The 8 AI words
to grasp
before the rest.

The analogy that clicks

Picture a personalized gym badge. You scan it on the way in, the gym knows it's you, and it charges your membership every time you visit. If you lose it or lend it out, you're the one who gets the bill. An API key is exactly that: a personalized badge for Claude. I keep mine in a password manager, never out in the open.

What it actually does for you

• Hook Claude up to Zapier, Make or n8n to automate writing emails, meeting summaries or product sheets.
• Connect Claude to an editor like Cursor, which helps you write documents, scripts or websites without starting from scratch.
• Run an automation that reads your Notion, your Google Sheet or your CRM and generates content on a loop while you sleep.

How to use it — step by step

1. Create an account at console.anthropic.com (it's separate from claude.ai, even if you already have a Pro subscription).
2. Add a bank card in the Billing tab. No mandatory monthly subscription, you only pay for what you use.
3. Go to Settings → API Keys, then click Create Key.
4. Give it a meaningful name (example: zapier-prod, n8n-test). The rule I stick to: one key per tool, never a single one for everything.
5. Copy the key immediately and paste it into a password manager. If you close the window, it's gone: you'll have to create a new one.
6. Paste it into the target tool (the API Key field in Zapier, Make, n8n, Cursor, etc.).
7. Go back to Billing → Limits and set a monthly spending limit. That's your safety net.

What it costs (April 2026, per million "tokens" — see entry #03): Haiku 4.5 at $1 input and $5 output, Sonnet 4.6 at $3 / $15, Opus 4.7 at $5 / $25. For most everyday automations, Haiku or Sonnet is plenty.

The analogy that clicks

It's the key ring tucked away in the hallway drawer. When someone visits your house (your code), they see the furniture but not the keys. The key ring stays hidden, and when the app needs to open a door (call Claude, send an email), it quietly reaches inside for it. The worst mistake would be sticking the keys on the front door — the equivalent of posting your code online with the passwords inside.

What it actually does for you

• Separate your secrets from your code. You can share your project with a technical person without handing over your paid access.
• Swap out a key (after a rotation or a leak) without touching a single line of code.
• Run several environments side by side: a .env.development with test keys, a .env.production with the real ones.

How to use it — step by step

1. At the root of your project, create a file named exactly .env (mind the dot at the start).
2. Add one line per secret, for example: ANTHROPIC_API_KEY=sk-ant-your-real-code. No spaces around the =, no quotes unless the value contains a space.
3. Before saving anything to GitHub, check that the .gitignore file does contain the line .env. That's what keeps your notebook from leaving with the rest.
4. Create a .env.example file next to it with the same names but without the real values (example: ANTHROPIC_API_KEY=your-key-here). That one can go to GitHub, it serves as a template for you or a collaborator.
5. Restart your app so it loads the new variables.

Resources that go further if one day you're managing several keys with a team: 1Password CLI (secrets never touch the disk, unlocked with Touch ID) or Doppler (automatic key rotation, handy from three people up). For a solo project, a plain .env is more than enough.

The analogy that clicks

Think of an Uber. You don't pay per trip, you pay by distance traveled. The meter ticks up every kilometer. Tokens work the same way: every word you send Claude turns the meter, and every word it sends back does too. A short email = a few cents. A fifty-page contract = several dollars. Me, I always estimate before kicking off a loop, to dodge the nasty surprise the next morning.

What it actually does for you

• Understand why your bill swings around: a hundred short emails cost far less than a single fifty-page PDF.
• Estimate the cost of an automation before you launch it. A worked example: 1,000 product sheets of 300 words processed with Sonnet 4.6 comes to roughly 400,000 input tokens, or about $1.20.
• Pick the right model for the task. Haiku 4.5 at $1 per million tokens is three times cheaper than Sonnet 4.6 at $3, and it's plenty for simple jobs (sorting, short summaries, extraction).
• Avoid the nasty surprises. A hundred-page PDF contract is roughly 125,000 input tokens. If you run it on a loop inside a scenario, the bill climbs fast.

