Content creation automation

Content creation automation that has to pass an audit

This is the system the articles on this site, and on client sites, are written with. It takes a brief and a keyword and produces a finished long-form article in the client's voice: researched, drafted section by section, proofread five times, fact-checked, and given its internal links and meta tags.

It runs in Claude Code rather than as an n8n workflow. Instead of one chain of prompts, it is a set of skills, rules and verifier scripts that a person runs and reads, which is what lets it hold one standard across hundreds of articles.

Why speed was never the hard part

Producing a draft is cheap now, a shift argued in how to scale content marketing. The expensive part is everything a reader can catch: a price that changed last month, a statistic quoted from a summary of a summary, the same paragraph turning up in four articles in one batch, a link to a page that 404s.

Those are the failures this system was built against, and most were found the hard way. Early batches shared as many as 450 identical eight-word runs between articles. After the checks described below, the best batch since has a worst pair of 4.

How to automate content creation without losing the editor

The pipeline runs in a fixed order, and no stage starts until the one before it has passed:

1. Intake: article type, client, the client's style rules and knowledge base, keywords and any angle that must be hit. Nothing is assumed.
2. Research: 15 results from Ahrefs and 10 from Jina, narrowed to the 10 most relevant and scraped, plus Reddit threads and two authoritative research sources.
3. Outline: one outline per source, merged into a master outline and checked twice.
4. Draft: written one section at a time against a word budget.
5. Proofread: five named passes, covering content, mechanics, a fresh read, consistency across the batch, and facts and rules.
6. SEO: internal links checked against the live sitemap, keyword counts, meta title and description.
7. Final QA: a written report confirming every gate, or the article is called a draft.

The order is the point. Keywords go in after proofreading rather than during drafting, so they land in sentences that already work, and final QA re-checks every earlier gate afterwards.

Why five proofreading passes, not three

The first three passes check that the rules were followed. They cannot tell whether the argument is wrong, and on two batches in a row it was.

Pass four reads a batch as a set. That is where product boilerplate repeated across three articles was caught, along with one article about to ship with no links to its siblings. Pass five audits facts against sources and against the client's own rules, and it caught an article recommending a product for a use the client explicitly rules out, even though every sentence in it was true on its own.

The checks that run as code

Rules a person has to remember get skipped on the fortieth article, so anything that can be counted is counted by one of eleven Python checkers. Five of them do most of the catching:

The rest cover word counts, update sections and whether the rule files still agree with each other. A checker is code, and code has bugs, so every finding gets printed and read before anyone changes an article. One early scan reported double spaces in every file, and they turned out to be paragraph breaks.

What stays with people

The system doesn't choose keywords, set positioning or decide what a client may claim. Those arrive at intake and come from people. It doesn't publish either. Finished articles go into a Google Doc for the client or into the CMS as drafts through automated content publishing, and a person approves them.

The faster sibling

SEO content automation is an n8n workflow that turns a keyword into an edit-ready draft in about 12 minutes, and it suits a team with an editor who wants a strong starting point. This system is slower and heavier. It fits work where every figure and link has to survive a fact audit before an editor sees the article at all.

The course built on the same rules

The rules in this system are what the free Claude Code course teaches, each one through the incident that produced it. It is the quickest way to see the reasoning before deciding whether you want a version of your own.

What an article costs to produce

The research tools bill per call and Claude Code runs on a subscription, so the variable cost per article is small. The real cost is the review time the system is designed to protect, which is why it's priced as capacity rather than per article.

Articles are produced under execution capacity, which is $2,000 a month flat when the AI agents route is added to a fractional Head of Content retainer. A new client's style rules and knowledge base are the bulk of a one-to-two-week setup.

Start with two articles you've published

Send one article you're happy with and one you're not. We'll use the first to draft your style rules and run the second through the passes on a call, so you can see what the checks flag in your own writing.