A Practical Guide to Claude Code (Even for Non-Developers)

In recent editions, I told you about Claude (the chat) and Claude Cowork. I had touched on Claude Code, but only briefly. I thought it would be worth dedicating a full edition to it, and here it is.

Especially for people who can't code and might think “this isn't for me”. On the contrary, there's a lot of good stuff to get out of it, even when you're non-technical.

Andrej Karpathy (ex-OpenAI) summed it up well, back in 2023

“English is the new code.”

This article, like the Cowork one, also applies to the other ecosystems, which increasingly resemble each other. Codex is just as good (or even better right now) as Claude Code.

OpenAI recently shared that more than 90% of its employees use Codex (Anthropic's alternative). And obviously it's not just developers. The goal of this edition: understanding how to install it, the best practices, and 3 concrete use cases anyone can set up as early as this week.

Quick refresher: Claude, Cowork, Code... what are we talking about?

For those discovering the ecosystem, here's the difference between the Claude tools:

In 2 sentences:

• Claude (the chat) is a chatbot, like ChatGPT
• Claude Cowork is a mix between Google Drive and Claude, working on your local files
• Claude Code is an agent capable of creating, editing and running code for you

Claude Code was initially reserved for developers. But code has become so accessible to everyone that the use cases for non-technical people have become very interesting.

Here we'll focus on the desktop app (or Claude Code Web), which is the simplest route. For those who want to go further, you can switch to an IDE like VS Code.

The 5-minute setup

Getting started is quick:

1. Download the Claude Desktop app (requires at least a Pro account at $20/month)
2. Open the app, click the “Code” tab
3. Select a folder on your computer (create a dedicated one: “claude-test”)
4. Connect your GitHub account (free) if you want to host your sites online

That's it. You can start prompting.

The 4 permission modes to know

When you use Claude, you have 4 possible modes, selectable in the bottom left.

For example, I selected the “Démo” folder on my computer in planning mode

A quick tip before prompting: take 2 minutes to structure your folder. At the root, create a CLAUDE.md file (the memory Claude reads at each session with the project context). Subfolders can be useful, like /input for the files you give it to analyze, or /output for the generated deliverables.

You can also add a few context files in .md format (design.md, voice.md, about-me.md). The better organized the folder, the more relevant Claude is. (I already talked about this in the Claude Cowork edition)

You can control these 4 modes to gradually shift to mode 4, where you hand over control to Claude. Unless you're a developer, chances are you won't check the generated code.

1️⃣ Ask on every action (default): safe but slow. You approve everything.

2️⃣ Auto-accept edits : Claude accepts file changes automatically.

3️⃣ Plan mode : Claude plans out the agenda to follow and presents it to you before execution. Very useful for big projects, especially at the start.

4️⃣ Bypass permissions : Claude reads, edits and runs everything without asking, including destructive actions like deleting files.

⚠️ If you enable mode 4, work in an isolated folder to limit the damage in case of error.

My advice to start: alternate between plan mode to frame what you want, then auto-accept to execute. It's the right balance of speed and control. To switch between modes, you can use shift+tab in the terminal.

Before moving on to 3 examples of basic Claude Code uses, here are the points to keep in mind about the tool:

3 concrete use cases to get started (without coding)

I'm sharing 3 simple cases that resonate with most workers. We come across them very often in the companies we support at Tandem.

1️⃣ Analyzing and manipulating documents

It's probably the most underrated case. Excel, PDF, CSV, reports... these time-consuming tasks can largely be delegated.

A few concrete examples:

• Analyzing an Excel sheet to generate an interactive HTML dashboard
• Automatically filling in PDFs (administrative forms, contracts)
• Cross-referencing several files to produce a summary report
• Cleaning up messy tabular data

The real gain is when Claude has all your company's context: your templates, your processes, your guidelines. It no longer just acts on an isolated file, it works like an assistant that knows your organization.

Here's an example with the analysis, categorization and dashboard generation of an expense PDF

2️⃣ Creating slides and reports

Another classic case: producing clean, understated materials that respect a brand's visual guidelines.

You describe the content (quarterly results, client brief, strategic plan), you give the brand constraints (colors, fonts, logo), and Claude generates the slides directly.

You can also create a design.md file in your folder with your guidelines (hex colors, typography, tone). Claude Code will refer to it every time it creates something.

3️⃣ Creating a website or a landing page

The most visible case. You can literally create a functional site in a few prompts.

You can start with “Plan Mode” (in blue), iterate conversationally in the chat (in yellow), then view the preview directly in the app (in green) via the display panel (in red).