How to estimate — step by step

1. Count the number of words in the document you want to send (right-click → Properties in Word, or at the bottom of Google Docs).
2. Multiply by 1.3 if it's in English, 1.5 if it's in French. That gives you the input token count. Good to know: French costs about 14% more than English for the same volume, because the tokenization is less optimized.
3. Estimate the length of the expected answer and multiply that by 1.5 too. Those are your output tokens.
4. Apply the model's pricing. Simple rule to memorize: output costs five times more than input. So a short prompt asking for a long summary costs more than a long document with a short question at the end.
5. If you want the exact figure, run your text through claude-tokenizer.vercel.app.

A few visual landmarks: 1,000 tokens ≈ 500 to 650 words in French, roughly one A4 page. The context window (what Claude can read in one go) is 1 million tokens on Sonnet 4.6 and Opus 4.6/4.7, so about 1,500 pages. Easily enough to feed it a whole book.

The analogy that clicks

Imagine texting back and forth with your computer. No buttons, no icons, no menus. You type a sentence, it replies. That's it. The first time I opened it, that black window scared me. In reality, it's just a chat box. Quick vocabulary: CLI = the principle (command-line interface), Terminal = the app that opens the window, Shell = the engine running inside it (bash or zsh on Mac, PowerShell on Windows). You don't need to memorize this, I'm just leaving it here in case you come across it.

What it actually does for you

• Install a tool a tutorial tells you to copy and paste (Claude Code, Node.js, Homebrew).
• Launch Claude Code in a folder on your computer by simply typing claude.
• Run a little program a contractor sent you, without going through a technical team.

How to use it — step by step

1. On Mac: press Cmd+Space, type "Terminal," confirm with Enter. If you want something prettier later: Ghostty (free) or Warp (paid, with built-in AI).
2. On Windows: open "Windows Terminal" from the Start menu, it's pre-installed. By default it launches PowerShell, which will do the job.
3. Copy the command the tutorial gives you, paste it into the window, press Enter, wait for the machine to respond.
4. If you see a red error, don't panic: copy it, paste it into Claude, ask it to translate it and give you the fix.

The 3 commands that cover 90% of cases: cd folder-name to enter a folder, ls to see what's inside, claude to launch Claude Code in it.

The analogy that clicks

It's the USB-C of artificial intelligence. Remember before: every device had its own cable. One for the iPhone, another for the Kindle, a third for the headphones, and you rummaged through a drawer full of wires to find the right one. USB-C imposed a single universal plug. You plug in, it works. MCP does exactly the same thing for AI: one plug, and any AI can talk to any tool you already use. No adapter, no custom project at €15,000. It was Anthropic (the company behind Claude) that launched the standard in November 2024, and OpenAI adopted it in turn in March 2025. Today Google and Microsoft too. It became the norm in under 18 months.

What it actually does for you

• Hook Claude up to your Gmail: it reads your emails, sorts them, drafts replies for you.
• Hook Claude up to Notion: it creates pages, updates a database, digs up a note anywhere in your workspace.
• Hook Claude up to Stripe or to your database: you ask "summarize this week's sales" and it queries the real data live, not an estimate.
• On the numbers side: more than 10,000 MCP connectors available by the end of 2025, 97 million downloads a month. It's no longer a gadget, it's a standard building block.

How to use it — step by step

1. Open Claude Desktop (the app you installed) or go to claude.ai in your browser.
2. Click your name at the bottom left, then Settings, then Connectors.
3. Click Add connector. You pick from the list: Gmail, Notion, Slack, Google Drive, Calendar, Stripe, etc.
4. You sign in to your Gmail account (or another) just like logging into a normal app. You click "I authorize." That's it.
5. Back in Claude, you test it by typing: "Summarize my 10 most recent unread emails." It does it.

The most useful connectors when you're an entrepreneur: Gmail and Google Workspace (read/sort/draft, calendar, Drive, Sheets), Notion (official, create and search your pages), Slack (47 actions available), Stripe (conversational sales dashboard), Supabase or HubSpot (query your customer database in plain English).

The analogy that clicks

Imagine three engines in the same garage. Opus is the racing V8: it can do anything, but it drinks fuel and it costs. Sonnet is the family hybrid: powerful when needed, reasonable day to day. Haiku is the electric city car: you take it out for errands, it's quick, it won't break the bank.

I don't pick the V8 to go grab bread. I don't take the city car to cross Europe. With Claude models, it's the same: you match the tool to the mission.

