Claude for Marketing: What It Does Well, and What It Can't Know

This week we asked Claude about our own company. With no web search, it answered: "I don't have specific information about a brand called 'Jinn' in my training data." With web search on, it found us in one answer. In the other, it decided the main Jinn was a London delivery company that happens to share our name.

The same week, six AI engines answered 119 buying questions in our own category, the kind a marketer types when shopping for a brand platform. Jinn came up zero times.

I'm not sore about it. It's the most useful thing I can tell you about using Claude for marketing. The model is very good at the work. It doesn't know your brand, and unless your company is famous, training data won't fix that for you. Everything that follows is about closing that one gap, because the rest of the job Claude can already do.

What Claude is already good at

I build with Claude every day. My loop is idea, research, prototype, comps, build, in hours where it used to take days. Point that loop at a marketing problem and the plain chat window already carries a lot. Claude drafts and rewrites copy for a channel. It reads a pile of customer reviews and tells you which complaint keeps coming back. It tightens a creative brief until a stranger could act on it. It argues with your positioning, which is the use I'd push hardest, because I'd rather hear the objection from a model than from the market.

None of that needs a setup. What it needs is context, and the quality of the output tracks the quality of the context almost exactly. Paste in your voice rules and three approved posts and the draft sounds like you. Paste in nothing and you get the average of every brand in your category.

So the real question behind "Claude for marketing" is where your context lives, and Claude now gives you four places to put it.

Give each brand its own Project

A Project is the simplest upgrade from a blank chat. Anthropic describes Projects as "self-contained workspaces with their own chat histories and knowledge bases" (Claude Help Center). You add documents to the Project's knowledge, write instructions for how Claude should answer inside it, and every chat in that Project starts with that background. Projects are available on every plan, free included, with free accounts capped at five.

For a brand, the knowledge worth adding is short and specific: the voice rules, the words you never use, the claims you can and can't make, a handful of approved pieces, and the customer language you hear most. If you run more than one brand, give each its own Project, so one brand's compliance rules never leak into another brand's launch.

This is where most teams should start, and many never need more for one-off work.

Turn a job you repeat into a Skill

The second place is a Skill. Anthropic introduced Skills as "folders that include instructions, scripts, and resources that Claude can load when needed," usable "across Claude apps, Claude Code, and our API" on paid plans (Anthropic). Each Skill is centered on a SKILL.md file that says when it should be used and what to do.

The test for a Skill is repetition. If you write ad variants under a character limit every week, or run every draft through the same brand red-line review, that procedure belongs in a Skill instead of a prompt you paste from a doc. We publish a set of marketing skills under an MIT license for this reason, a campaign brief and a brand guardrails review among them, because a procedure written down once runs the same way for everyone on the team.

Claude Code for marketing

Claude Code is the one marketers skip, because it runs in a terminal and the name says "code." That's a mistake. Claude Code works on files, and a lot of marketing lives in files: ad exports, product feeds, landing pages, email templates, a spreadsheet of every claim on every package.

That is the shape of Claude Code for marketing: bulk work over real files, with the rules written down. Three pieces make it practical.

A CLAUDE.md file for the brand. Claude Code reads a CLAUDE.md file at the start of every session. Anthropic's docs say each session "begins with a fresh context window," and CLAUDE.md is how your standing instructions carry across (Claude Code docs). Put your voice rules and banned words there, and every session in that folder starts with them.

Scripts that run without you. Pass -p and Claude Code runs a prompt non-interactively, so a weekly job like "read last week's ad export and draft replacements for the bottom ten" can run from a script (Claude Code docs).

Connections to the tools you already use. Claude Code connects to outside tools and data through MCP, which I'll come back to, because it's the piece that changes the most.

Where all of this breaks: five copies of your brand

Follow a team that adopts everything above for six months. The Project has a brand guide someone uploaded in the spring. The CLAUDE.md in the landing-page folder has its own list of banned words. The ad-variant Skill has the product claims typed into it. A teammate keeps a personal set of instructions with an older tagline. The email platform's built-in assistant has a fifth version.

That's five copies of your brand, and copies drift. The day your legal team strikes a claim, four of the five copies still make it. Nobody did anything wrong. The facts were pasted, and pasted facts don't update.

Anthropic is candid about the limit here. Its own docs say Claude treats CLAUDE.md files "as context, not enforced configuration" (Claude Code docs). That's the right design for a general tool. It also means a stale fact in a context file goes out with the same confidence as a current one.

When I talk to brand founders, what I hear comes down to one sentence: before AI it was slow and took forever to get creative, and now with AI it doesn't work, because it takes so much time to learn the tools and they're wrong so often. Most of the time, wrong traces back to the facts the model was given, or wasn't.

What changes when the agent can reach the brand itself

MCP, the Model Context Protocol, is the fix for the copies problem. Anthropic launched it in November 2024 as "a new standard for connecting AI assistants to the systems where data lives" (Anthropic). The point that matters for marketers is simple. With MCP, the agent asks for your facts at the moment it works, from one source, instead of carrying a copy you pasted months ago.

Both Claude surfaces support it today. In Claude Code, adding a remote server is one command, claude mcp add --transport http, and a server can be saved at project scope in a .mcp.json file so the whole team shares it through version control (Claude Code docs). In the Claude apps, custom connectors use remote MCP on every plan, with free accounts limited to one, and you can attach a bearer token as a request header (Claude Help Center). Anthropic's same page warns that custom connectors reach services it hasn't verified, so connect only servers you trust.

This is what we built Jinn Agents to be, and we're its first customer. One bearer token connects Claude, Cursor, or any MCP client to your brand's verified record. The tools return your voice rules, including the words you never use, your positioning, and your design system, from the same record every Jinn product runs on. In Claude Code the connection is one line:

claude mcp add --transport http jinn https://app.jinn.works/api/mcp --header "Authorization: Bearer $JINN_MCP_TOKEN"

Two details matter more to me than the tool list. The first is that one tool, ask_brand, takes a plain question about your brand and answers with the facts that bear on it, and names what your record can't answer, so the agent says "we don't have that" instead of inventing it. Inventing is the failure from the top of this piece, and naming the gap is the cheapest way I know to stop it.

The second is the line on spending. Only two tools write anything, and neither can spend money. An agent can queue a purchase. A person completes it. AI generates, humans decide, and here the boundary is built into the product, so nobody has to remember it.

You can try it without an account. A free demo token needs no login, reads a few showcase brands, and expires after sixty minutes, which is long enough to connect Claude to a real brand record and watch what it reads. If you want the longer version of what AI marketing agents can do once they read your brand, we wrote it up here.

What I'd do this week

And review the work. The job isn't done when the AI is done. It's done when you've reviewed it and it is what you expected.

Jinn Agents is in early access, with open sign-up.