MAKE MODE

One moment…

·
A getting-started guide

Start with
Claude Code

An AI collaborator that lives in your terminal. You tell it what you want in plain English, and it builds it. By the end of this guide you'll have a real website of your own, live online.

~10 min setup~30 min first projectNo coding experience

Modules

open +
01Introduction

What AI is, the terminal, the smart-intern mindset, and agent vs. skill.

What AI is, the terminal and why we choose it, the smart-intern mindset, agent vs. skill, and a demo of a personal skill (your DNA).

01What you need
01

What you need

  • A computer or laptop. Not sure which one you need? Check which laptop you need for working with AI ↗.
  • Your own Claude subscription — the Claude Pro plan (about €17/month) includes Claude Code. Make your own account at claude.ai and pick Pro; you'll log in with it in step 2.
  • A free GitHub account, where your code lives.
  • A free Vercel account. It puts your projects on the internet; sign up with GitHub.
  • A terminal. Mac: the Terminal app. Windows: PowerShell.
  • Optional: Supabase, a free database for when a project needs to remember things. Skip it for now.

You don't need to install anything else yourself. Claude sets up the developer tools for you in step 5.

02Install & launch
02

Install & launch

Mac — open Terminal and run:

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

Windows — open PowerShell and run:

PowerShell
irm https://claude.ai/install.ps1 | iex

Then make a folder and start Claude inside it:

Terminal
mkdir my-first-project
cd my-first-project
claude
Logging in

When Claude starts, it opens your browser to log in with your Claude account — use the one with your Pro subscription. No API key needed.

Mental model

Claude works in the folder you started it in, and you talk to it in plain English. One folder per project.

03Bypass permissions
03

Turn on bypass permissions

By default, Claude checks with you before every edit and command. For an expert that's useful. While you're learning it just gets in the way, so turn it off:

Paste into Claude
Turn on bypass permissions permanently — change my settings so I don't have to approve every action.

Claude edits its own settings file. Then restart it: type exit, run claude again, and accept the one-time warning.

Why this is safe enough

You work in one folder per project, so there's nothing here Claude can break that you couldn't rebuild. Git keeps a history you can roll back to, and your CLAUDE.md tells Claude never to delete anything without checking first.

04Your CLAUDE.md
04

Your CLAUDE.md

CLAUDE.md is a file of standing instructions Claude reads at the start of every session. Copy the starter, then tell Claude: "create a CLAUDE.md with this content" and paste.

Starter CLAUDE.md
# My Claude Code Preferences

I'm a beginner learning to build things with AI. Help me learn, not just ship.

## How to communicate
- Explain *why* you chose an approach before doing it.
- Make changes in files; tell me in plain English what you did.
- When something errors, explain what it means before fixing it.
- Use simple, readable solutions over clever ones.

## Safety
- Commit to git after every working change. Git is my undo button.
- Never delete files or run destructive commands without asking.
- If about to do something irreversible, stop and explain first.

## Working style
- For anything bigger than a small fix, give me a 2-3 sentence plan first.
- Verify changes actually work before saying it's done.
- One task at a time. Finish and verify before moving on.

## Memory
- Keep a NOTES.md: plan, decisions, what's done, what's next.
- When I say "gnite": commit + push, update NOTES.md, short summary.

Or have Claude write the file for you. This prompt (from Derek) does that, and adds one thing the starter above misses: an honesty rule — Claude should push back when you're wrong, and never present a guess as fact.

Paste into Claude — Derek's version
Create a CLAUDE.md file in my project folder with the standing preferences below, so you read them at the start of every session. When it's saved, show me the file and give me a one-line recap.

# My Claude Code Preferences

## Teaching
- Explain what you're doing and why, in plain language. Define jargon the first time you use it.
- When there's a better or more standard approach than what I asked for, tell me and explain the trade-off.
- End each task with a short recap: what changed, why, and what I could try next.

## How to work
- Be autonomous: do everything you can yourself. Don't ask me to run commands or make edits you can do.
- Before any large or multi-file change, show me a short plan first and wait for my OK.
- Keep changes small, simple, and reversible. Prefer boring, well-documented solutions I can learn from.
- If something fails, try another approach before asking me, then tell me what went wrong and how you fixed it.

