You opened Claude, asked for a LinkedIn post or an offer page, and got something correct, polite and interchangeable with any other company's copy. It is the complaint I hear most often about using Claude AI for marketing, and it was mine at the start too. The model is hardly to blame: it did not know who it was writing for, what is true about your business, or what good copy from your brand looks like.
I use Claude every day for my content machines. This blog is written with one of them, and the one at Moonworkers, my own company, publishes two articles a day. Here is what I give Claude before asking for any text, how that groundwork becomes reusable, how I stop it inventing things, and what I keep for myself.
01What Claude does well in marketing, and what it does badly
Claude is good at holding a voice once it has been described, following a long structure without dropping a piece, reworking content for another format (an article into a post, a post into an email) and checking a text against a list of rules. It is also quick at producing three versions of a headline or a subject line, which helps you decide.
It is bad at what it cannot know: your figures, your customers, the real reason people choose you, the nuance that makes a sentence ring true in your market. Without those, it fills the gap with whatever is most likely. The result reads like the average of what already exists, and your audience has learnt to scroll straight past that kind of text.
It can also state something false with the same confidence as something true. Good intentions are not enough here, you need a method.
02Context before the prompt
Anthropic's documentation uses an image I find accurate: treat Claude like a brilliant new employee who knows nothing about your norms or your ways of working. Its guidance on writing prompts suggest a simple test: show the instruction to a colleague who lacks the context. If they would be lost, Claude will be too.
What goes into my context pack
For any marketing job, I prepare the same pieces before writing a single instruction:
- The reader, in one sentence: who they are, in what situation, with what problem.
- What the text must achieve: a sign-up, a conversation, an understanding.
- The voice: a few adjectives, the words you use, the words you never use.
- The approved facts: what is true about the offer, figures included, and nothing else.
- Examples: three to five pieces of brand copy you think work, and one or two that ring false.
The same Anthropic page recommends three to five examples, and explains that giving the reason behind an instruction helps the model aim well. "Don't write a list" works less well than "this text will be read aloud in a meeting, a list would not work". It also advises wrapping each type of content in its own tag (instructions, context, examples), so that the model does not mistake an example for an instruction.
Good and bad examples
In my machines, the writing instructions include sentences that sound right and others that sound wrong. For this blog, the example not to follow is something like "I support companies with a unique, tailored approach". Showing the flaw corrects it faster than banning it in abstract terms. I go into more detail in my article on the AI content machine and human review.
03Making the groundwork reusable: projects and skills
Pasting that pack into every conversation soon becomes tedious, and eventually you forget. Two Claude features solve this, and they do different jobs.
A project holds context. Anthropic presents it, in its announcement of Projects, as a space where you upload your documents (style guides, interview transcripts, past work) along with instructions specific to that project. For a marketing team, it is the natural home for the brand platform, product sheets and reference copy.
A skill holds the way of working. According to the introduction to Agent Skills, it is a folder of instructions, scripts and resources that Claude only loads when it is relevant to the task. In a HubSpot Marketing tutorial whose transcript I read, the distinction is neatly put: the project tells Claude who you are, the skill tells it how to do the work. The same tutorial stresses the skill's description, which decides whether it fires at all: too vague and it never gets used.
That is how I work. This blog has its own skill: it says which references to read before writing (the voice, the list of usable facts, the structure, the rules), in what order to proceed and which commands to run. The hard rules are not in the skill. They live in a configuration file that the tool reads back: a rule repeated in a prompt gets forgotten, a rule in the configuration blocks.
The HubSpot tutorial lists mistakes I have made almost all of: one skill that does everything, a skill never updated, a third-party skill installed without being read. It adds one I keep in mind: a skill speeds up your standards, it does not replace them. If your criteria are vague, you will get vague copy, just faster.
04Stopping Claude from making things up
This is the risk that worries me most in marketing, because a false claim about your offer or an invented market figure ends up in a public document. Anthropic devotes a page of its documentation to reducing hallucinations, and one of its examples is precisely a press release.
The techniques it describes are easy to apply:
- explicitly allow Claude to say it does not know;
- ask it to use only the documents provided, not its general knowledge;
- after drafting, have it find an exact quote in those documents for each claim, and remove any claim without one.
The page notes that these techniques reduce errors without removing them, and that anything important still has to be checked. I take two rules from this in my machines. Each source enters the brief with the precise claim it supports, and only after it has been read. The personal angle can only draw on a closed list of facts already published: if a topic calls for an experience that is not on the list, the article waits for my decision instead of being written.
On the search side, Google is clear about the substance. In its guidance on AI-generated content, it explains that it judges the quality of content rather than how it was produced, and treats automation used mainly to manipulate rankings as spam. The problem is low-value text produced in bulk, whatever tool wrote it.
05What I don't hand over to it
I don't hand Claude the choice of what to say. The angle of an article, the promise of a page, the decision to cover one subject rather than another: these are marketing decisions, and they require knowing the customer and the market. Claude can suggest options, it does not decide.
I don't hand it the figures either. Every published figure comes from a source I have opened, with its link in the same paragraph. My automated check verifies that rule before anything goes live.
Finally, nothing is published without a check. On this blog, a gate reviews each article against its brief before it goes online: structure, sources, links, writing tics, confidentiality. The same principle applies to reporting: I describe elsewhere how I automated a monthly SEO report, and the text is only useful because someone who makes decisions reads it, as in a monthly marketing committee.
This groundwork has a cost. Preparing the context pack, writing a skill and a check takes time, more than typing an instruction. For a one-off piece, I would not bother. It pays off on the formats that come back every week.
06Where to start this week
If you already use Claude for your marketing, I would start small:
- Pick the task you explain most often, such as your LinkedIn posts or your newsletter.
- Write the context pack: reader, goal, voice, approved facts, three to five examples.
- Put it in a project, produce three pieces with it and note every correction you make by hand.
- Turn the corrections that keep coming back into written rules, and only then into a skill.
- Keep a human review before anything is published, checking facts and figures first.
After a few weeks, you will know what can be automated and what should stay with you. If you would like to set up this kind of groundwork for your team, or a full content machine, it is part of my work, described on the services page. Happy to talk it through.