Claude Code,
the complete guide.

What is
Claude Code?

In concrete terms, Claude Code is the terminal version of Claude. You open your terminal, type claude, and you land in an interactive session where the AI can do what a human assistant would do for you: open a file, edit it, run a test, validate, start over. You can tell it "Add a signup form to my page, hook it up to Resend, test that it works," and it chains the steps on its own.

Claude Code vs Claude.ai (the chat)

Claude.ai is the conversation — you ask questions, Claude replies with text. Claude Code is the execution — you give it an intent, Claude reads your project, handles files, runs shell commands, and delivers a result. It's the same AI underneath (Anthropic's Sonnet and Opus models), but the interface changes everything. On Claude.ai you copy-paste code by hand; with Claude Code you never touch a programming keyboard again.

Claude Code vs Cursor, Aider, Codex, and GitHub Copilot

All these tools use an AI to write code. The difference plays out on three axes: where it lives (terminal vs editor), who acts (you vs the agent), and the model underneath (Claude vs GPT vs others). Cursor is a full code editor with a chat; Aider is an open-source CLI that edits files; GitHub Copilot is an autocomplete built into VS Code; Codex is OpenAI's CLI agent. Claude Code combines Aider's simplicity, Codex's agency, and Claude's reasoning quality.

You'll find the detailed comparison table a bit further down this page. If you want the short version: go with Cursor if you want to replace VS Code, go with Claude Code if you want to hand off whole tasks to an autonomous agent.

Who is it for?

Officially, it's a tool for developers. In practice, I use Claude Code every day without being one. If you can open a terminal and copy-paste a command, you can use it. I know consultants who use it to generate client reports, marketers who build their landing pages with no dev, lawyers who automate contract review. The barrier to entry is lower than people think.

Why Claude Code
changes everything.

For years, AI for code meant autocomplete: you type a function, the AI proposes what comes next, you accept. Useful, but you stay the pilot. Claude Code breaks that pattern. You describe what you want to get, and it picks the files to touch, the commands to run, the tests to write, all by itself. You go back to being the decision-maker, not the operator. That's what I call an agentic workflow.

What it changes for me, concretely

I built this site without writing a line of HTML myself. I built a sales tool I use every day. I have a Hermes Agent that automatically sends me Telegram messages when my builds pass. I launched a Gmail agent that filters my emails. All of it was built by Claude Code, and it's maintained by Claude Code.

Claude Code's real strengths

• It reads your whole codebase. With 1 million tokens of context on Opus 4.7, it can hold an entire project in its head before acting.
• It spins up subagents in parallel. You ask it to analyze five documents, it launches five at once, then synthesizes. Five times faster than a sequential agent.
• It integrates through MCP. You can plug in Sanity, Stripe, Telegram, your Postgres database, your Meta Ads account — all through the open MCP standard.
• You can control it finely. With hooks and the CLAUDE.md, you can stop Claude from touching certain files, impose rules on it, and trigger actions at specific moments.

What Claude Code does less well (the honest part)

I'm not going to sell you Claude Code as a miracle cure. Here's what it does less well in my experience: it sometimes invents library functions that don't exist (the context7 plugin fixes 80% of the problem by feeding it up-to-date docs). It over-writes code by adding pointless comments (the code-simplifier skill cleans that up). It sometimes loses the thread on very long sessions (past 500,000 tokens, I'd rather open a new session). And it doesn't replace a real senior dev for critical architecture decisions — for that, I ask a human.

My recurring use cases

• I ask it to write my newsletters in my own voice (the copywriting skill)
• I ask it to build web pages like this one (the landing-page-generator skill)
• I have it audit the security of my projects (the security-review skill)
• I have it do my research and curation (the deep-research skill)
• I have it write and send cold emails (the cold-email skill)
• I have it process my podcast audio and generate the transcripts
• I have it watch my Vercel builds and alert me when there's a regression

All of this comes from skills. We'll cover how to install them in step 04 below.