## Safety
- Warn me clearly before anything hard to undo: deleting files, force-pushing, changing settings, or anything risky.
- Use git as a safety net: commit each working step so we can always undo. Don't commit secrets or API keys.
- Never enter my passwords or payment details, hand those steps to me.

## Honesty
- Push back when I'm wrong. Don't just agree with me.
- Actually run or test changes and show me the real result before saying it's done.
- If you're unsure or guessing, say so. Don't present a guess as fact.
The most important thing

The file grows with you. Every time Claude does something you didn't like, add a rule for it. By the end, the CLAUDE.md is yours.

05GitHub & Vercel
05

Set up GitHub & Vercel

Now let Claude set up your developer tools — this is where those free accounts come in:

Paste into Claude
Set up my computer for web development. Install the GitHub CLI and the Vercel CLI, then help me log in to both. Make sure git knows my name and email. Explain each step — I'm a beginner.
Handy trick

If Claude asks you to run a command yourself, type ! before it (e.g. ! gh auth login). The ! runs it inside the conversation, so Claude sees the output.

06First project
06

Your first project

You're set up. Now build something real:

Paste into Claude
Build me a simple one-page personal site: my name, a short bio, and three project cards. Keep the code simple enough that I can read it. When it works, show me how to see it in my browser, then deploy it to Vercel so I have a public URL.

Thirty minutes from now you'll have a real URL you can send to anyone.

07Survival skills
07

Survival skills

The things experienced users do that beginners don't know exist.

EscStops Claude in the middle of a task. Use it any time, as often as you like.
/clearWipes the conversation. Start a new task with a clean one; long sessions get slower and pricier.
/exit + /loginWhen Claude acts strange, quit, restart, run /login. Fixes a lot.
claude -cClosed the terminal? Run this in your folder to continue your last conversation.
Shift + TabCycles permission modes, including plan mode — approve a plan before it builds.
Paste the errorCopy the whole error in and ask "what does this mean?" before a fix. This is how you learn.
"Did it work?""Done!" is not proof. Open the page, click the button, verify yourself.
Paste a pictureCtrl + V (yes, Ctrl — even on Mac) pastes a screenshot. Claude can see images.
"do it"Short replies are fine: "yes", "do it", "keep going".
/helpShows everything else.
08Level up
08

Level up

Once the basics feel comfortable, a few habits that make you faster.

  • Use multiple windows. Open several terminals, run claude in each, work on different parts at once — it's how experienced people work.
  • The web is your reference library. See a component you love? Screenshot it or paste the link, and ask Claude to recreate it in your style.
  • Make Claude write things down. Keep memory in files, not the chat. Your CLAUDE.md keeps a NOTES.md, so you can /clear freely.
  • Talk like a director. Nearly half of expert prompts are under 30 characters: "ok", "do it", "why?", "wait". The skill is steering with short nudges.
09Unlocks
09

Unlocks & when you're stuck

Claude can use almost any tool or service. Each account you connect gives it a new power.

  • A database — Supabase. Lets your projects remember: sign-ups, favourites, comments. Free account at supabase.com, then: "Add a database to this project using Supabase — walk me through connecting it."
  • Eyes — let Claude see your site. Normally Claude builds blind. Give it the ability to open your page and look, so it checks its own work.
When you're stuck

Paste the full error and ask what it means. Reopen the terminal. Run /exit then /login. And remember — you can always ask Claude itself.

Questions from the group — session 1
How many windows can I open?No limit from Claude — the brake is your laptop and your own overview. Rule of thumb: one window per project.
"Meet Sonnet 5…"?A system message, not an error. A newer model is available; you don't have to do anything. Switch with /model if you like.
App or terminal?Code in the Claude app is the same Claude Code as in the terminal — different jacket, same engine. Chat is for conversations, Cowork works alongside you in your files and mail. Pick whichever feels comfortable.
Do I need an API key?No. You log in with your Claude account; the Pro plan (about €17/month) is all you need. Open the terminal, type claude and follow the login step.
Pasting acts weirdText from Word or email drags hidden formatting along. Paste into Notes first (plain text), then into the terminal. Images paste with Ctrl + V — yes, Ctrl, even on Mac.
Homework

Gather material for your own DNA skill.

open +
02Your own skill

A skill that writes and sounds like you. Your first personal win.

