The Four-Layer Stack for Claude

Most teams use Claude wrong. They open a chat window, type a prompt, copy the output, paste it somewhere else, and call that "using AI." That works. It also wastes 90% of what Claude can do.

The teams getting real leverage aren't smarter. They've just built infrastructure. Four specific layers, each doing one job, each compounding the others. We call it the four-layer stack: Projects, Skills, MCP, and Cowork.

This is what separates "I use Claude sometimes" from "my team runs on Claude." It's also why some marketing teams report 5-10x output gains while their peers struggle to save an hour a week. Same tool. Different infrastructure.

This chapter walks you through each layer. By the end you'll know what each one does, how they fit together, and where to start building.

The Stack as Real Infrastructure

If you've ever worked with a really good freelancer, you know the pattern. The first few projects are painful. You explain your brand voice. You send the style guide. You answer the same questions twice. You re-upload the same logo files. You re-share the same access credentials.

By month three, none of that is happening. The freelancer knows your brand. They have access to your tools. They follow your process. You send a one-line brief and a finished asset comes back.

What changed isn't the freelancer. What changed is the infrastructure around them.

That's the stack. Projects are the brand book and asset library you handed over. Skills are the playbook you wrote together. MCP is the access you granted to Slack, Drive, and Shopify. Cowork is the freelancer working without you in the loop.

You're not just using Claude. You're onboarding it.

Layer 1: Projects (Persistent Context)

A Project is a workspace inside Claude that remembers everything you put into it.

Anthropic launched Projects with a simple promise: stop re-uploading the same files, stop re-explaining your context, stop starting from zero every conversation. Each Project holds a knowledge base (with a 200K context window, roughly 500 pages of content), custom instructions, and a chat history that all stays put.

For a CPG marketing team, one Project can hold the artifacts that define how you work.

Your brand voice and tone guide.
Your customer personas with demographic and psychographic detail.
Past campaign briefs and the creative they produced.
Product positioning and category rationale.
Top-performing assets and what made them work.
Your standard formatting preferences for briefs, posts, and emails.

Every conversation inside that Project starts with Claude already knowing all of it. You don't have to explain who your customer is again. You don't have to remind Claude that you spell it "Co-Wash" not "co wash." You ask for an Instagram caption for the new SKU and Claude already knows the brand voice, the audience, and the campaign positioning.

At Tymoo, we run a separate Project for each brand we work with. The Project holds the persona documents, the content waves, the brand visual system, every previous brief, and standing instructions on what we mean by "in voice." When a writer kicks off a new blog post, Claude isn't starting from scratch. It's starting from where we left off last week.

The mistake most teams make is treating Projects as folders. They aren't. They're context. The point isn't to organize files. The point is to teach Claude your business once and have it remember forever.

Three things to optimize when you build a Project.

Curate aggressively. Garbage in, garbage out. Don't dump every PDF you have into the knowledge base. Pick the documents that actually define how you work.
Write strong custom instructions. The instruction field at the top of every Project is where you turn Claude into a specialist. Be specific about voice, format, and the words you never want to see.
Update as you learn. Every Project should evolve. New persona research, new voice rule, new campaign learning, all of it goes back in. The Project is a living asset, not a snapshot.

Done well, a Project is the single biggest leverage point in the stack. Skip it and every other layer is doing more work than it should.

Layer 2: Skills (Reusable Workflows)

Skills are the playbooks Claude runs automatically when relevant.

Anthropic shipped Agent Skills in late 2025, and they're the most underrated layer in the stack. A Skill is a folder containing a SKILL.md file with instructions plus any supporting files (templates, scripts, examples). When Claude detects a task that matches a Skill, it loads the Skill's instructions and follows them.

The shorthand: prompts are what you ask. Skills are what Claude knows how to do.

For a CPG marketing team, Skills can encode the workflows you already run manually every week.

Writing a content brief from a raw topic idea.
Formatting a blog post to brand standards with NLP keyword integration.
Building a creative brief in your team's exact template.
Optimizing a PDP against the 10-element rubric.
Generating image prompts for featured images in your brand's visual style.
Adding source citations and contextual internal links to a draft.

The win is that Skills run consistently. Every blog post your team produces hits the same NLP keyword density. Every brief follows the same structure. Every PDP audit scores against the same rubric. The variance that comes from "did the writer remember to do step 7" disappears.

We've built more than thirty Skills at Tymoo. The blog-workflow-coordinator skill, for example, takes a raw topic and runs eight steps in sequence: extract content, fix flow, optimize chunking, integrate NLP keywords, format, add internal links, add citations, grade against a 14-category rubric. The whole sequence runs without a human touching it between steps. That's the kind of leverage Skills create.

The trick to good Skills is the same as good SOPs. Write them like you're explaining the job to a new hire on day one. Be specific. Include examples. Tell Claude what good and bad look like. The skill-creator Skill that Anthropic ships with Claude will walk you through building your first one.

Skills compound with Projects. The Project gives Claude your brand. The Skill gives Claude the procedure. Together you get consistent, on-brand output without prompt engineering every time.

Layer 3: MCP (Tool Integration)

MCP is how Claude reaches outside its own walls.

Model Context Protocol launched in late 2024 and is now the de facto standard for connecting AI agents to external tools. Anthropic donated it to the Linux Foundation in December 2025. ChatGPT adopted it. So did most of the rest of the industry. It's the USB-C port of AI infrastructure.

For a marketing team, MCP is what lets Claude actually do work in your existing stack. Not "give me a Notion-formatted output you can paste in." Actually create the Notion page. Actually post the Slack message. Actually pull the data from Shopify.

