Claude for Marketing: Skills, Projects and MCP

Here is a scene from almost every marketing team. Two colleagues give the same AI assistant the same job, say a newsletter draft for the winter campaign. Two drafts come back with nothing in common: different greeting, different length, one with the discount in the subject line, one without.

The obvious reaction is to write better prompts. So the prompt grows, gets copied into a shared document and picks up examples, rules and a list of things to avoid. Three months later there are four versions of it and nobody is sure which is current.

The problem is not how the question is phrased. It is the missing context behind it. That is what Skills are for: folders of instructions, templates and optional scripts that Claude pulls in by itself when a task matches. This guide shows how retail marketing teams work with them, how Skills differ from Projects and from MCP, and where to start. Everything about the product here was checked against Anthropic's documentation on 1 November 2026.

The prompt is not the problem

A prompt lives for one conversation. It disappears when the chat is closed and survives only as long as somebody keeps copying it forward. Everything your team knows about its own brand therefore sits in individual people and in clipboards.

Context is the opposite: stored once, available on every matching task, versioned, identical for everyone. In practice it splits into three layers that are easy to mix up.

Projects are the shelf with the reference material, Skills are the standard operating procedures, and MCP is the line into the systems holding current numbers. Putting all of it into prompts tries to replace all three at once.

What a Skill actually is

A Skill is a folder with a file called SKILL.md. The top of that file carries two pieces of metadata; below it sits the instruction as plain text.

---
name: newsletter-briefing
description: Produces a newsletter briefing in our house style. Use when a newsletter, a promotional email or a subject line needs writing.
---

# Newsletter briefing

1. Ask for occasion, audience and promotion mechanic if not given.
2. Subject line: 45 characters maximum, no discount figure, no exclamation mark.
3. ...

name and description are required. The name may run to 64 characters and takes only lowercase letters, numbers and hyphens. The description matters more: Claude matches your request against it to decide whether the Skill applies, so it has to say both what the Skill does and when to use it. "Newsletter template" is not enough. "Produces a newsletter briefing in our house style, use for promotional emails and subject lines" works.

The loading model is the clever part. Anthropic calls it progressive disclosure, and it is why a team can keep many Skills around without the assistant getting slower or more expensive.

Level three is the part marketers underrate. Anything that normally lives in attachments can go into that folder: the style guide as Markdown, a list of banned words, twelve months of subject lines as a table, a compliance checklist. Those files cost nothing until they are read, and with scripts only the output reaches the conversation, never the code.

Your first Skill: brand voice

If you only ever build one Skill, build this one. It answers the question every AI draft in marketing fails on: does this sound like us?

A style guide written for people is made of attitudes ("approachable but never matey"). One written for an assistant is made of decisions. Rewrite your rules so they can be checked:

• Address and form. Formal or informal, and how you handle inclusive language.
• Banned words. The vocabulary that never appears in your copy. Ours rules out "seamless" and "bespoke", along with the em dash that models reach for.
• Sentence construction. Maximum sentence length, no nested clauses, active before passive.
• Numbers. Where discount figures may appear and where they may not, for instance never in the subject line.
• Two examples. One paragraph that lands, and one that almost lands, each with a sentence on why.

That last element does the most work. Rules describe the boundary; examples show it. Put both texts in the Skill folder as their own file and point to them from SKILL.md, so they are read only when copy is being written.

The side effect is organisational: the style guide finally has a place where it gets used. A document nobody opens changes no copy. A Skill that fires on every draft does.

Four Skills a retail team genuinely uses

Four shapes have proven durable with online retail teams: small enough to fire reliably, big enough to save real time.

Newsletter briefing. Takes occasion, audience and mechanic and returns subject line, preview text, structure and call to action in your house style. Bundled: banned words and subject patterns.

Weekly competitive read. Summarises which promotions ran in the market and puts them in context. This one needs current data, so it needs an MCP server alongside it.

Campaign post-mortem. A fixed set of questions after every promotion: what was the assumption, what happened, what do we keep, what do we drop. The output is a paragraph, not a document, which is why it gets written at all.

