Nine Hundred Blogs A Month, Without The Prompt Grind

The real cost was the reprompting

Nine hundred posts a month is a volume problem, but volume was not what made it painful. The pain was that each post took several rounds. Write, read, notice it was generic, reprompt, read again, fix the format, paste it in.

Every one of those rounds was a person waiting on a chat window. Multiply three or four rounds across hundreds of posts and the team was spending most of its month supervising a tool rather than doing anything that needed judgement.

The output was also inconsistent in a way that is hard to fix by prompting harder. Two writers using the same starting prompt got different structures, different depth, and different levels of specificity, because the follow-up prompts were improvised each time.

Research before writing, not during

Generic writing is usually a research problem wearing a style costume. A model asked to write about a service in a town it knows nothing specific about will produce something that could be about anywhere, because that is genuinely all it has.

So the pipeline researches first. One agent studies the client, what they actually do, and which services they sell. Another researches the specific topic the post is about. A third researches the locality, both for regional phrasing and for concrete local detail the post can legitimately reference.

Only then does a writing agent draft, and it drafts against gathered material rather than from a standing start. That single reordering is most of the difference between copy that reads like it came from an industry and copy that reads like it came from a template.

Two quality gates that do different jobs

The first gate checks substance. Is the topic actually covered, is the information accurate against the researched material, and has anything been invented. Fabricated statistics, invented credentials, services the client does not offer, and confidently wrong local detail are all failures. This gate exists because the fastest way to destroy trust in a content system is one hallucinated claim published under a client name.

The second gate checks format independently. Heading structure, section ordering, length, internal conventions, and whatever house style the client expects. It is deliberately separate from the substance check, because a reviewer looking for both at once reliably does neither well.

A draft has to clear both. A pass on substance with a broken structure is still a fail.

Failure triggers a rewrite, not an alert

When a draft fails either gate, nothing lands in a human queue. A prompt-rewriting agent takes the specific failure, composes revised instructions targeting exactly what was wrong, and sends the piece back through.

The rewritten draft then faces both gates again from scratch. It does not inherit a pass from the earlier round, because a fix for a format problem can easily introduce a substance one.

That loop is the reason the human at the end is proofreading rather than repairing. The reprompting that used to consume the team still happens on every post that needs it. It just happens without anybody watching it.

The human stays, with better tools

Every post still reaches a person before publication, and they can act at whatever granularity the problem deserves. Regenerate a single heading. Rewrite one paragraph they do not like. Send the entire post back if it missed.

That granularity matters more than it sounds. Without it, a reviewer who dislikes one paragraph either accepts it or throws away a whole acceptable post, and in practice they accept it. Being able to fix exactly the sentence that is wrong is what keeps quality from drifting downward over hundreds of posts.

Across several hundred posts in testing, the full-rewrite path was never used. Individual paragraph and heading regeneration was, which is the outcome the design was aiming at.