The Four-Layer System Running a $500K Business, Built Entirely on Text Files

I run a company doing over $500K in revenue entirely on text files.

CLAUDE.md files, markdown documents, JSON configs are the infrastructure. There's no proprietary platform or enterprise software with a hundred-seat license. It's just text files in folders, read by AI, executed at speed.

I've trained more than 1,000 people across 25 cohorts. I manage marketing, sales, customer delivery, operations, SEO, a newsletter, podcast outreach, journalist outreach, speaking outreach, event production, lead generation, and content creation with a team of under 10 people.

This is the secret, four-layer operating system that allows me to do it.

I built it all with Claude Code, and I'm going to walk you through exactly how it works.

The SaaS Stack That Was Slowing Me Down

Before this system, I had the same problem every small company has. Too much to do, not enough people, not enough time. Every task required me to open a different tool, remember a different workflow, and context-switch between dozens of things.

I tried the SaaS approach: CRM for deals, project management for tasks, mail tool for outreach, analytics dashboard for metrics, etc. Each tool had its own login, its own interface, its own learning curve.

And none of them talked to each other.

I'd spend an hour setting up a lead in the CRM, then have to manually copy that same information into my outreach tool. I'd draft a newsletter in one app, then log into another to schedule it. Every process was disconnected.

So I stopped trying to find the right tool, and started building a system. And the foundation of that system turned out to be a text file called CLAUDE.md.

Layer 1: Master Guides (CLAUDE.md Files)

A CLAUDE.md file tells Claude Code how a part of your business runs: the steps, the standards, the things not to do. It's a plain markdown file that sits directly in a folder, and Claude Code reads it automatically before doing anything else in that folder.

Global CLAUDE.md. This file lives at the top level and applies everywhere. Mine contains my personal workflow conventions, data integrity rules, brand asset guidelines, and task completion standards. Every time Claude Code starts working on anything, it reads this file first. It knows how I want things done before it touches a single file.

Project CLAUDE.md. Each major folder in my business, marketing, sales, operations, customer delivery, has its own CLAUDE.md, and each one holds the specific rules for that domain. The marketing file knows the content pipeline, the outreach workflows, the brand guidelines. The sales file knows the deal stages, the pricing rules, the CRM conventions.

Subfolder CLAUDE.md. Inside each project, specific workflows get their own CLAUDE.md too. LinkedIn, newsletter, SEO, podcast outreach all have one. Each one contains the exact instructions, templates, naming conventions, and quality standards for that specific workflow.

When Claude Code enters a folder, it reads the CLAUDE.md automatically and knows the rules before I say a word.

Here is a real example. My LinkedIn CLAUDE.md contains:

• Voice guidelines (first person, conversational, under Grade 5 reading level)
• Forbidden phrases (never say "AI-First," never say "use AI," use em dashes sparingly)
• Post structure rules (hook in first line, short paragraphs, end with engagement question)
• Posting strategy backed by data (0-1 hashtags max, Sunday outperforms other days)
• The exact Notion database IDs for syncing content to my content calendar
• The workflow for exporting visuals to PNG and uploading to our CDN

One text file. Everything Claude Code needs to produce LinkedIn content that sounds like me, follows my brand guidelines, and syncs to the right systems. I don't repeat these instructions every time. They're baked into the folder.

Multiply this by every workflow in my business. I have CLAUDE.md files for events, journalist outreach, speaking outreach, podcast outreach, newsletter production, SEO, visual assets, video production, YouTube scripts, lead generation, sales playbooks, customer delivery, and operations.

Each one is the accumulated knowledge of how that part of the business works. Written once, read every time, and continuously improved as I learn what works better.

Every time I improve a CLAUDE.md file, every future interaction in that folder gets better. I don't train people on new processes. I update a text file. The system absorbs the change instantly.

Layer 2: Skills (Pre-Built Workflows)

A skill is a pre-built workflow that I trigger with a slash command. Think of it as a recipe. I say "make this," and Claude Code follows a detailed set of steps to produce a specific output.

I currently have over 50 skills. Here are the three that show the range.

• /featured. This skill answers questions on Featured.com for SEO backlinks. When I trigger it, Claude Code reads my content inventory, finds a relevant question, drafts an expert answer in my voice, follows the word count requirements, and formats it for submission. The result? 22 published placements so far, from publications with domain authority ranging from 43 to 72.
• /podcast-outreach. This skill manages a 9-stage pipeline, from first identifying a podcast through outreach, scheduling, recording, and publishing. When I trigger it, Claude Code reads the current state of the pipeline, figures out what needs to happen next, drafts personalized outreach emails, and logs every interaction. One command, and the whole pipeline moves forward.
• /visual-asset. This skill generates branded HTML graphics, like cheat sheets, comparison charts, framework diagrams, stat graphics, and carousel slides. I give it a topic, and it produces a polished HTML file that matches my brand colors, uses my real logo, and follows my design conventions.