Pre-send check. Runs a finished draft against the things that have bitten you before: mandatory disclosures on discounts, validity dates, minimum order values, exclusions, labelling of AI imagery. No substitute for legal review, but it catches the recurring slips.

One rule holds across all four: one job per Skill. A single large "marketing Skill" gets matched less reliably than four small ones with sharp descriptions.

Where Skills run

Skills are not a standalone product but a feature available in three places. For a marketing team the difference is about distribution.

In the app the feature needs code execution switched on, and Anthropic lists it for plans from Free through Enterprise. Alongside your own Skills there are prebuilt ones from Anthropic and partner Skills for the matching connectors.

One detail catches teams out: Skills do not sync between those places. What sits in the app is not available through the API, and Claude Code works off the filesystem either way. So you maintain them per surface, and if you want one place to do that, use the repository: there they sit in version control, go through your normal review and reach everyone with one command.

Projects are a different thing. A Project is a workspace with its own knowledge base and instructions, available on every plan including free accounts, with the larger knowledge capacity reserved for paid plans. Lasting knowledge belongs there; the procedure belongs in the Skill.

Where the data comes from

Skills and Projects describe how something is done and what holds true. Neither says what happened in the market this week. For that the assistant needs a line into the systems, and that line is MCP: see What is an MCP server? and the overview of MCP servers for marketing.

Together that gives you a division of labour that holds up day to day:

• Project: what always applies here.
• Skill: how this particular job gets done here.
• MCP: what is true right now.

The weekly competitive read shows all three at once. The Skill sets the shape, three paragraphs and a recommendation. The Project supplies the framing, your assortment and price position. And Reyo MCP supplies the observation: the server connects Claude, ChatGPT or Claude Code to your Reyo account, reads your competitors' campaigns and emails, and on request creates entries in the Campaign Planner or writes comments, always only within your own account. The guide covers all three clients; the server is included from the Professional plan.

One limit belongs inside any such Skill, so the assistant does not write past it: with competitive data we see what was sent, not how it performed. Nobody has other retailers' opens, clicks or revenue. And we do not name the shops, because the point is the patterns, not the senders. Where those observations come from is covered by competitor monitoring.

Permissions and trust

A Skill is an instruction, and with bundled scripts it is executable code. Anthropic is explicit: use Skills only from trusted sources, meaning ones you wrote yourself or got from Anthropic. A malicious Skill can direct Claude to use tools in ways its name does not suggest.

For a marketing team that means three things. Read third-party Skills before using them, bundled files included, and look hard at anything that calls out to the internet. Hand out write permissions deliberately: a Skill that only drafts copy is harmless, while one that creates or sends something through a server belongs behind an approval. And give the rules a home, in a short policy covering which data may go into which tool, which connections read and which write, and who signs off on a new one.

Four early mistakes

• The do-everything Skill. One file for all of marketing rarely fires, because its description matches nothing precisely. Split it up.
• The description as a title. "Style guide" does not say when to use the Skill. Put the trigger in writing.
• The style guide as attitude. "We are approachable" cannot be checked. Turn it into rules plus two examples.
• No owner. A Skill that belongs to nobody is stale within a quarter.

How to start

Pick a job that comes up every week and depends on one person. Write down the steps as you would explain them to a new colleague, put the trigger in the description, and try it on two real cases. If the second draft needs no rework, you have a Skill.

After that the order sorts itself out: brand voice first, then whichever procedure eats the most time, then the data. How recurring flows get built in general is covered in the guide to marketing automation.

Takeaway

Skills move the work with AI from phrasing to preparation. What your team knows about its brand no longer lives in four versions of a document, but in one place that fires on every matching task.

The three questions for every new Skill stay the same: is it exactly one job? Does the description say when to use it? And does it hold a decision you would otherwise make from scratch every time?

Last updated 1 November 2026. Details on Skills, Projects and plans from Anthropic's documentation, checked on 1 November 2026.