness against what was gathered.

Before writing: clustering

With around twenty-five sources on the input side, the same story almost always arrives in several copies, told by different outlets. The newsroom's first job isn't writing but clustering: entries about the same event are merged into a single group, so the later steps work on one story and not ten variants of it. It's also where it decides what doesn't deserve an article — most of what comes in stops here.

A newsroom that runs on a PC

There is no paid AI service behind MioLink. Every role is an open model run locally with Ollama — Llama, Mistral, Qwen, Gemma, used across the board depending on the task. The n8n flow runs on my own PC: it starts when I open n8n or launch it by hand, and it's written to be moved as-is onto a server and run on its own. The production cost of one article, across its eight model steps, is therefore zero — apart from the electricity of the machine. It's also a concrete proof: a full automated newsroom doesn't necessarily need expensive APIs, just ordinary hardware and models that run on it.

Why n8n and WordPress

All the orchestration lives in n8n: each node is a step of the newsroom, the arrows are the piece moving from one desk to the next. Having it as a visual flow, rather than a monolithic script, lets me change a single role — swap the fact-checker's model, add an RSS source, move a check — without touching the rest. The destination is WordPress for a practical reason: it is a framework that automates well. It has a full API that n8n can drive — create the draft, assign categories and tags, upload the image, publish — and it brings a ready-made editorial ecosystem (feeds, SEO, archive) I didn't have to reinvent.

What went wrong

Early on the newsroom produced plenty of bad articles: pieces that, along the way, lost the link to the original story and ended up making no sense against the sources they were built from. The fix wasn't a better model but many guardrails and re-checks added along the flow — constraints on what each step can and cannot do, and intermediate checks that stop the piece if it has drifted too far from the gathered material. A second problem was the cover images: the first image-generation models, also local, produced nonsensical covers; they improved by moving to newer models better suited to the task. And there's a simple safeguard on operation: every run sends me a summary email, including on error — if the email doesn't arrive, that itself is the signal something broke.

The numbers so far

MioLink.it is not a theoretical demo: it actually publishes. As I write it has 35 published articles, the first in late November 2025 and the latest in late July 2026, organized across dozens of categories and tags, drawn from around twenty-five tech sources on the input side. The pace is that of a small automated outlet: a few well-worked stories, not a wall of content churned out at speed — which is exactly the trap this project is built to avoid.

What a finished piece looks like

An article that comes out of the flow isn't a three-line summary: it's a structured piece, with a headline, a lead, a body split into paragraphs and the sources it started from. The typical length is that of a news-outlet story — enough to explain the fact and the context, not so much that it dilutes it. For a portal that also hosts other projects, keeping MioLink on a separate domain (miolink.it) is part of the same choice: the automated news lives in its own house, with its own label, without mixing into the hand-written pages.

Everything disclosed

It's the most editorially "sensitive" project on the portal, because it touches directly on the quality and originality of AI-generated content. That's why two things are structural, not optional. First: fact-checking is its own node, with a model dedicated solely to that, before any article goes out. Second: the site says everywhere — in the text and in dedicated banners — that the content is created and processed by an AI system, so no reader can mistake it for human editorial work. Transparency about the use of AI and a truth check are the two conditions that make an experiment like this acceptable.