The Four C's framework

Module: 1, Foundations of thinking | Time: about 25 min theory + 20 min practice

The gist

Picture your business as an apartment. You know where everything is, you remember whom to call and what to do in any situation. Now put an AI assistant in that apartment. Without preparation it's like a person with amnesia: standing in the middle of the room, not even knowing its own name. The Four C's are instructions for turning a blank Claude Code into a smart partner that knows your business, pulls the data you need, can actually do things, and works while you sleep.

Key concepts

Context (what the AI can see in the conversation): what the AI knows about you, your business, your voice and your goals
Connections: which tools and data the AI can access (MCP, the Model Context Protocol; APIs, application programming interfaces; databases)
Capabilities: what the AI can do with that data (Skills, which are reusable instructions; SOPs; agents, which are autonomous helper programs)
Cadence: when the AI acts on its own: schedules, triggers, routines
AI OS: an operating system built on the Four C's, a smarter partner than an empty chat
CLAUDE.md: the master file where all four C's live

Theory

C1, Context: the system's brain

🎨 Picture this: Claude without Context is like a new doctor you're seeing for the first time. Smart and experienced, but forced to start from scratch: "What brings you in? Any chronic conditions? Any allergies?" Context is your medical chart, which the doctor reads before the appointment.

Open a new Claude Code session right now and ask it something about your business. It will answer like a stranger you met five seconds ago. Because without Context, it knows nothing.

What goes into Context:

Who you are and what you sell
Whom you sell to (your ICP, ideal customer profile)
Your communication style (voice, tone, sample texts)
Priorities for the quarter / sprint
An archive of past decisions

Where it's stored: a context/ folder in the project, made of Markdown files (Markdown is a simple text formatting language). For example:

Code

context/
  about-me.md        # who you are
  about-business.md  # what you sell, to whom, how
  priorities.md      # Q3 goals
  voice-samples.md   # samples of your writing

The pass test: ask Claude Code "what should I focus on this week?" If it answers like your partner and not like a search engine, Context is set up.

The fail test: the answer starts with "I don't know the details of your business, but..." That means Context is empty.

C2, Connections: the system's hands

🎨 Picture this: Connections are like giving an employee access to your systems. A smart person without the password to the CRM (Customer Relationship Management) is useless for working with clients. Connections give Claude "hands" to reach your business's real data.

By default, Claude Code works with project files, the command line and web search. But your important business data isn't on the internet: it's in ClickUp, Notion, Google Sheets, your CRM, your email, your databases.

Connections = bridges between Claude and your tools.

Three ways to connect a tool:

Method	What it is	Pro	Con
MCP server	A ready-made plugin	Fast, works out of the box	Lots of tokens (tokens are the smallest units of text for AI); you don't always need every function
API endpoint	A direct request to the API	Fewer tokens	You need to study the documentation
Browser automation	Controlling a browser	Works anywhere	Slower, more fragile

A recommendation from practice: for many tools an API comes out cheaper in tokens than an MCP. Ask Claude Code to research the tool's documentation and save a list of endpoints to a Markdown file. After that it will read this file instead of going online every time.

Storing keys: never paste API keys into the chat. Only through a .env file:

Code

CLICKUP_API_TOKEN=pk_xxxxx
CLICKUP_TEAM_ID=12345678

Seven priority areas for connections:

Area	What goes here	Why
Revenue	Stripe, QuickBooks, Lemon Squeezy	P&L, runway, growth
Customers	CRM, Skool, course platform	Health of the customer base
Calendar	Google Calendar, Calendly	What's scheduled
Comms	Gmail, Slack, ClickUp DM	Incoming messages
Tasks	ClickUp, Notion, Linear	Priorities, deadlines
Meetings	Fireflies, Otter.ai	Transcripts, action items
Knowledge	Local files, YouTube transcripts	Context from the past

C3, Capabilities: what the system can do

🎨 Picture this: Skills are like recipe cards in a kitchen. Without a recipe, the cook reinvents the dish every time and gets a different result. With a recipe, it's consistent, repeatable and can be improved. A Capability turns a one-time "did it well" into an ongoing "always does it well."