The basics of a skill, and you build your first one. We also pick up the jargon you keep hearing but don't fully know yet. This is the overview; some topics get a deeper look in later modules.

01What is a skill?
01

What is a skill?

The official explanation, straight from AnthropicWatch on YouTube ↗

A skill is a reusable recipe. You write down once how you want a task done: the tone, the steps, the format. From then on Claude follows that recipe every time, so you get the same result without explaining it again.

In practice a skill is just a text file with instructions, sitting in a folder on your computer. Nothing mysterious: no code, no installation. You can open and read it yourself, just like a Word document.

The clever part is how Claude handles it. A skill sits quietly until you need it. Only when a task matches what the skill does, Claude picks it up and reads the instructions. That means you can have dozens of skills without weighing down your conversation.

Not unique to Claude Code

Anthropic builds skills to an open standard, Agent Skills. The same skill you make can work in other AI tools too, not just Claude Code.

Anthropic: "Introducing Agent Skills" (official announcement) ↗Anthropic engineering: how skills work under the hood ↗
02When do you use a skill?
02

When do you use a skill?

A skill was built by Anthropic for two reasons: you get better, more consistent results, and you don't have to type the whole prompt every time. You save the how-to once, and from then on you just call the skill.

Claude Code is an agent: you give it a goal and it works out the steps on its own. A skill makes sure a specific part is always done your way. So use one for anything you do more than once, like a quote, a newsletter, or a report.

You can also just call a skill in a chat, but most of its value is in standardized output. Think of a presentation that always looks the same, an email in your fixed tone, or a report in a fixed format. Without a skill, Claude decides that structure itself, and it differs each time. With a skill, you hand it the structure, so you get it back every time.

The handy part is also where you can use it. You can call a skill in an app, you can hand an agent a skill to work with, and you can even chain skills together. More on that in the later modules.

CLAUDE.md versus a skill

Your CLAUDE.md file is always “on”: Claude reads it in every conversation, so it's for fixed facts about you or your project, like your style, your company, or your preferences. A skill works the other way round: it sits quietly until the task calls for it, so it's for a procedure, a format, a step-by-step recipe.

Rule of thumb: if a piece of your CLAUDE.md grows into a recipe with steps, instead of just a fact, that's a sign it should really be a skill.

03How do you use this?
03

How do you use this?

Just ask in plain language, and name the skill you want used. For example: research my competitors, then write it up as a report with my report skill. Claude does the research and delivers it in your saved format.

A few more examples of what that looks like in practice:

  • Write this email with my writing skill. — the same tone as always, no doubt about how formal it should be.
  • Make a quote for Lisa with my quote skill. — your fixed structure and pricing, only the client name changes.
  • Turn this into a presentation with my brand-deck skill. — the same house style and slide structure as your previous decks.
  • Write a LinkedIn post about this project with my DNA skill. — it sounds like you, not like a generic AI post.
  • Do my monthly bookkeeping with my ZZP finance skill. — the same overview every time, without explaining the steps again.
04How do you make a skill?
04

How do you make a skill?

You make a skill by giving examples. Put the texts you were happy with in one place, like a Word document of emails, or gather your old LinkedIn posts into a file. Then tell Claude you want to make a skill from them. The skill creator (built by Anthropic) kicks in automatically and walks you through it.

Try the skill on something real. Not happy? Fine-tune it in plain language: 'make it shorter', 'a bit more formal'. It adjusts.

A smart question to ask Claude

Ask: 'what questions can I answer to make my goal as clear as possible?' Claude then shows you where your skill is still ambiguous, so you can sharpen it.

Good first skills: a brand guide (how your brand looks and sounds), a DNA skill (writes in your voice), or a recurring business task like a quote or a report.

The one thing that matters most

Context. The more you tell Claude, the better the skill.

05Tips & tricks
05

Tips & tricks

Keep skills small and specific. One skill that writes in your voice, one that knows how your LinkedIn posts work. That works better than one giant skill that tries to do a whole workflow.

