How to Set Up Claude Code: Complete Guide for Founders (2026)

Most founders who try Claude Code give up within a week. Not because it doesn't work. Because nobody explained the full system — just the surface layer.

This guide covers everything in the right order: sandbox safety, memory that compounds, MCP connectors, skills, live dashboards, and scheduled tasks. By the end you'll have a working AI employee setup, not just a tool you open occasionally.

Claude Code vs Claude Chat: Brain vs Brain With Hands

This is the most important distinction to understand before anything else.

Claude chat is a brain. You ask it a question, it thinks, it responds. Smart. Useful. But you still have to do everything with the answer yourself.

Claude Code is that same brain with hands. It reaches into your actual files, your Gmail, your Notion, your databases, your connected tools — and comes back with a finished thing.

The same model. Completely different output.

When you understand that Claude Code has hands, you stop talking to it like it only has a brain. That shift in how you prompt it changes everything about the output quality.

The Output-First Prompt Formula

The most common mistake with Claude Code: describing a task instead of defining a deliverable.

Chat mindset (task-first): "Review my lead list and recommend a follow-up priority order." You get advice. You still do the work.

Claude Code mindset (output-first): "I have six inbound replies in my Gmail. Classify each as interested, not now, or not interesting. For every interested lead, write a recommended next action. Export the results as a CSV sorted by priority and save it to my Notion pipeline view." Walk away. Come back. Work is done.

The formula to use every time:

Here is what I have [files / data / source] + here is where it is [location / tool] + here is the finished thing [deliverable + format + quality bar + output location]

Two examples that make this concrete:

Prospect research: "I have 40 prospect research PDFs in my Leads folder. Give me an Excel with company name, headcount, funding stage, and ICP fit score 1–5. Flag any missing data as 'verify'."

Content strategy: "I have an Apify actor connected via MCP with a LinkedIn content scraper. Deliverable: find all posts with 500+ likes or 200+ comments from the past 30 days. Save content title, full body, and a summary of what patterns I can use — output as a structured Notion database entry."

Same tool. Completely different output quality. The formula is the difference.

Safe Setup: The 10-Minute Sandbox

The number one reason first sessions go wrong: people give Claude Code access to their entire computer and panic when it starts moving files.

Fix this in 10 minutes.

Step 1: Download the desktop app. Claude Code only runs in the desktop app, not the browser. Claude Code in your browser panel is different — you need the standalone app.

Step 2: Create a dedicated playground folder. Go to Documents or Downloads, create a new folder called "Claude Playground" or similar. In the app: Settings → choose different folder → select your new folder → "Allow Claude to change files in this folder."

That folder is now the sandbox. Claude Code will not touch anything outside it.

Step 3: Add global instructions. In Settings, find the global instructions field. Add these guardrails:

"Before deleting, overriding, or writing to any existing file, show me exactly what will change and wait for my confirmation."
"Always work only inside the folder I've pointed you to."
"When a task has multiple steps, list your plan first and wait for my go-ahead before executing."
"Flag anything you're unsure about as 'verify' instead of guessing."
"At the end of every task, tell me what changed and where the output lives."

These run every session automatically. No repeating yourself.

Step 4: Enable memory. Settings → Capabilities → enable both memory checkboxes. This is what makes Claude Code compound over time instead of starting from zero every session.

Step 5: Enable "Load tools when needed." This prevents Claude Code from loading every MCP tool at session start, which burns tokens before you've typed a single character.

Ten minutes. Done. You now have a setup that won't cause panic.

Memory: The Compounding Advantage

Memory is the only capability that makes every session better than the last.

Without it: smart tool, starts from zero every time. With it: a teammate that gets better the more you work together.

Two files do the work:

CLAUDE.md — Your Onboarding Document

Lives in your Claude Code folder. This is everything about how you like to work: your roles, preferences, defaults, standing instructions. Think of it as the onboarding document you'd hand a new employee on day one — except this employee reads it before every single session.

Keep it under 300 lines. Shorter is more reliable. The longer it gets, the more likely Claude Code skims rather than absorbs it.

What to put in it:

Your name, company, and role
Your working style preferences
File naming conventions you use
Tools and platforms in your stack
Output format defaults
What you never want done without confirmation

MEMORY.md — The Learning Log

This is what Claude Code has learned about you over time. Every correction, every preference tweak, every feedback loop you give it — it writes to this file so it remembers next session.

These are real text files you can read, edit, and back up. No black box. Full transparency.

The result: after a few weeks of normal use, Claude Code knows your tone of voice, your vocabulary, your punctuation style, your pricing rules, your signature format. Not because you told it once — because it learned and remembered.

Quick setup to bootstrap memory fast: Connect Gmail (or paste in content samples) and say: "Read my last three months of emails / my recent LinkedIn posts. Extract my tone of voice and writing style. Save the patterns you found as writing principles in my memory file."