Claude Code
vs the alternatives.

The market for AI coding assistants exploded in 2024-2026. Every six months, a new tool shows up, claims to be the best, and three months later another one takes its place. Here's my comparison table, refreshed in May 2026, based on what I've actually tested (not copy-paste from marketing pages).

My honest take after testing them all

Claude Code stays my main tool for the heavy stuff — refactoring, page generation, audits, agents. I use Cursor on the side for when I need to see the diff live in an editor. Aider, I tested it for two weeks then dropped it: too manual for my use. Codex CLI does a good job, but the shorter context limits me on big projects. Continue, I like it for short local tasks on an open-source model when I want zero API cost.

To dig into a real-world case, read my hands-on comparison Limova vs Claude Code (€840 a year apart): I lay out why I ditched a commercial tool for Claude Code, numbers and all.

Install Ghostty.
The terminal I use.

Ghostty is the terminal I use on Mac. It's free, fast, good-looking, and it supports emoji. You don't have to use Ghostty — the default macOS terminal works too — but I highly recommend it.

echo "Salut Ghostty"

Install Claude Code.

Open your terminal (Ghostty or Terminal), and paste this command:

curl -fsSL https://claude.ai/install.sh | bash

Here's what you should see when it works

Once the install is done and you've restarted the terminal, type claude --version then claude. You should see exactly this:

$ claude --version
claude 1.2.3

$ claude
Welcome to Claude Code!
Ready. Type a message.
>

If you see that, you've nailed it. If you see command not found, restart your terminal (quit it fully with Cmd+Q, reopen it) and try again. In most cases, that's enough.

claude

mkdir ~/Desktop/test-claude
cd ~/Desktop/test-claude
claude

Drop in your CLAUDE.md.

The CLAUDE.md is a small file Claude reads at every session. You put your rules, your style, and your tech stack in it. It saves you from repeating "use Next.js" or "reply in English" every single time.

I'm sharing my anonymized CLAUDE.md as a starting point. You open it, adapt it to your style and your stack, and that's it.

mkdir -p ~/.claude
cp ~/Downloads/CLAUDE.md ~/.claude/CLAUDE.md

open ~/.claude/CLAUDE.md

Install my 26 skills.

Skills (in plain words: ready-made "professions" for Claude) are what turn Claude Code from a plain assistant into a specialist. A skill = a folder that holds instructions for a specific job: cold email, SEO, landing page, security audit, and so on.

When you ask Claude "write me a B2B cold email", it automatically loads the cold-email skill and follows the rules defined inside it. You save a ton of time.

mkdir -p ~/.claude/skills
cp -R ~/Downloads/jeremy-claude-pack/skills/* ~/.claude/skills/

claude

Which skill(s) for you?

37 skills at once is a lot. You don't need to use them all — they load automatically based on what you ask. But to help you figure out where to begin, here's my breakdown by profile.

My 8 favorites if you're starting out.

I have 37 skills installed on my machine in total. Here's the truth: 90% of people use 8 at most. Here are the 8 I recommend when you're just starting — they cover writing, marketing, and research, with no dev skills required.

How I use it
every day.

Morning — The watch

I launch claude in my research folder and ask it: "Use the deep-research skill to summarize the 10 most important AI news items this week". While it works, I grab my coffee. 10 minutes later, I've got a clear brief.

Midday — Building

When I need to write a landing page, an email, a page — I launch Claude Code. It automatically loads the right skill (copywriting, landing-page-generator, cold-email). I give it the brief in 3 sentences, and it produces a first draft.

Evening — Cleaning up the code

Before I commit, I run security-review on the diff (the list of changes made since the last save). In 30 seconds it tells me whether I left an API key (the "access card" that lets your code talk to a paid external service) lying around, or whether a SQL query is vulnerable.

The official plugins worth knowing.

Anthropic offers more than 140 official plugins on its marketplace (the official "app store" — claude.com/plugins). A plugin is a ready-made pack of skills, commands, and agents that installs in one line. Nothing to configure, nothing to copy by hand. Once installed, it's available in all your projects.