You can chain them in one chat: 'first use my writing skill, then my LinkedIn skill.' That's a small workflow of your own.

Find your skills

Type / in Claude to see your commands and skills.

Give Claude a file (more in module 5)

Copy the path, not the file itself. On Mac: hold Option, right-click the file and choose 'Copy as Pathname'. Paste that path into Claude.

open +
03Prompting & Plan Mode

Making your goal clear, breaking a task into steps, and plan mode.

Claude works best when you tell it clearly what you want. In this module you'll learn how to write a good request, and how to use Plan Mode so Claude shows you its plan before it does anything.

01A prompt is: explaining well
01

A prompt is just: explaining well

Prompt is a big word for a small thing: it's simply what you type. The best ask sounds like how you'd explain something to a smart intern on day one — they get a lot, but they don't yet know anything about your business.

Tell it too little and it guesses. Tell it exactly what you mean and you get exactly what you need. The difference isn't fancy words, it's how much you tell it.

Vague versus clear

"Make a newsletter" leaves Claude to guess the rest. "A short newsletter about our summer sale, warm tone, three paragraphs, with a button to the shop" gives it direction.

02The three things
02

Three things every good ask has

You don't need to memorise a checklist. Picture that intern, and tell it the three things you'd tell any stand-in:

  • The goal — where it should end up, and how you'll know it's done.
  • The context — which file, which client, which example. Point at it.
  • The steps — for a big job: break it up. The smaller the step, the less there is to guess.

Rule of thumb: the more you'd have to explain to a stand-in, the more you write down.

03Break a job up
03

Break a big job into small steps

Say "sort out the whole admin" and anyone stalls — person or Claude. Say "first this, then that" and it flows. Think of a recipe: not "make a cake" and walk off, but weigh first, then mix, then into the oven.

Just list the steps under each other in one ask. Copy this template and fill it in:

Copy & fill in
Ik wil [wat je wil bereiken]. Belangrijk: [context — welk bestand, welke klant, welke stijl].

Doe het in stappen:
1. eerst [stap]
2. dan [stap]
3. dan [stap]

Laat het na elke stap even zien.
04Plan mode
04

Plan mode: the sketch before the build

For something big, you don't want Claude diving straight in. You want to see the plan first. That's plan mode: Claude writes down what it's going to do, and changes nothing until you say yes.

Think of a builder: a good one doesn't knock down a wall without putting a sketch on the table first. You check it's right, and only when you nod does the first hammer swing.

Copy & fill in
Maak eerst een plan voor [de klus]. Verander nog niks — laat alleen de stappen zien die je zou nemen, en wacht op mijn oké.
Shortcut

Press Shift + Tab until "plan mode" shows at the bottom. The same combo switches it back off.

05Is your ask good?
05

How you know your ask is a good one

Four things to run past before you hit enter:

  • You said what "done" means.
  • You pointed at the right things — file, client, example.
  • You broke a big job into separate steps.
  • For anything risky, you asked for a plan first.
Homework

Think of one small thing you'd like to make. Write it down as a clear request in one or two sentences, then break it into three steps. Next, try turning on Plan Mode and let Claude show you its plan first.

open +
04Install Cursor

Get Cursor onto your computer — Mac and Windows, up to your first terminal.

Cursor is a friendlier window onto the same Claude you use in the terminal. In this module you get it onto your computer in ten minutes: download, install, open it the first time, and start your first terminal. Nothing you can break.

01Get it
01

Get it from the official site

Cursor comes from one place: cursor.com. Go there and click the big Download button. The site already sees whether you're on a Mac or Windows. A short list of versions appears; you only pick one:

  • On a Mac: "Mac (Universal)" — works on any Mac. You download a .dmg, think of it as a moving box with the app inside.
  • On Windows: the "User" version — installs without admin rights. You download a .exe, the installer file.

The file lands in your Downloads folder. Wait a moment for it to finish.

02Install it
02

Install it

On a Mac: double-click the .dmg file. A little window opens with the Cursor icon and an "Applications" folder. Drag the icon into that folder — you're putting it in the cupboard where your computer keeps its programs, so it can find it from now on.