Claude Code will write a detailed voice profile to your MEMORY.md. Every content task from that point forward will use it automatically.

Pair this with a second brain tool like Obsidian for an additional RAG layer — so you can bring in your research, clippings, and swipe files as reference context on demand.

Connecting Your Stack: MCP Connectors

MCP connectors are how Claude Code reaches into your actual tools.

Before MCPs: copy email → paste into Claude → get output → copy output → paste back into Gmail. Every step manual.

After MCPs: "Draft a follow-up email for the proposal I sent yesterday." Claude Code reads your Gmail, finds the thread, drafts the reply, and saves it as a draft in Gmail — without you opening the app.

The connectors to set up first:

Gmail — Search, read, draft, and send emails. Useful for follow-up automation, inbox classification, and campaign sequences.

Notion — Read pages, databases, and views. Write new entries. Update existing records. The combination of Claude Code + Notion replaces a significant amount of manual CRM work.

Blotato — Content distribution across LinkedIn, X, Threads, and Facebook from a single prompt. Generate a post → Claude Code formats for each platform → schedules it → you approve or reject. No platform-switching required.

Stripe — Useful for payment monitoring, invoice tracking, and basic financial reporting inside Claude Code.

Obsidian — Direct file access to your notes vault for research context, clippings, and knowledge base queries.

Each connector you add multiplies what Claude Code can finish autonomously. The goal is eventually getting to a state where you give an outcome and Claude Code touches every tool needed to deliver it — without you doing the handoffs.

Skills: Packaging Repetitive Workflows

A skill is any repetitive workflow wrapped into a reusable instruction file.

The mental model: think about a task you do repeatedly. Now think about how you'd explain it to an intern on day one — step by step, what to check, what the output looks like, where it goes. Write that out. Wrap it in a skill file. Now that workflow is one command.

How to create one:

Identify the repetitive task
Document it as if briefing an intern
Save the file in your skills folder
Test it, iterate on it, improve it

In Claude Code, type / to see all available skills. Type /follow-up-checker or /newsletter-writer to invoke them.

Why skills matter beyond convenience: they're portable. If you use Claude Code today and want to switch to Claude.ai routines tomorrow, your skill files transfer directly. You don't rewrite anything — the new tool reads the same instruction files.

Skills also make scheduled tasks dramatically simpler. Instead of writing 200-word instructions in each scheduled task, you write one line: /follow-up-checker — and it runs the full workflow.

Live Dashboards With Live Artifacts

Claude Code can build you custom HTML dashboards — charts, tables, KPIs, action buttons — without writing a line of code yourself.

Before this: non-technical founders either hired a developer for a few thousand dollars or used a SaaS tool with limited customization. Now you describe what you want and Claude Code builds it in minutes.

Practical dashboards worth building:

Sales pipeline dashboard — Pipeline by stage, follow-ups due today, deal velocity. Pulls from your Notion CRM or Supabase.

Marketing performance dashboard — Social calendar, SEO position changes, new content opportunities, recent publish log. This is essentially a live view of your AI CMO operations.

Agency operations dashboard — Client deliverables by project, deadlines, status flags. Replace the weekly status meeting with a dashboard everyone can see.

The workflow:

Set up scheduled tasks that gather and aggregate data into your folder or database
Claude Code builds an HTML artifact that reads and visualizes that data
Add action buttons to the dashboard that trigger Claude Code tasks on demand

You don't need to understand any of the code. You need to clearly know what you want to see and what actions you want to take from it. Claude Code handles the rest.

Scheduled Tasks: What "AI Employee" Actually Means

A lot of people say Claude Code is not an AI employee. The argument worth making back: it depends entirely on how you set it up.

A Claude Code setup with the right memory, connectors, skills, and scheduled tasks — running on a timed cadence without you initiating it — is functionally an employee. It works while you sleep. It reports to you. It escalates when it needs a decision.

How to create a scheduled task:

Go to the scheduled tasks tab in the app
Create a new task: name, description, full instructions
Use /skill-name syntax in the instructions to chain skills
Set the schedule (daily, weekly, specific time)
Keep the desktop app open — settings → General → enable "Keep computer awake"

Five scheduled tasks worth building immediately:

1. Follow-up checker (daily, morning) — Checks LinkedIn, email, and your Skool community for overdue follow-ups. Drafts recommended next actions for each. You wake up to a prioritized action list, not a manual review process.

2. Swipe file reviewer (daily, morning) — Reads all your saved clippings from Obsidian or Notion. Summarizes what's waiting for your attention. You decide what to act on in two minutes instead of browsing for twenty.

3. Inbox SDR (daily) — Checks your LinkedIn inbox for unread messages. Classifies intent: business inquiry, resource request, connection noise. Flags the opportunities. Saves you from going through 50 DMs to find the two that matter.

4. Weekly performance report (weekly, Monday morning) — Pulls social performance, SEO position changes, content pipeline status. Saves everything to your reporting database so you can compare week over week. Your team sees it without a meeting.