After several weeks of testing, here are the 6 I use. Each one adds real value for a non-dev running concrete projects.

claude plugin install superpowers
claude plugin install context7
claude plugin install claude-md-management
claude plugin install frontend-design
claude plugin install code-review
claude plugin install code-simplifier

What they bring

To understand the difference between a plugin, a marketplace, a skill, and an agent in plain words, check out entry #08 in the glossary.

Quick troubleshooting.

The claude command isn't recognized

Restart your terminal. If that changes nothing, manually add the Claude path to your ~/.zshrc:

echo 'export PATH="$HOME/.claude/bin:$PATH"' >> ~/.zshrc
source ~/.zshrc

Claude isn't loading my skills

Check that the folders are actually there:

ls ~/.claude/skills/

You should see the list of skills. If the folder is empty, rerun the install from step 04.

Claude isn't reading my CLAUDE.md

Check the location:

cat ~/.claude/CLAUDE.md

If it prints content, you're good. Claude reads it automatically at every session.

I'm really stuck

Reply to my newsletter — your email comes straight to me, and I'll unblock you. I read everything.

The right workflow
for your job.

There's no single "right way" to use Claude Code. The best way to use it depends on what you do day to day. Here are the four profiles I run into most often, with a recommended setup for each.

The essential articles
on Claude Code.

This page is a pillar guide: it covers everything, but not in depth. For each specific topic, I've written a dedicated article. Here's the selection to read first, ranked by relevance for a beginner.

The vocabulary
of Claude Code.

Before diving deeper, here are the terms that come up all the time. Short definitions, with a link to the full entry in my AI glossary.

• Claude Code Anthropic's CLI agent that runs tasks in your terminal. See the entry →
• CLAUDE.md A context file at a project's root or in ~/.claude/, read by Claude at every session. See the entry →
• Skill A folder of instructions that turns Claude into a specialist for a job (cold-email, copywriting, security-review). See the entry →
• Subagent A mini-Claude the main agent launches for a parallel task. See the entry →
• Hook A script that fires automatically on a Claude Code event (before a tool, after a response, on stop). See the entry →
• Slash command A manual shortcut starting with /, e.g. /brainstorm, to launch a specific workflow. See the entry →
• MCP Model Context Protocol, the open standard for plugging Claude into external services (DB, API, tools). See the entry →
• Agentic workflow A pattern where the AI chains several steps autonomously, using its tools with no human involvement. See the entry →
• Context engineering The art of building the right context (files, examples, rules) so the AI produces the right answer. See the entry →
• Tool use A model's ability to call external tools (bash, file, web). The heart of agency. See the entry →
• Context window The number of tokens the model can hold in its head at once. Sonnet 4.x: 200k, Opus 4.7: 1 M. See the entry →
• Hallucination When the AI invents a function or a fact that doesn't exist. Enemy number one of agents. See the entry →
• API key An access card that lets a program talk to a paid third-party service. Keep it secret. See the entry →
• .env file A local file that holds your secrets (API keys, DB passwords). Never to be committed to Git. See the entry →
• GitHub Actions GitHub's built-in CI/CD system. You can run Claude Code there in non-interactive mode on events. See the entry →

You've got the full list on the AI glossary page — more than 60 short entries, each with a definition, an example, and a link to the use in practice.

Everything people ask me
about Claude Code.

Twenty questions I get by email or DM every week. Direct answers, in first person, with no marketing pose.

Ressources
utiles.

Quelques liens à mettre en favoris pour aller plus loin sur Claude Code. J'ai mis en priorité les sources officielles et celles que je consulte vraiment.

Et si tu veux que je te tienne au courant des sorties Claude Code, des plugins qui valent le coup, et de mes propres trouvailles — la newsletter AI Playbook tombe tous les vendredis matin à 9 h. Lien d'inscription dans le footer.