The approaches that work well, I find:

• Describe the end result (not the technical steps): “I want a clean landing page for my consulting firm, with a main section, 4 services, testimonials, a footer”, rather than explaining to Claude how to do it.
• Share a screenshot of a site that inspires you and ask: “Reproduce this style but for my business”
• Point Claude to files you've prepared (about-me.md, voice.md, services.md), exactly like with Claude Cowork (I talked about this in that edition)Claude Cowork: the tool you're not using (and that changes everything)Louis Graffeuil·Apr 10Read full story

But as you can see, the design is too generic. To fix that, I recommend using Claude Design (recently released by Anthropic) which brings a real design touch, and can even respect your design (more on that later). You generate the visual on Claude Design, then bring the project back into Claude Code to iterate on the logic.

For example, I generated and iterated on Claude Design (where I can also edit, comment.. easily). Then I click on “Share” to share the result with Claude Code:

At this point, I have 2 options:

1. share the content with Claude Code via the app (or Codex, Antigravity)
2. share the result with Claude Code Web

I recommend staying on the app and you'll be able to retrieve the entire content to make use of it.

You'll then need to go through GitHub to host your site (or Vercel, Netlify) to get exactly the same rendering (otherwise the font might not load, nor the images, etc)

That's typically the winning combo: Claude Design for the visual x Claude Code for the logic and the iterations.

Practical tips to go faster

Here's a series of techniques that make a real difference when you use Claude Code regularly.

Have Claude ask questions before it executes

Probably the most impactful piece of advice. Rather than diving in headfirst, ask Claude to question you to fully understand what you want.

A simple prompt: “Keep asking me questions until you're 95% confident about what you need to do. Use your ask_user_question tool.”. If you're in “Plan Mode”, there's a good chance Claude will ask you questions directly.

The result: fewer back-and-forths, fewer errors, time saved. You clarify what's missing rather than heading in the wrong direction with rough assumptions.

Create a CLAUDE.md file, your project's memory

The CLAUDE.md is a markdown file that Claude Code reads first at each session. It stores what it has learned about your project there: structure, design choices, preferences, conventions.

So it's important to have a Claude.md file at the start of a session if it's useful, and to keep the document updated. You can make it a habit at the end of each session: “Create a CLAUDE.md file with everything you've learned about this project: folder structure, design choices, my preferences, existing sections.”

Note: this file is loaded into every conversation and consumes context. Stay under 200 lines max. To go further, you can reference sub-files in the CLAUDE.md (style-guide.md, business-context.md, etc) that Claude will only fetch when needed.

Here, for example, in the file, I point to other documentation for a demo project:

Watch your token consumption

The /context command shows you exactly what's consuming your context window (system prompt, files, MCP servers). Very useful when you feel your session is getting slow.

When you reach 60% of context, use /compact to compress the conversation. You keep the history but free up space.

A few commands to know

• /init: scans an existing project and creates a CLAUDE.md automatically
• /rewind: go back in the conversation without redoing everything. Below each prompt, you can go back or fork a conversation to move forward on 2 tracks in parallel, for example.

• /usage: see your token consumption

On the app, use voice dictation to share your ideas and save time.

Work in steps

One task per prompt, not 12. Rather than “Create the whole site with authentication, blog, dashboard and payment”, break it up: the homepage first, validate, then the contact page, etc.

The same goes for fixes: take a screenshot of the visual bug and ask “Fix this”. Faster than a long paragraph of explanation.

The friction points to know (and the good news)

Claude Code is powerful but not magic. There are still important limitations to know:

1. It burns tokens. A Code session consumes the equivalent of 20+ classic chats. On the Pro plan at 20€/month, you'll quickly hit the limit. You can quickly move up to a plan at $100/month.
2. Claude Code can also hallucinate. It can tell you a task is done when it isn't, or delete code while saying “I solved the problem”. Always check the result yourself: open the site, click the buttons, test like a user.
3. It can create bugs while fixing bugs. When you see the same error twice, stop. Ask: “Stop, explain to me what's happening and suggest 2 different approaches.”
4. The design is mediocre by default. We talked about it, hence the value of Claude Design + well-stocked brand guideline files.

Very recently, Anthropic significantly increased the token limits (official article here). Even on the Pro plan at $20, you now have more room to experiment with peace of mind (the limits are doubled + the limit reduction during peak hours on Claude Code is removed for Pro and Max accounts)

If you want to dig deeper, let me know in the comments. I can do an edition dedicated to techniques for saving tokens, to using design.md, or to Claude Code plugins (skills, sub-agents, etc) that I didn't cover.