What it actually does for you

• Steer three levers at once: intelligence, speed and cost. The more powerful the model, the more it costs per request.
• Opus saves you the back-and-forth when you want clean strategic reasoning right the first time (an acquisition plan, a contract review, a complex trade-off).
• Sonnet is my default setting for about 90% of uses: writing, research, synthesis, replying to emails.
• Haiku saves up to 80% on repetitive high-volume tasks: sorting 500 emails, extracting data from a list of invoices, tagging leads.

How to use it — step by step

1. On claude.ai: click the model name at the top of your conversation, then select the one you want from the list that opens.
2. In Claude Code: type /model in the window, a list appears, pick the one you're after.
3. On a third-party tool (n8n, Zapier, Cursor, an API integration): copy the exact identifier (for example claude-sonnet-4-6) and paste it into the configuration's "model" field.
4. Test the same prompt on two models when you're unsure: you'll immediately see the difference in quality and speed, and you'll decide with full knowledge.

The analogy that clicks

You walk into a restaurant. On the cook's counter there are several manuals: Italian cooking, pastry, sushi. The cook opens the right manual the moment they need it. Those manuals are the skills. On their own, they cook nothing. They make the cook competent in a specific area the moment they get to it.

The agent is the chef you send to do the shopping, come back, cook, plate up and serve. You say "make dinner," you go back to your own business, and they run through the steps on their own until the plate lands on the table. And while they cook, they can absolutely open a manual (a skill) for tonight's Italian recipe.

Summary: an agent can use skills. A skill, on the other hand, never acts on its own.

What it actually does for you

• A skill installs an expertise once and for all. Example: I have a "prospecting email" skill. Every time I ask "write me an email to this prospect," Claude automatically applies the skill's best practices without me having to remind it.
• An agent carries out a multi-step mission while you do something else. Example: a watch agent that reads 100 sources every morning, filters out the noise, summarizes the essentials and pings you when it's done.
• The combo: my watch agent uses an "editorial synthesis" skill when it writes the final summary. The agent drives, the skill provides the method.

How to put them to work — step by step

1. For a skill: download or copy the skill's folder, drop it into ~/.claude/skills/skill-name/. Inside, there's a SKILL.md file with a description. Claude reads that description automatically and activates the skill as soon as it recognizes the right context. You don't need to call it by hand.
2. For an agent: either you use Claude Code (which is already a working agent day to day), or you go through a tool that wraps it up cleanly (n8n, Zapier, a SaaS solution), or you have a custom agent built via Anthropic's Agent SDK.
3. To check a skill works: fire off a request that matches its specialty and watch whether Claude applies the skill's method without you reminding it. If so, it's in place.
4. To check an agent works: give it a simple mission with a clear finish line ("scan these 10 sources, send me a summary"), check the result, tweak its instructions if needed.

The analogy that clicks

Imagine a smartphone straight out of the factory. It works, but there's not much on it: phone, texts, browser. You open the App Store, you download WhatsApp, Notion, Spotify. In 3 minutes your phone has 10 new superpowers.

It's exactly that for Claude Code. On a base install, it knows how to code and answer. With a plugin, you add a pack of expertise in one command: a working method, a framework's up-to-date docs, automatic review, and so on. Anthropic's official marketplace lists over 140 of them, all tested and signed.

What it actually does for you

• Install a complete methodology in one command (superpowers) instead of rewriting it yourself in your CLAUDE.md.
• Wire a framework's up-to-date official docs into Claude Code (context7) · no more outdated answers from Claude about Next.js or Supabase.
• Get polished front-end interfaces without briefing 20 lines of prompt (frontend-design triggers on its own).

How to install — step by step

1. Open your terminal (see entry #04 CLI / Terminal if you're not comfortable with it).
2. List the plugins already installed with claude plugin list to know where you stand.
3. Install a plugin with claude plugin install <name>. For example claude plugin install context7.
4. Restart Claude Code (close and relaunch) so the new slash commands are recognized.
5. Test it. For superpowers: type /brainstorm in Claude Code on an article idea. For claude-md-management: /revise-claude-md.

You can disable without uninstalling with claude plugin disable <name> · re-enable with claude plugin enable <name>. Always tested and reversible.