Connections give access to data. Capabilities define what to do with it.

Three levels of Capabilities:

Level 1, Skills (recipes): A folder in .claude/skills/ with a SKILL.md file that lays out step-by-step instructions for a specific task. For example, a "linkedin-post" skill: research the topic → find three angles → write a draft → check it against your voice.

Type this into the chat

.claude/skills/
  linkedin-post/SKILL.md
  weekly-report/SKILL.md
  client-onboarding/SKILL.md

You can call a skill as /linkedin-post, or let Claude pick it up on its own from the description. Your own commands in .claude/commands/, if you already have some, are merged with skills and keep working.

Without a skill → Claude guesses, and the result is unpredictable. With a skill → Claude follows the recipe, and the result is consistent.

Level 2, SOPs (standard operating procedures): If you already have an SOP for some process (client onboarding, a weekly report), give it to Claude Code and ask it to turn it into a skill. You get automation of a process you already have.

Level 3, Agents (specialized workers): Subagents that run in parallel for specific specialized tasks. More on this in the lessons on agents.

The sign of good Capabilities: instead of writing a two-paragraph brief every time, you say "do the quarterly report" and the system does it on its own, because it knows the SOP.

C4, Cadence: the system works without you

This is the most exciting part. Cadence = the schedules and triggers that run the system while you're away from the computer.

Three types of Cadence:

Schedule:

Code

Every morning at 7:00 AM → the "morning briefing" routine runs
Every Monday → the "competitor analysis" routine runs
First day of the month → the "financial report" routine runs

Event trigger:

Code

A new client lands in the CRM → the "client onboarding" skill runs
A metric drops in Stripe → an alert arrives
A new comment → the "reply to comment" skill runs

Webhook (an HTTP notification about an event; an external trigger): Another tool (Zapier, Make, GitHub) sends a request → Claude Code carries out the task.

⚠️ What's available for this as of October 2026. Routines handle schedules, API calls and GitHub events: they run in the cloud, even when your computer is off (create them with /schedule or at claude.ai/code/routines; available on paid plans, in research preview). For schedules on your own computer with access to local files, there are scheduled tasks in the Claude app, and /loop repeats a request inside an open session. The current list of features: What's current.

Important about Cadence: without Context, Connections and Capabilities in place, Cadence is pointless. You can't put an empty system on a schedule. The order is strict: C1 → C2 → C3 → C4.

🎨 Picture this: the Four C's are like hiring an employee. First explain who you are and what the company is (Context). Then give them access to the systems they need (Connections). Then train them to do specific tasks (Capabilities). Only then can they work on their own, on a schedule, without you (Cadence).

How the Four C's live in CLAUDE.md

CLAUDE.md is the project's master file (created with the /init command). Claude Code reads it at the start of every session, and it also picks up an AGENTS.md file. Without them, it starts from zero every time.

The structure of CLAUDE.md in Four C's terms:

Type this into the chat

# [Name] AI Operating System

## Who I am (Context)
You are my personal AI OS. You help me think, decide and act faster.
Here's what you know about me: [link to context/about-me.md]

## My tools (Connections)
- ClickUp: tasks and projects [see references/clickup-api.md]
- Gmail: inbox [MCP: gmail-mcp]
- Google Calendar: schedule [MCP: gcal-mcp]

## What you can do (Capabilities / Skills)
- /linkedin-post: write a LinkedIn post
- /weekly-report: weekly report
- /audit: check the state of the AIOS against the Four C's

## Where things are (Folder Structure)
- context/: information about me and the business
- references/: API documentation for tools
- skills/: instructions for specific tasks
- outputs/: work results
- archives/: archive of old material

🎨 Picture this: the Four C's are like opening a restaurant. First the menu (Context: what you cook). Then the food suppliers (Connections: where you get ingredients). Then recipes and standards (Capabilities: how you cook). And only then the opening hours (Cadence: when you're open). Opening a restaurant with posted hours but no suppliers is absurd.