Each skill reads the CLAUDE.md files for context, which is why the layers build on each other rather than existing in isolation. A skill knows the brand guidelines, the naming conventions, and the quality standards because that's what the CLAUDE.md already established.

Building a skill is straightforward. You write a SKILL.md file that describes the workflow step by step. What to read, what to produce, conventions to follow, and quality checks to run. Then you register it so Claude Code can find it when you type the slash command.

The key insight: skills make complex tasks repeatable. I figured out how to write a great Featured.com answer once. I encoded that knowledge into a skill. Now I can produce them at scale without thinking through the process each time. The quality stays consistent because the instructions are consistent.

Layer 3: MCP Servers (Plugins)

This is the layer that connects the text files to everything else.

MCP stands for Model Context Protocol. In simple terms, an MCP server is a plugin that connects Claude Code to an external tool. Instead of being limited to reading and writing files, Claude Code can now interact with databases, websites, APIs, and services.

The three that do the most work in my system:

1. Notion. My content calendar, customer databases, media tracking, and project management all live in Notion. The Notion MCP lets Claude Code create pages, update databases, query records, and sync content. When a LinkedIn draft is done, Claude Code pushes it directly to the Notion content calendar. When a deal moves forward in sales, Claude Code updates the Notion pipeline, without manual data entry.

2. Sanity. Our website runs on Sanity CMS. The Sanity MCP lets Claude Code publish content, update pages, manage images, and deploy changes to the live website. When I write an article, Claude Code can publish it to the website without me opening a browser.

3. Playwright. This is browser automation. Claude Code can navigate to any website, fill out forms, click buttons, take screenshots, and interact with web pages. I use this for scheduling LinkedIn posts, scraping data, completing directory submissions, and any task that requires browser interaction.

I also run MCPs for Slack, Zapier, and Apify, handling team notifications, calendar and email automation, and web scraping at scale.

On its own, Claude Code is a smart text window that can read your files and write new ones. With MCPs, it becomes an operating system. It can understand your context, connect to your tools, and execute across your entire business.

A real example: I tell Claude Code to prepare for a webinar. It reads the event context file (Layer 1: CLAUDE.md for events). It follows the event landing page skill (Layer 2). It generates the HTML page, deploys it through the Sanity MCP, creates a Notion page for tracking registrations, schedules reminder posts through the LinkedIn workflow, and sends invite emails through the Zapier MCP. One command, five systems touched, no browser opened.

In short, MCPs turn intelligence into action.

Layer 4: Agent Teams (Parallel Execution)

Everything up to this point runs sequentially. One Claude Code instance, working through tasks one at a time. Agent teams change that.

An agent team is multiple Claude Code instances working at the same time, coordinated by a team lead. They share task lists, send messages to each other, and work in parallel.

Here is how it works:

• I create a team with a task list.
• The team lead assigns tasks to agents.
• Each agent works independently on its assigned task.
• When an agent finishes, it reports back.
• The team lead coordinates the results.

Use Case: SEO Orchestration

My SEO orchestration system is a good example.

I've created six specialized agents, each focused on a different SEO domain. A technical expert runs audits and fixes meta descriptions, while a content expert atomizes one guide into 20+ distributable assets. On the outreach side, a directory expert handles submissions on platforms like DesignRush and Clutch, and a link building expert writes Featured.com articles and pitches journalists. A competitive expert tracks what competitors rank for that I don't, and a reporting expert generates the daily, weekly, and monthly performance reports.

Before this system, running that same six-domain SEO process myself meant working through each one in sequence, one after another, over the course of a day. Now all six run at once, and I get the combined output in the time it used to take to do just one.

The team lead coordinates everything, making sure agents don't duplicate work and resolving dependencies (if Agent B needs something Agent A produces, Agent B waits). It collects the final outputs and presents them together.

Agent teams are the multiplier.

How the Layers Work Together

The layers are not independent. They stack.

A skill triggers (Layer 2). The skill reads the CLAUDE.md for context (Layer 1). It connects to MCPs for external tools (Layer 3). And it can spawn an agent team for parallel work (Layer 4).

Here is a complete flow.

I type /seo-link-building-expert find opportunities.

