From Chat to Code: Making the Switch to Claude Code
🎙️ Listen to this article
From Chat to Code: Making the Switch to Claude Code
What this guide covers: The mental model shift from Claude.ai to Claude Code — what changes, what stays the same, and how to start working with folder-based AI agents.
Estimated time: 10 minutes to read, same day to get working
Prerequisites: You've used Claude.ai with Projects. You have Claude Code installed (see: How to Install and Set Up Claude Code on Your Mac). You don't need to be technical.
The Shift Nobody Explains
This guide comes from a training session I ran this week. One of the team had been using Claude.ai with Projects for weeks and getting real value — then hit the wall everyone hits. "I have no idea what Claude Code is. I thought it existed within Claude Pro." Twenty-nine minutes later, they had an agent pulling Google Ads data and building reports. Here's the path we walked.
You've been using Claude.ai. You've set up Projects, uploaded files, written instructions. You're getting real value from it.
Then someone mentions Claude Code and you realise it's a completely different thing. Not an upgrade to what you've been doing — a different way of working entirely.
Here's the distinction that matters:
Claude.ai is a consultant you call for advice. You describe problems. It gives solutions. You go implement them yourself.
Claude Code is that same consultant sitting at your desk, with access to your files, your folders, your tools. It doesn't just advise — it works. It reads data, writes scripts, builds reports, connects to external platforms. You're not implementing its suggestions. You're directing it while it implements.
That's the leap. And it's bigger than the jump to ChatGPT.
The Architecture: Folders, Not Chat Windows
In Claude.ai, your world is Projects and Chats. Instructions at the top, files attached, conversations below.
In Claude Code, your world is folders on your computer. Everything lives in a folder. The folder is the project.
Here's what a Claude Code project looks like on your machine:
shark-ninja-reporting/
├── CLAUDE.md ← The agent's briefing
├── .claude/
│ └── agents/ ← Specialist agent definitions
├── skills/
│ ├── copywriting.md ← How to write copy
│ ├── webprofits-brand.md← Brand guidelines
│ └── notion-connector.md← How to pull from Notion
└── data/
├── ad-performance.csv ← Raw data you drop in
└── campaign-brief.pdfThe key folders: CLAUDE.md at the root is the agent's brain. A skills/ folder holds reusable capability files. A data/ folder is where you drop your CSVs and spreadsheets.
When you launch Claude Code in that folder, it reads the CLAUDE.md file first. That's the agent's brain.
Everything else in the folder is context. Data files, skills, agent definitions. Claude can see and use all of it.
[Screenshot: A real Claude Code project folder open in Finder, showing CLAUDE.md, skills/, and data/ directories]
CLAUDE.md: The One File That Matters
The CLAUDE.md file is the single most important thing in any Claude Code project. It's the prompt that defines the agent.
Think of it as the difference between hiring someone and briefing them. Claude.ai is the interview — you explain what you need each time. CLAUDE.md is the job description, the onboarding document, and the operating manual rolled into one.
A real CLAUDE.md might say:
# SharkNinja Reporting Agent
You are a reporting agent for SharkNinja paid media campaigns.
## What You Do
- Pull data from Google Ads via the MCP connection
- Analyse ad-level investment across product categories
- Build visual reports with Webprofits branding
- Deploy reports as HTML files
## Data Sources
- Google Ads MCP (live connection)
- CSV exports in /data folder
- Notion database for campaign briefs
## Standards
- Use Australian English
- Follow Webprofits brand guidelines (see skills/webprofits-brand.md)
- Always verify data before presentingIn plain terms, you're telling the agent: you're a reporting agent, here's where your data comes from, here are your standards.
[Screenshot: A real CLAUDE.md file open in a text editor, showing agent instructions]
Go to Claude.ai and say: "I want to build a Claude Code project that does X. Make me the CLAUDE.md file and any skills I'll need." It generates the files. You put them in a folder. You launch Claude Code in that folder. You're working.
That's the workflow. Brief Claude to build the briefing. Then let the agent run.
Skills: Reusable Capabilities
Skills are markdown files that teach the agent how to do specific things. They're like giving your new hire a procedures manual.
A copywriting skill defines your tone of voice and quality checklist. A brand skill holds colour codes and messaging frameworks. A connector skill handles authentication with Notion or Google Ads.
The agent reads these when it needs them. You don't have to reference them in every conversation — they're just there, part of the agent's toolkit.
What makes skills powerful: They're reusable across projects. The same copywriting skill works for SharkNinja reports and Ripple reports. Write it once, drop it into any project folder that needs it.
Skills handle what the agent knows. MCPs handle what the agent can connect to.
MCPs: Connecting to Everything
MCP stands for Model Context Protocol. In practical terms, it means Claude Code can connect to external platforms — not just read files in a folder.
With MCPs configured, your agent can:
- Pull Google Ads data directly from the API — live campaign performance, spend by ad group, conversion data
- Read and write to Notion — pull from databases, update pages, create new entries
- Access Meta Ads data — campaign performance, creative analysis (this one's still getting set up — Jay's configuring it and I'm waiting to hear back on the authentication. Not everything's plug-and-play on day one.)
- Connect to Slack, Supabase, and more — the list keeps growing
Here's what this looks like in practice. During the session, I typed this prompt live:
"Use the Google Ads MCP to pull a three-month performance report for Ripple and build it into an analysis with Webprofits branding."
