Turning content production from a six-hour job into a 30-minute review

The problem

Writing a good article properly takes most of a working day. Research. A structure. A full draft. Editing. Fact-checking. Formatting. Done right, each piece takes between 5 and 7 hours.

When the same person is also handling client work, business development, and everything else, that time doesn't exist. The choices are to write less often, lower the bar, or hand it off to a writer who doesn't know the business the way you do.

None of those options work when you're trying to build authority in a specific space and compete on search against teams with dedicated writers.

What we built

A production flow where the operator moves a card in their project board and the article shows up ready to read. Research, drafting, quality checks — all handled before the operator sees anything.

The system researches the topic, writes a full draft in the brand's voice, then runs it through three independent quality checks — one for accuracy, one for search structure, one for whether it sounds like a real person wrote it. If a check flags a problem, the draft is revised before it reaches the operator.

Once approved, the article publishes automatically. The topics database updates so the same angle doesn't come up again. Each article informs the next — structure, sources, what worked.

What we had to figure out

• Early drafts were accurate but flat The first version produced well-structured, factually correct articles that read like they'd been assembled rather than written. We added a dedicated editing pass whose only job is making the writing sound like a person — shorter sentences, a clear point of view, no filler. It runs after the draft and before the quality checks.
• One quality checker was flagging almost everything One of the review passes was failing nearly every article and sending it back for another round. We capped the revision cycles at two. If a draft still has issues after two passes, it reaches the operator with a note explaining what to look at. Infinite loops don't ship anything.
• The system kept proposing similar topics Without a log of what had already been written, the system kept proposing similar angles. We built a coverage log that updates after each publish. Any topic already covered — or in progress — gets flagged before it becomes a draft.
• A provider update changed how the system read source material One of the external tools we use updated how it handles certain document formats. It caused misreads for a short period. We caught it during a review, locked the version, and set a quarterly reminder to check for this kind of drift.