5. End-of-day shutdown (daily, evening) — Summarizes what changed in your projects and notes that day. Writes a brief for tomorrow morning so you start clean instead of re-orienting.

Other scheduled use cases that work well: new lead alerts, CRM stage updates, competitor content monitoring, weekly content performance pulls, client status updates for service businesses.

Token Management: The Practical Reality

Claude Code with Opus and multiple MCPs active burns tokens fast. This is not a reason to avoid it — it's a reason to be intentional.

The model selection approach that works:

Sonnet by default for most tasks: research, drafting, file operations, MCP calls
Opus only for hard reasoning: complex strategy, multi-step analysis where accuracy matters more than cost
Haiku for light/fast tasks: classification, quick summaries, simple file reads

Enable "Load tools when needed" from day one. This single setting prevents Claude Code from loading all your MCP connections at session start — which can consume a significant chunk of your context window before you've given it a single task.

For browser automation specifically: use Firecrawl or a similar tool for quick webpage scans rather than spinning up a full browser agent. Cheaper, faster, same result for most tasks.

The Full System Stack

Put it all together and this is what a functional AI employee setup looks like:

Sandbox folder + global instructions — safety layer, never touch files outside the sandbox
CLAUDE.md + MEMORY.md — onboarding document + learning log, compounds every session
MCP connectors — Gmail, Notion, Blotato, Stripe, Obsidian plugged in
Skills — repetitive workflows packaged as one-command invocations
Scheduled tasks — skills chained and running on a timed cadence
Live artifact dashboard — data from scheduled tasks visualized, action buttons wired to Claude Code

Most tutorials cover steps one through three. Steps four through six are where the actual leverage lives.

The full course in the video above covers every step in detail with live demos. The Skool community has the CLAUDE.md and MEMORY.md templates if you want a starting point rather than building from scratch.

For more on building an autonomous marketing system with AI agents, read What Is Agentic Marketing: The 2026 Guide.

FAQ

What is the difference between Claude Code and Claude chat?

Claude chat is a conversational AI — you ask questions, it responds, but you execute everything yourself. Claude Code has the same intelligence plus the ability to take actions: reading and writing files, calling APIs, sending emails through Gmail MCP, updating Notion databases, scheduling content through Blotato, and running processes on a timed schedule. The practical difference is that Claude chat gives you advice and Claude Code delivers finished work.

Is Claude Code safe to use on my computer?

Claude Code is safe when set up correctly. The key steps: create a dedicated sandbox folder and point Claude Code only to that folder, add global instructions that require confirmation before any file deletion or override, and enable the memory features so your safety preferences persist. With this setup, Claude Code cannot touch files outside your designated folder, and it will always ask before making irreversible changes inside it.

What is CLAUDE.md and do I need it?

CLAUDE.md is a plain text file in your Claude Code folder that acts as a standing onboarding document. You put your preferences, working style, tools, output format defaults, and any rules you want followed every session. Claude Code reads it at the start of each session. Without it, you repeat yourself constantly. With it, Claude Code already knows your defaults before you type the first character. Keep it under 300 lines for reliability.

What are MCPs and which ones should I connect first?

MCPs (Model Context Protocol) are integrations that give Claude Code access to external tools. Instead of copying and pasting between tools, Claude Code reads from and writes to them directly. Start with the tools you use daily: Gmail for email automation, Notion for task and CRM management, and whichever content distribution tool you use (Blotato works well). Add Stripe if you do subscription billing or need payment monitoring. Each connector multiplies what Claude Code can complete without you doing the handoffs.

How do I set up a scheduled task in Claude Code?

Go to the Scheduled Tasks tab in the Claude Code desktop app. Create a new task, give it a name and description, and write the full instructions including any skill invocations using the /skill-name syntax. Set your schedule. Keep the desktop app open and enable "Keep computer awake" in Settings → General — scheduled tasks only run when the app is active. Start with one simple task (like a morning follow-up checker) before building complex multi-step sequences.

How much does Claude Code cost to run daily?

Cost depends on which model you use and how many MCP tools are active. Using Sonnet for most tasks with selective Opus for hard reasoning keeps costs manageable for typical founder workflows. The highest cost scenarios are Opus with multiple MCP connections running simultaneously — those can drain your token budget fast. Enable "Load tools when needed" to prevent MCPs from loading at session start, and scope your scheduled tasks precisely to avoid burning tokens on broad searches.

Can Claude Code replace a full marketing team?

Claude Code with the right setup handles a significant portion of repetitive marketing execution: content drafting, lead follow-up, social scheduling, performance reporting, and inbox classification. What it does not replace: strategic judgment, relationship building, and creative direction. The better frame: it eliminates the execution bottleneck so a lean team can operate at the output level of a much larger one. AI Topia runs lean marketing infrastructure for founders using exactly this kind of setup — see the AI CMO platform for what that looks like in practice.