Claude Code connected to Google Ads, found the Ripple account ID, pulled the data, and started building the report — all while the conversation continued. No export. No spreadsheet. Live data, straight from the platform, into a branded analysis.
That changes the speed of everything.
[Screenshot: Claude Code pulling live Google Ads data via MCP during the session]
[Screenshot: The finished Ripple Google Ad Performance report output]
Getting Started: Desktop App in Five Steps
You don't need Terminal to use Claude Code. The Claude Desktop App has a Code mode built in.
Here's how to get started:
- Open the Claude Desktop App — the one you already use for Claude.ai
- Switch from Chat to Code — look for the Chat/Code toggle in the interface
- Select your project folder — click the folder icon at the bottom and navigate to the folder with your
CLAUDE.mdfile - Choose your model — switch from Sonnet to Opus if you need the most capable model
- Start working — type your first prompt
[Screenshot: Claude Desktop App showing the Chat/Code toggle and folder selection]
That's it. The app reads your CLAUDE.md, loads the skills, connects to any configured MCPs, and you're talking to your agent.
When you trust the folder (a one-time prompt you'll see the first time you open a project), you're telling Claude Code: yes, you can read and change files in this folder. Once you do this, the agent has full access to everything in the project.
Drop a CSV, Get a Report
One of the most immediately useful things about Claude Code: you can drop data files into the project folder and point the agent at them.
Export a CSV from Meta Ads. Download a spreadsheet of campaign performance. Save a PDF of a client brief. Put them in the /data folder.
Then tell Claude what to do with it. In the same session, I dropped a SharkNinja ad-level investment CSV into the data folder and prompted:
"Find the SharkNinja ad-level investment data year to date. Analyse which products and creative elements we're investing most and least in. Build a visual report with Webprofits branding."
Claude read the CSV, wrote a script to parse and analyse it, and started generating the report — while I continued the conversation. The analysis was structured around the actual data, not generic templates.
They'd tried something similar before in Claude.ai — exported Meta data and asked for an analysis. The data got confused because they asked too many things at once and the naming conventions tripped it up. Claude Code handles this better because of the folder structure and the ability to write proper parsing scripts, but the lesson stands: clean data in, clean analysis out.
[Screenshot: Claude Code agent mid-analysis, writing a script to parse the SharkNinja CSV]
The key insight: You don't need live MCP connections to get value. A CSV export and a clear prompt will get you 80% of the way there on day one.
The Workflow: Think What, Not How
The mental shift that makes Claude Code click:
One of the team summed it up mid-session: "Don't think of how. Think of what you want to achieve."
You don't need to know how to write a script that analyses CSV data. You don't need to know how to build an HTML report with branded styling. You don't need to know how MCP authentication works.
You need to know what you want the output to look like. You need to be specific about the outcome.
The process:
- Define what you want — a report, a dashboard, an analysis, a tool
- Brief Claude.ai — ask it to generate your
CLAUDE.mdfile and skills - Set up the folder — put the generated files in a new directory, add any data
- Launch Claude Code — open the folder in the Desktop App or Terminal
- Direct the work — tell the agent what to build, approve its steps, refine the output
You're not coding. You're not configuring. You're directing an agent that has the skills and connections to do the work.
Every Project After the First Gets Faster
The skills you build compound. The CLAUDE.md files you write get sharper. The MCPs you configure work across every project. Each new project starts faster than the last because the infrastructure carries over.
This is where it stops being a tool and starts being a way of working. Not "I use Claude sometimes" but "I have an agent that builds my SharkNinja reports while I focus on the strategic recommendations inside them."
They went from "I have no idea what Claude Code is" to building a SharkNinja ad analysis in under thirty minutes. The infrastructure carries over. The second project takes half the time.
Common Gotchas
"I opened Claude Code but it's not reading my instructions."
Your CLAUDE.md file must be in the project root — the top level of the folder you selected. If it's nested inside a subfolder, Claude Code won't find it. Check that the file is named exactly CLAUDE.md (capital letters, no spaces).
"The analysis pulled wrong data from my CSV."
This usually means the naming conventions in the CSV are ambiguous. One of the team hit this — they asked Claude.ai to analyse Meta Ads data and it picked up the wrong terms from the naming conventions. Be specific about which column or metric to analyse. Clean data in, clean analysis out.
"I don't see the Code toggle in the Desktop App."
Make sure you're running the latest version of the Claude Desktop App. The Code mode was added in a recent update. Check for updates in the app, or download the latest version from claude.ai.
Cheat Sheet
| Concept | What It Is | Where It Lives |
|---|---|---|
| CLAUDE.md | The agent's briefing — who it is, what it does | Project folder root |
| Skills | Reusable capability files | /skills/ or /.claude/ folder |
| MCPs | Connections to external platforms | Configured in Claude settings |
| Data files | CSVs, spreadsheets, documents for analysis | /data/ folder |
| Desktop App | Easiest way to launch Claude Code | Toggle Chat → Code in the app |
| Terminal | Direct way to launch Claude Code | cd folder && claude |
Next Steps
Pick one thing you do manually every week. Export the data. Write one paragraph describing what you want the output to look like. Ask Claude.ai to generate the CLAUDE.md. Put it in a folder. Open Claude Code.
You'll have a working first version before your next meeting.
