Using Claude for the Marketing Work That Repeats

Most marketing teams use language models in exactly the way they are weakest and ignore the way they are strongest.

The weak use is asking for an opinion. Write a post about why our category matters. The output is fluent, immediately familiar, and indistinguishable from what a competitor gets by asking the same question. It has no point of view because it was not given one.

The strong use is transformation. Here is a thing we already believe, expressed once, properly. Now move it.

The work that genuinely repeats

Look at a marketing operation and separate the tasks where judgment is being applied from the tasks where a decision already made is being re-expressed.

• Repurposing. One good piece becoming a newsletter, five short posts, a script outline and an email. The thinking happened once. The rest is format.
• Summarising. Turning a long call, a research document or a competitor's page into something a person can act on.
• Briefing. Converting a rough idea into a structured brief a writer or designer can work from.
• First drafts of the tedious. Meta descriptions, alt text, image captions, product variants. Nobody wants this work and it still has to be right.
• Variants. Ten versions of an ad headline to test, where the point is coverage rather than brilliance.

Every one of these has the same shape: the judgment is in the input, the model is doing mechanical work, and a human is checking output that is easy to check.

What happens when you skip the human

I have direct evidence on this and it is worth being blunt about.

This site previously ran a programmatic content pipeline. Over a hundred posts, generated and published at volume, differentiated mostly by city name. Over three months those posts produced 148 impressions and zero clicks, and Google had indexed thirteen pages out of roughly a hundred and forty.

The model was not the problem. Publishing without judgment was. The output was competent, generic, and answered questions nobody was asking in a voice nobody recognised. That is the predictable result of removing the person from the last step, and it took a full rewrite to undo.

Show it your writing

The most common tone instruction is a list of adjectives. Professional but approachable. Confident, not arrogant. These produce the category average, because that is what those words describe across every business that has ever written a brand guideline.

Give it your actual material instead. Ten things you have written that sound right. A model with real examples matches voice far better than one working from a description, and the difference is immediately obvious in the output.

If you have nothing you are happy with, that is the real problem and no tooling fixes it.

Where the build cost sits

Model usage for a marketing team is trivial. Tens of dollars a month for meaningful volume.

The cost is the pipeline: where source material comes from, how a draft reaches a reviewer, what happens to rejections, where approved output lands, how it gets scheduled and how anyone measures whether it worked. That is integration work and it is priced on the systems involved, not on the model.

Teams that buy the model and skip the pipeline end up with a very capable assistant that everyone stops opening after a fortnight.

The rule I would hold

Use it wherever a decision has already been made and something needs re-expressing. Keep it away from the decision itself. Put a person on the last step before anything is publicly attributed to your business.

That is less exciting than full automation and it is the difference between leverage and volume. If you want the pipeline around it scoped properly, that is what an MVP Roadmap is for.