Claude Code is more than a chat window that writes code. It has an extension layer made of CLAUDE.md, skills, MCP, subagents, hooks, and plugins, and the names are easy to confuse. Marketers who learn which tool does which job get far more out of it, because each one solves a different problem.
Everything below follows Anthropic’s own Claude Code documentation, which changes as the product updates. Check the docs for current details before you build anything important. The marketing uses are our suggestions, not features of the tools.
The one-sentence version of each
- CLAUDE.md: standing instructions Claude reads at the start of every session.
- Skills: reusable instructions or workflows Claude loads when needed, or that you trigger with a
/command. - MCP: a connection from Claude to an outside service, such as a database or another app.
- Subagents: helpers that work in their own separate context and return only a summary.
- Hooks: automatic actions that run every time a specific event happens.
- Plugins: a package that bundles skills, hooks, subagents, and MCP servers so you can reuse a setup.
CLAUDE.md: rules Claude should always know
According to Anthropic, CLAUDE.md is loaded into every session automatically and holds persistent context such as project conventions and “always do X” rules. For a marketing or website project, that might be your brand voice rules, naming conventions, and things never to do.
Keep it short and specific. Anthropic’s guidance is to target under 200 lines per file, because longer files use more context and reduce how well Claude follows them. We cover how to write one in our guide to writing a CLAUDE.md for website and marketing projects.
Skills: repeatable workflows you can name
A skill is a markdown file called SKILL.md with two parts: a short header that tells Claude when to use it, and the instructions to follow. You can invoke one with a command like /skill-name, or Claude can load it when it is relevant.
Use a skill when you keep pasting the same playbook into chat. Marketing ideas that fit this pattern:
- A monthly reporting checklist that always follows the same steps.
- A meta-title and description review that applies your length rules. Our SERP preview tool is a good way to check the results.
- A UTM-tagging routine that follows your naming convention. Pair it with our free UTM builder and UTM parameters guide.
A skill is read as instructions, not enforced code, so check the output.
MCP: connect Claude to your tools
MCP (Model Context Protocol) connects Claude to external services and tools. Anthropic’s examples include querying a database, posting to Slack, or controlling a browser. For marketers, the value is that Claude can read live data from a connected tool instead of working from a file you exported by hand.
MCP is the right choice when you keep copying data from a tool Claude cannot see. Be careful about permissions: connect only what you need, prefer read-only access for reporting, and keep humans in the loop for anything that spends money or sends messages. Our guide to MCP servers for marketing and our post on MCP server workflows go deeper.
Subagents: keep the main conversation clean
A subagent runs its own loop in an isolated context and returns a summary. That makes it useful for tasks that read a lot of material, such as scanning dozens of pages or exports, when you only want the conclusions back in your main conversation.
Anthropic describes the key benefit as context isolation: the work happens separately, and only the summary comes back. If a side task keeps flooding your conversation with output you will not reference again, route it through a subagent. See our post on Claude Code agents for marketing analytics for more.
Hooks: automation that always runs
A hook runs your script, request, or prompt when Claude Code reaches a specific event, such as after a file is edited. Anthropic’s own example is running a linter after every edit.
Use a hook when something must happen every time and you do not want to rely on Claude remembering. Anthropic notes that CLAUDE.md instructions and output styles are followed as guidance, not enforced, so anything that has to happen every time belongs in a hook. For a website project, that might be formatting or a check that runs after each edit.
How they fit together
| Need | Use |
|---|---|
| Claude keeps forgetting a rule | CLAUDE.md |
| You paste the same procedure repeatedly | A skill |
| You copy data from a tool Claude cannot see | MCP |
| A side task floods the conversation | A subagent |
| Something must happen every time | A hook |
| A second project needs the same setup | A plugin |
Anthropic’s docs describe the same triggers: add a rule to CLAUDE.md when Claude gets something wrong twice, save a prompt as a skill when you keep typing it, connect MCP when you keep copying data, and write a hook when you want something to happen every time.
The order we suggest for marketers
- Start with CLAUDE.md. It is the cheapest and most useful step.
- Add two or three skills for tasks you repeat weekly.
- Connect one MCP server for your most-used data source, read-only first.
- Use subagents when research tasks get heavy.
- Add hooks for checks that must never be skipped.
- Package it as a plugin only when you need the same setup in other projects.
You do not need all of it on day one. If you are brand new, read what Claude Code is and our setup guide first.
Safety habits that apply to everything
- Review before you publish or send. AI output is a draft.
- Never paste secrets such as passwords or API keys into prompts or instruction files.
- Start with read-only access and add write access only when needed.
- Keep a human approval step for spend, sends, and publishing.
- Write down what each skill and hook does, so teammates are not surprised.
Frequently asked questions
What is the difference between a skill and a subagent?
A skill is reusable instructions that load into your conversation. A subagent is a separate worker with its own context that returns a summary.
Do I need MCP to use Claude Code?
No. You can start without it. MCP becomes useful when you want Claude to read from or act in other tools directly.
What is the difference between CLAUDE.md and a skill?
CLAUDE.md loads every session and holds always-on rules. A skill loads on demand and suits reference material or workflows you trigger by name.
Can a hook replace a skill?
They do different jobs. A hook runs automatically on an event, while a skill is a set of instructions Claude uses when invoked or relevant.
Need help building this?
We design and review AI workflows for marketing teams, including the guardrails. See our AI and automation service or contact us.