The MCP servers worth knowing about for CPG marketing teams cover the systems where your work actually happens.

Notion for creating pages, updating databases, and pulling wiki content.
Slack for posting messages, drafting messages, and reading channels.
Gmail for searching threads, drafting replies, and managing labels.
Google Drive for finding docs, reading content, and creating files.
Airtable for reading and writing records across your bases.
Shopify for managing products, pulling orders, and running analytics queries.
Linear or Asana for creating issues and updating tasks.
SEMRush for running keyword research and pulling traffic data.
HubSpot or Klaviyo depending on your CRM and email stack.

Each connection turns a manual step into an automated one. Without MCP, asking Claude to "draft a brief and put it in Notion" means Claude writes the brief and you copy-paste it over. With MCP, Claude writes the brief and Notion has a new page when you're done.

The compounding effect with the rest of the stack is significant. A Project tells Claude what you sell and to whom. A Skill tells Claude how to write a brief. MCP lets Claude pull live SEMRush data into that brief and drop the finished version into your Notion content calendar. No copy-paste. No tab-switching. No human handing assets between systems.

One word of warning. MCP gives Claude real access to your business systems. Permissions matter. Start with read access. Add write access deliberately. Audit which servers each Project can reach. The skill at Tymoo isn't "connect everything." It's "connect what's needed and nothing else."

Layer 4: Cowork (Autonomous Execution)

We covered Cowork in Chapter 1. Here's how it sits on top of the other three layers.

Cowork is the desktop app that turns a stack of layers into an autonomous worker. The Project gives Cowork context. Skills give Cowork procedures. MCP gives Cowork hands. Cowork itself is the part that walks across all of them, makes decisions about which to use when, and produces a finished deliverable.

Without Cowork, you can absolutely run Projects, Skills, and MCP. You just have to be in the chat window driving each step. "Run the brief skill. Now pull SEMRush data. Now update the Airtable record." That works, but you're the orchestrator. You're spending your time on the boring connective tissue between Claude's actual capabilities.

Cowork removes that. You give it the outcome. It figures out the sequence. It runs the Skills, pulls the data through MCP, references the Project context, and hands you back the finished thing.

The legal team at Anthropic, with zero coding experience, used this combination to cut contract review from three days to twenty-four hours. They didn't write code. They built up the stack. Then Cowork made it autonomous.

This is why Cowork without the other layers is a curiosity. And why the other layers without Cowork are nice-to-have, not infrastructure.

How the Layers Compound

The four layers aren't additive. They're multiplicative.

A team running only Projects (persistent context) gets maybe 20-30% productivity gains. Better outputs because Claude knows the brand, but you're still copying and pasting everything.

Add Skills (reusable workflows) and the gains double. Now the outputs are consistent and you're not re-prompting the same procedure every time. Call it 50-70% gains.

Add MCP (tool integration) and you start removing tab-switching. The outputs land where they need to land. Call it 100-150% gains for tasks that touch multiple systems.

Add Cowork (autonomous execution) and the gains compound again. Now full workflows run end-to-end without human orchestration. Anthropic's own internal data shows 50% productivity gains in roles using the full stack. Some teams report much higher.

The math matters because most marketing teams stop at one or two layers. They build a Project, declare victory, and wonder why they're saving an hour a week instead of a day. The leverage is in the stack, not in any one piece.

The Tymoo Stack

Here's what our setup actually looks like, for transparency.

We run a Project per brand client. The Project holds the persona documents, brand voice guide, content waves, past creative briefs, and standing instructions. Custom instructions cover writing style, brand-specific terminology, and approval workflow.

We've built more than thirty Skills covering the work we do most often. Creating content briefs from raw topics. Generating featured image prompts. Optimizing blog drafts against SEO and AI visibility rubrics. Building PDP audits. Generating personas from interview transcripts. Drafting pre-meeting reminder emails. Processing client meeting notes into Linear tickets and follow-up emails.

MCP connects the stack to Notion (where we store briefs, personas, and meeting notes), Airtable (where we track every blog post in the pipeline), Slack (where we coordinate with clients), Gmail (where the client conversations live), SEMRush (for keyword research), Linear (for project management), and Shopify (for client e-commerce data).

Cowork is what runs the daily operations. A new client meeting wraps. The meeting-notes Skill kicks off, pulls the transcript from Notion, updates Linear with new tickets, drafts a client follow-up email, and posts a summary to Slack. We weren't in the loop for any of that. The full sequence runs in about four minutes.

That's the stack. The pieces aren't exotic. The discipline to build them up over time is what matters.

Where to Start

You don't build the full stack on day one. You build it the same way you build any operations layer. One piece at a time. Each piece earning the right to the next.

Here's the order we recommend.

Week one: spin up a single Project for one team or one brand. Put your most-used reference documents in it. Write strong custom instructions. Use it daily.
Weeks two and three: identify the three procedures you run most often. Build a Skill for each. Don't optimize for elegance. Optimize for "Claude does this consistently now without me re-prompting."
Month two: connect one or two MCP servers. Whichever systems you spend the most time copy-pasting in and out of. Probably Notion or Slack for most marketing teams.
Month three: install Cowork. Run your first end-to-end workflow autonomously. Pick something with a clear input and output, not your most complex process. Build trust in the system before delegating the hard stuff.

By month four, you're operating on infrastructure. The team is shipping more. The variance in quality has dropped. The boring orchestration work is gone.

The next chapter is the one that keeps you from over-trusting all of this. We call it the jagged frontier. It's the specific places where Cowork helps, and the specific places where it tanks quality. Know that before you delegate anything that matters.