A content pipeline where agents write and people sign off

The operator runs many content sites, each with its own domain, language, brand voice and topics to avoid. Every site needs a steady flow of search-led articles, and each article has to hit its target keyword, carry the right links, stay accurate and read like a person wrote it.

Doing this by hand did not scale across sites and languages, and handing it to a language model in one step gave text nobody could vouch for. The operator needed a system where models do the writing, every step has a check that decides whether the work moves on, and a person stays responsible for what is published.

We designed the work as a fixed pipeline of stages. An SEO lead opens a ticket with a target keyword, search intent, length, required links and taboos. The ticket then moves through outline, research, draft, readability, verification, humanizing, information check, normalization and quality control. Each stage refuses to run unless the ticket is in the stage it owns, and a test asserts that the chain of stages has no gaps and no dead ends.

Each stage is an agent with its own prompt, built on an orchestration layer. Prompts ship as version-controlled defaults, and the operator can override them per stage from the admin panel without a release. A site's brand voice and taboos are fed into every stage.

Checks that can be counted are done in code, not by a model: length, keyword frequency, required links and structure. When a model asks to send an article back for a costly rewrite, the pipeline confirms the reason against its own measurement first. If the two disagree, the article goes to a person with both readings instead of looping.

People sign off at the points that matter. Every article stops at ready for review, where an editor approves it or sends it back. Generated legal pages are published only with a review flag that a person has to clear, and the review date appears on the page only after that.

Around the pipeline we built the cockpit the team works in: a web app for tickets, sites, a block-based article editor, bulk approval, site launch, redirects and language versions. Articles are stored as structured blocks and rendered to static pages at publish time, so a change to shared data updates every article on the next rebuild. An SEO dashboard brings Search Console data into the same place.

We run the pipeline with the same discipline as our own delivery process: cheap checks before expensive ones, one definition of done that lives in code, and a record for each stage of how often it fires and whether it pays for itself.

The platform is in production and runs the operator's content sites day to day. Articles reach publication only through staged checks and a human sign-off, and the operator's team can change prompts, sites and checks from the cockpit while the system keeps running. Between late August and mid-October the team shipped a steady run of changes the operator asked for, each as its own reviewed change, without stopping the pipeline.