A language model drafted this article; a person reviewed and published it. That division of labor is not a one-off: it is the editorial system at d0a1.es, and it works this way for every article on the blog. We describe it in detail because a publication that uses AI to write should say so, and because the process has interesting parts and real limitations.
From Prompt to Draft
The journey starts with a prompt. The AI Agent Bridge plugin, version 0.21.1, receives the assignment inside d0a1’s WordPress and queues it as a pending job for Ollama. A local model picks up the job, writes the full article, and then makes a second self-review pass to polish it.
The final filter is automatic: a validator scores the text against a minimum score of 70. With the default configuration, anything below the threshold is rejected; on d0a1.es, with auto-publishing disabled, every draft waits for a human no matter what, and the score is simply a quality signal for the editor.
Human Review Is Not Optional
From the draft onward, the workflow changes hands. A person reviews the facts, fixes the structure, adjusts the tone, and checks the links before hitting publish. The plugin does not publish anything on its own: an editor always presses the publish button.
This division of labor reflects a technical limit. The model writes fast, but it only sees the prompt it receives and cannot verify a fact on its own. That check requires human judgment, which is why the pipeline always ends in a draft, never in publication.
The Editorial Work the Plugin Does Do
When creating the draft, AI Agent Bridge automatically links the article to its counterpart in the other language via Polylang, the system that keeps the Spanish and English versions of the blog connected. It also fills in the SEO metadata, sets the featured image, and inserts internal links.
That work is mechanical and repeatable, which is why it fits inside a plugin. Taking it off the editor’s hands leaves the person with the tasks that require judgment: reading, verifying, and deciding what gets published.
The Limitations, Unvarnished
The system has known flaws. Translation sometimes flattens the heading structure and loses links along the way, so a person reviews every translation before signing off on it.
The AI also does not browse the web or consult its own sources: it writes solely from the prompt it receives. The facts in the draft come from the assignment, and if any of them are wrong, human review is the last barrier before publishing.
Why We Explain the Process
Disclosure is part of the deal with the reader. Knowing that an AI wrote the first draft lets you evaluate the text with full information; hiding it would turn an editorial shortcut into a transparency problem. We explain the plugin version, the validator cutoff, and the translation flaws so the process can be audited.
The AI Agent Bridge plugin is free software and part of the d0a1 packs, so anyone can install it, read its code, and adapt it to their own WordPress.
The Next Step
The WordPress pack lives at d0a1/packs, and the blog has the step-by-step deployment guide. If you set up your own instance, you get the same pipeline: prompt, Ollama queue, self-review, a validator with a minimum score of 70, and a draft waiting for a person. The clearest proof is comparing the initial draft with the published text, because the difference between the two is exactly the human work behind this blog.