By the ConnectLabz Systems team · 23 September 2026
Ask most marketers how they sorted out Claude skills vs projects and you’ll find they started with the piece that should come last. You watch a demo where “an AI agent ran the whole campaign,” open Claude, and find four things with confusing names: skills, projects, connectors and agents. Forty minutes of developer documentation later, you paste your brand guidelines into a chat. Again.
This guide untangles all four in plain marketing terms. You’ll get what each one does and a quick test for telling two of them apart. Then a sensible sequence for this week, so the agent you were promised actually has something good to run.
Key Takeaways
- A skill is a playbook Claude loads when a task matches it; a project holds what’s true about one client or initiative.
- MCP connects Claude to live data, and an agent works through a task by itself. Neither one tells Claude what good work looks like.
- Set them up in this order: skill, project, MCP, agent. An agent can only amplify the playbook it’s given.
- Two layers are enough to start: a single skill plus a single project. Skills work on Claude’s individual plans, Free included, once code execution is switched on.
What Skills, Projects, MCP and Agents Each Do
Anthropic’s own explainer, Skills explained (5 March 2026), defines all of these. Here they are translated for marketing work.
What is a Claude skill?
A Claude skill is a folder holding a playbook: a name, a one-line description of when to use it, and step-by-step instructions, plus any examples or scripts. Claude reads only the name and description until a request matches, then loads the full instructions. The result: one job, done the same way every time.
Think of the brief template you rewrite from memory for every new campaign. As a skill, it gets written once, properly, and Claude reaches for it whenever you ask for a brief.
What is a Claude project?
A Claude project is a workspace with its own chats, custom instructions and uploaded documents, all about one initiative or client. Put simply, a project tells Claude what’s true in that context. In Claude Code, a project instruction file (usually called CLAUDE.md) sitting in the folder you work in does the same job.
Your client’s price list, their three service areas, the claims legal won’t let them make: that’s project material.
What is MCP?
MCP, the Model Context Protocol, is an open standard for connecting Claude to the places your data lives, such as analytics, a CRM or Google Drive. With it, Claude reads current numbers instead of working from memory or from whatever you pasted in last week.
What are Claude agents and subagents?
An agent is Claude running a sequence of steps on its own: planning, doing, checking, adjusting. A subagent is a specialised helper with its own instructions, its own working memory and limited tools, handed one part of a bigger job. Agents execute. They don’t decide what good looks like.
That last sentence is the one to remember, and the next two sections explain why.
Claude Skills vs Projects: The What-to-Know / How-to-Do Test
When you’re unsure where something belongs, ask one question. Does it describe this client, or how you always do a kind of job?
If it describes this client or this campaign, it goes in the project. Describing how you do a type of work, no matter who it’s for? That becomes a skill.
| It’s about this client (project) | It’s about how you work (skill) |
|---|---|
| “They sell in Texas, Oklahoma and Arkansas.” | “How we write a local-service landing page.” |
| “Never mention financing; legal said no.” | “Our five-step ad review before anything launches.” |
| “Their best month is March.” | “How we turn a month of numbers into a one-page report.” |
Anthropic draws the line the same way: projects provide what to know, and skills teach how to do things. Getting this right is what stops the classic mess, where a client’s price list ends up baked into your landing-page playbook and then quietly shows up in another client’s draft.
How to Set Up Claude for Marketing in the Right Order
Here’s the order we’d follow if we were starting from nothing on Monday. Each step makes the next one worth more.
Step 1: Write the playbook as a skill
Start with the job you repeat most. For most marketers that’s the brief template, so write it the way you’d train a sharp new hire: what goes in a brief, in what order, and what a bad one looks like.
Spend real care on the description line, since it’s all Claude sees before deciding a skill is relevant. “Helps with marketing” will almost never fire. “Use when writing a campaign brief for a local-service client” will. Don’t worry about installing several over time, either. Only the relevant skill’s instructions get pulled in, so a library of twenty playbooks doesn’t crowd the conversation.
Where people slip: writing the skill around one client. Once it mentions a specific price or place, it has started doing the project’s job.
Step 2: Give it a project to work inside
Now make one project per client. Put in the facts the playbook needs to do a good job for them: offers, service areas, tone notes, banned claims, their best-performing past work.
With the skill and the project together, “write a brief for the spring campaign” produces something in your structure and in the client’s reality. Either one alone gets you half of that.
