How Marketing Teams Actually Use Claude Code

The work that actually fits

Marketing teams reach for AI expecting a copywriter and find something more useful: an engineer for the tasks nobody had capacity for.

Reporting and reconciliation. Pulling exports, joining them, normalizing campaign names, and producing a weekly view. This is the strongest starting point because the output is checkable. Numbers either reconcile against the source or they do not, so an error announces itself rather than hiding in prose.

Landing pages and experiments. Building the page, wiring analytics events, adding structured data, checking accessibility, keeping the thing fast. Most of these steps get skipped under deadline, and skipping them is what makes campaign results unreadable later.

Content operations. Not writing the article. The pipeline around it: formatting, internal linking, metadata, image variants, publishing, and checking that what shipped matches what was approved.

Technical SEO and AEO. Crawling your own site for broken links, missing canonicals, thin metadata, orphan pages, invalid structured data. This is tedious, mechanical, high value work, and it is exactly what an agent with a shell is good at.

The guardrail that matters most

A language model asked to support a point with a statistic will supply one. It will be plausible, well formatted, and attributed to a source that sounds real. This is the single largest risk in using agents for marketing, because marketing output is public, attributable, and sometimes regulated.

The control is not an instruction to be careful. It is a rule about where facts come from: every number, customer name, capability claim, and comparison must be retrieved from a document you control, and anything that cannot be sourced does not ship. Write that as a hard constraint the agent operates under, and give it a corpus to retrieve from so the constraint is satisfiable.

That second half is what most teams miss. Forbidding invention without providing a source of truth just produces vaguer copy. The agent needs your product facts, your approved claims, your pricing, and your case study details available for retrieval, or it will fall back on general patterns because that is all it has.

Brand voice is a document, not a vibe

Every team that scales content with agents hits the same wall: output is competent and slightly off, in a way that is easier to feel than to explain. Then someone rewrites everything, which erases the time saved.

Drift happens because the standard exists in a few people's judgment rather than in a document. Fix the input, not the output. Write down the voice as rules with examples: sentence length, what you never say, how you refer to competitors, whether you use hype words, the exact phrasing for your core claims, and two samples of copy that were approved as they were written.

Once that exists, it should be retrievable rather than pasted. RDK indexes files from local vaults, docs, and code as encrypted private chunks, and agents search those chunks before querying a model. The agent pulls your approved claims and voice rules for the current task instead of carrying a summary of them in every request, and token spend drops 80 to 90 percent on this repeated, reference-heavy work because the answer is retrieved instead of regenerated. Stacked retrieval covers most of it: a private vault answers 40 to 65 percent of queries, the public network 15 to 20 percent, and the model handles the remaining 5 to 10 percent.

What stays human

Three things, and they are the three that decide whether the marketing works at all.

Positioning. What the product is for and who it is against. An agent has no view on this and will confidently extend whatever positioning it finds, including a stale one.

The claim itself. What you are willing to say publicly is a business decision with legal weight. Generation is not the constraint here, accountability is.

Judgment about the audience. Whether a message lands with the people you are actually selling to comes from talking to them. No amount of retrieval substitutes for that, and the teams that automate everything else are usually the ones with the most time to do it.

Everything else in the list above is production work. Automate it, verify it cheaply, and spend the recovered time on the three things that cannot be delegated.