The skill activates (Layer 2). It reads the SEO CLAUDE.md for current targets, priority algorithms, and quality standards (Layer 1). It queries Notion for my existing backlink inventory through the MCP (Layer 3). It searches for new opportunities, cross-references domain authority scores, and drafts outreach for the top priorities. If I want to execute multiple outreach pieces at once, it spawns agents to work them in parallel (Layer 4).

With one slash command, four layers are activated, and multiple systems coordinated. Output produced.

Every layer builds on the ones beneath it. Better CLAUDE.md files make skills more accurate. More MCP connections give skills more reach. Agent teams give skills more throughput.

Why the System Outlasts Any App

When people ask me what tools I use, they expect me to name five SaaS products. Instead, I describe this system.

Because the tools are not the point, the architecture is. Any individual tool can be swapped out, but the CLAUDE.md files, skills, and agent team structures would survive.

The architecture is portable. The knowledge lives in text files, the workflows in skills, the connections in MCP configs. Nothing is locked inside a proprietary platform.

It's also honest about what it requires. You can't write a good CLAUDE.md for a process you don't understand, or build a good skill for a workflow you haven't done manually first. The system forces clarity.

That's why I've documented every major process in my business this way. The CLAUDE.md file IS the process. Claude Code reads it and does the work.

If you want more layer by layer breakdowns like this, sign up for the newsletter, where I go deeper on the systems behind AI Operator.

How to Start Building Your Own Version

To get started, create your first CLAUDE.md file.

Step 1: Write Your First CLAUDE.md

Pick the workflow you do most often. It could be writing emails, creating reports or updating a spreadsheet. Write down exactly how you do it, including the steps, conventions, quality standards, and common mistakes to avoid.

Put that in a CLAUDE.md file in the relevant folder. Next time you ask Claude Code to help with that task, it will read those instructions automatically.

Spend a week refining it. Every time the output isn't quite right, update the CLAUDE.md. Within a week, you'll have an instruction set that produces consistent, high-quality output for that workflow.

Step 2: Build Your First Skill

Take a multi-step process you repeat regularly. Break it into explicit steps. Write a SKILL.md file that walks through those steps. Register it as a slash command.

Start simple. A skill that drafts a specific type of document. A skill that formats data a certain way. A skill that produces a weekly summary from your notes.

The goal: take something you do manually every week and make it one command.

Step 3: Connect Your First MCP

Pick the external tool you interact with most. If you live in Notion, connect the Notion MCP. If you live in Slack, connect the Slack MCP. If you need browser automation, connect Playwright.

One connection is enough to start. Once Claude Code can reach outside its file system, the possibilities change dramatically.

Step 4: Experiment with Agent Teams

Once you have a few skills and MCP connections working, try spawning your first agent team. Give two agents related but independent tasks, then watch them work in parallel.

The Setup Cost Is Real, But the Payoff Compounds

Writing good CLAUDE.md files takes thought. Building reliable skills takes iteration. Coordinating agent teams takes practice. It can take weeks of refining before any of it runs smoothly.

But once it's set up, the returns compound. Every CLAUDE.md file you improve makes every skill in that folder better. Every MCP connection makes every skill more capable. None of that upfront work has to be redone, it just keeps paying out.

Compare that to the alternative: doing everything manually, hiring more people to cover the gap, or running 20 SaaS tools that don't talk to each other. After a year of building this, the math isn't close. My entire business runs on text files, and it produces more output than a team five times my size, built from four layers working as one system.

I teach this system, layer by layer, inside the AI Operator cohort. If you want it built for your business, book a call.

Frequently Asked Questions

What is a CLAUDE.md file?

A plain text file that lives in a folder and tells Claude Code how to work there: the steps, the standards, the mistakes to avoid. Claude Code reads it automatically before doing anything else.

Do I need to code to build this?

No. If you can write clear instructions for how you do a task, you can build a CLAUDE.md file or a skill.

CLAUDE.md file vs. skill: what's the difference?

A CLAUDE.md file is context. A skill is action. One tells Claude Code how your business runs, the other triggers a workflow that produces a specific output.

Do I need an MCP server to start?

No. Start with a CLAUDE.md file and a skill using just your files. Connect an MCP once you're ready for Claude Code to reach outside your file system.

How long does setup take?

Weeks, not days. A working CLAUDE.md file takes about a week of refining. Skills and MCPs layer on after that, and none of the work has to be redone.

Does this only work with Claude Code?

I built it with Claude Code, but the core idea, plain-text instructions instead of scattered tools, works with any AI assistant that can read files consistently.

Does this work for a team, or just solo founders?

Better with a team. Mine runs under 10 people across marketing, sales, delivery, and operations. The instructions live in the files, not in your head, so everyone gets the same consistent output.