Watch for this: using the project as a dumping ground. A hundred uploaded files dilute the ten that matter, and Claude has no idea which are which.
Step 3: Connect live data with MCP
Reports are where people feel the gap first. Without a connection, Claude writes from last month’s numbers you pasted in, or from nothing, and the report comes back confident and stale. A connection through MCP gives it a proper line to the source: your analytics, your CRM, your shared drive.
Connect read-only first. A connection that can only look is safe to learn with. One that can change things deserves far more thought. A confident assistant with write access can edit a live campaign faster than you can notice.
The trap here: connecting everything at once. Start with the one source your weekly report actually needs.
Step 4: Only then hand it to an agent
This is where the demo magic happens, and it’s genuinely useful. Hand an agent “prepare next week’s campaign” and it can run the brief, the drafts, the checks and the report in sequence, handing parts to subagents as it goes.
But an agent only multiplies what’s underneath it. Give one a strong playbook, a well-kept project and live data, and you get a week of good work in an afternoon. Give it none of those and it will cheerfully produce twenty posts that sound exactly on brand and promote an offer the client ended in July.
What trips people up: judging the agent by how fast it works. Judge it by whether it stops when a step is missing.
What This Looks Like Inside a Working System
Our own SEO system is built in exactly these layers, which is the easiest way to see how an AI marketing system fits together.
Each job has its own skill folder: research, outlining, writing, checking. A standing instruction file tells the whole system how to behave, and a separate configuration file per client holds that client’s details, so their facts never leak into the playbooks. For data, it reads Google Search Console through a read-only connection. (Ours is a direct connection rather than MCP, but it does the job MCP standardises.) Split the work across subagents and each one hands over a written note carrying every number, name and approved line. Summaries, we found, quietly drop exactly the details the next step needs.
A confession, since it’s useful: those hand-off rules were among the last things we added, and they’re still the part that needs the most rules to stay honest. That’s the agent layer for you. Most powerful piece, easiest to get wrong.
Rather start from a finished stack than build one? Our SEO Growth System ships with all four layers already in place.
Where to Start If You Only Set Up Two
One skill and one project. That’s it for the first week.
Write your brief template as a skill, and make a project for your biggest client. Use the pair for every brief you write. You’ll feel the difference by Friday, and you’ll also find out which parts of your brief were never written down, which is useful all on its own.
Add MCP when a report keeps going stale. Add an agent when the same multi-step job repeats every week and you’re tired of starting each step by hand.
What These Pieces Still Can’t Do
A few honest limits, so nothing surprises you.
A skill with a vague description may never trigger, and installing lots of skills at once makes Claude’s choices less predictable. Grant a connector permission to write and it can change live things you didn’t mean it to touch. Agents can loop on a task or stop early when an instruction is ambiguous. And none of the four knows your offer is weak. They’ll execute a bad plan beautifully.
Conclusion
Agents are genuinely the most powerful of the four. They’re also the worst place to start, and both of those things are true at once. Build the playbook first, give it context, connect the data, and the agent you add last will finally have something worth amplifying.
Now you know the order. What’s left is the part no article can do for you: turning your own best work into playbooks. Start with a finished playbook: see the five ConnectLabz systems →
FAQ
Are Claude skills available on the free plan?
Yes. According to the Claude Help Center, skills come with the Free, Pro and Max plans, so you can try them without paying. You manage them under Customize, and code execution has to be switched on in your settings, because skills run in that environment.
Do I need MCP to use Claude skills?
No. A skill works on its own with whatever you give it in the conversation or the project. Adding MCP makes sense once a skill keeps needing current data, such as this week’s ad results, that you’d otherwise have to paste in by hand.
Claude skills vs agents: which should I use?
Use a skill for any job you repeat, and add an agent only when you want several of those jobs run in sequence without starting each one yourself. Without skills, an agent works fast but generically; skills without an agent work well, one step at a time.
Are projects and skills the same thing?
Not quite. A project is where one client’s documents, facts and custom instructions live. A skill captures a method you’d use for any client. The quickest test: if it names a particular client, a price or a town, it belongs in the project.
Can I use Claude skills outside Claude?
Often, yes. Skills follow the open Agent Skills standard, which other tools such as Gemini CLI, Cursor and GitHub Copilot also read. We compare them with Gemini Gems and custom GPTs in our guide to Claude skills vs Gemini Gems.