A Four C's audit: how to check that everything works

The /audit skill is an example of a custom skill (Claude Code has no such command by default; you can build one using the outline in this lesson). It scores your AI OS on each C out of 25 points (100 in total).

A sample result on day 1 after setup:

Type this into the chat

Context:     18/25  ✓ basic information is there
Connections:  16/25  ⚠ only ClickUp connected out of seven areas
Capabilities: 12/25  ⚠ no custom skills
Cadence:       8/25  ✗ not a single routine

TOTAL: 54.5/100
Priority 1: connect Fireflies (meeting transcripts)
Priority 2: create your first skill
Priority 3: set up at least one routine

54 out of 100 on day one is normal. The system grows along with you.

Practice

Exercise: sketch your own Four C's map

Open a blank document (Google Doc, Notion, paper, doesn't matter)
Under each C, write what you already have:
- Context: what should Claude know about you and your business? Write 5–7 points
- Connections: list 7 tools you check every week (Revenue, Customers, Calendar, Comms, Tasks, Meetings, Knowledge)
- Capabilities: which recurring tasks do you do every week? Write 3–5 of them: these are candidates for Skills
- Cadence: which of those tasks would you like to run automatically? When?
Open Claude Code and ask a question about your business: see whether it answers like a partner or like a stranger
Tell Claude Code: "I want to set up an AI OS using the Four C's. Create the folders context/, references/, skills/, outputs/ and a draft CLAUDE.md, then ask me questions to fill in the Context"

Goal: understand your Four C's map before setup, so onboarding is thoughtful rather than mechanical.

Common mistakes

❌ Mistake: Starting with Cadence (schedules) before setting up Context and Connections. ✅ Right: Strictly in order: C1 (Context) → C2 (Connections) → C3 (Capabilities) → C4 (Cadence). Automation without context and connections is an empty shell.

❌ Mistake: Pasting API keys straight into the Claude Code chat or into claude.md. ✅ Right: Keys go only in a .env file. The code Claude Code writes for you reads keys from .env, and they never end up in the chat. Add .env to .gitignore so the keys don't end up in git (the version control system for code).

❌ Mistake: Connecting every possible MCP server "just in case." ✅ Right: Every MCP consumes tokens on every request. Connect only what you actually use. For rarely used APIs, it's better to save the documentation in a Markdown file: the agent will read it only when needed.

Tools and resources

CLAUDE.md: the project's master file, read at every start → See the lesson CLAUDE.md: your project's system prompt
context/: a folder with information about you and your business
references/: a folder with API documentation for connected tools
.env: a file with keys and tokens (never in the chat)
/audit: an example custom skill for scoring the Four C's (0–100)
/onboard: an example custom skill for initial setup through 7 questions
Claude Code docs: official documentation
MCP Protocol: the Model Context Protocol spec, for Connections

→ See the lesson Installation and setup: how to install and set up your workspace → See the lesson CLAUDE.md: your project's system prompt: a detailed guide to the system prompt → See the lesson The WAT framework: the structure of an agent project → See the lesson What Skills are: how to create Capabilities for your system

Key takeaways

The Four C's aren't just a framework. They're a strict order: you can't jump to Cadence without Context. A system that works while you sleep is the end of the road, not the start.

The quality test for Context is simple: ask a question about your business. If Claude answers like a stranger, you need more Context. If it answers like a partner, you're in good shape.

For many tools, Connections are cheaper to build through an API than through MCP: fewer tokens. One well-configured API is worth more than five MCP servers eating up your context.

Next lesson

→ Installation and setup: Claude Code and VS Code (a popular code editor from Microsoft)