On Windows: double-click the .exe file and click through a few times (Next, Install). When it's finished, Cursor sits in your Start menu.

03Open it the first time
03

Open it the first time

Open Cursor like any other program: on a Mac from your Applications folder or Launchpad, on Windows from the Start menu.

On a Mac it asks a check question the first time ("Are you sure you want to open this?"). That's normal for anything you download from the internet; your computer is playing doorman. Click Open.

Then a few setup questions (leave the defaults), and finally sign in or make a free account in one click. Then you're in.

04Your first terminal
04

Open your first terminal

The terminal is the black strip where you give Claude Code its instructions. It isn't open by default; you bring it up yourself.

Click Terminal → New Terminal at the top, or press Ctrl + ` (the key top-left, left of the 1) — same key on Mac and Windows. A black strip slides open at the bottom.

For a second terminal, click the + at the top-right of that strip. This strip is your starting point: it's where you start Claude Code.

Homework

Install Cursor, sign in, and open a terminal once with Ctrl + `. That's all — you're ready to build.

open +
05Make Cursor yours

Make Cursor yours — settings, safe commands, your terminals and your statusline.

Cursor's installed, you're in. Now you make it yours: a calmer window, your own settings, and a few conveniences so it fits the way you work.

01The friendlier window
01

The friendlier window

The terminal does the job, but it's a bare place to live: one black window per task. Cursor is a calmer, more visual window onto the same Claude. Nothing changes about what it can do — it's just a friendlier place to work.

You might hear two words. The IDE is just the window you see. The CLI is the assistant working in the terminal — the same Claude Code you already know, now in one panel instead of your whole screen. That's all you need to know.

Do I need to move everything over?

No. Your Claude Code still works: open the terminal inside Cursor and run it like before. Your CLAUDE.md keeps working too, so your preferences carry over.

02Make it yours
02

Make it yours

Once you're in, you make it fit you. In Settings, these are the ones worth touching:

  • Rules for AI (CLAUDE.md) — the stuff Claude should always know: your business, your style. It reads them every chat.
  • Model — which Claude answers you. Keep it on the strongest one.
  • Theme — the colours of the editor. Pick what's easy on your eyes.
  • Format on save — tidies the code automatically when you save. One less thing to think about.
03Let the safe steps through
03

Let the safe steps through

Claude checks with you before it does things — good for safety, but it adds up when it keeps asking for the same harmless steps. Wave the safe ones through, so it only stops you for what actually matters.

Run this in Claude Code
/fewer-permission-prompts

It looks at what you keep approving and lets the everyday, safe commands through — and still asks first before anything risky.

04Name your terminals
04

Name your terminals

You'll open a few terminals at once. Give each one a name — for what it's doing — so you never lose your place. Paste this into Claude Code and change the wording to your own tasks:

Copy & fill in
Houd de tab van deze terminal genoemd naar waar we mee bezig zijn. Zet de tabtitel op een kort label voor de huidige taak, en hernoem 'm zodra we op iets nieuws overstappen. Houd het onder de vier woorden.

Think of names like June newsletter, Quote for Lisa, Website fix or Plan SmartGirls — short and concrete, so you know at a glance which terminal is for what.

In Cursor you can also just right-click a tab and rename it by hand.

05Your statusline
05

Your statusline

The statusline is the little strip at the bottom of a session. Fully customizable — show only the few things you actually glance at: your project, the model, how much context is left. Ask Claude to build yours:

Copy & fill in
Stel mijn statusline in om [project], het [model] en hoeveel [context] er over is te tonen. Houd het op één regel, met een klein balkje voor de context.

For example, a statusline might look like this:

makemode · Opus 4.8 (1M) · 62% context left · manual mode

A minimal version works too: just project · model · context — exactly the three things you glance at most.

Remember

There's no right setup. Switch on what gives you calm or time, and leave the rest off. It should fit your day, not someone else's.

open +
06Worktrees

One site, two changes at once — a separate copy to experiment on safely.

Your site is live. Sometimes you want to try two things at once — new colours and a new headline — without them getting in each other's way, and without touching your live site. A worktree is a separate copy of your project on its own desk, so you can.

What a worktree is

Nothing magical: a separate copy of your project in its own folder, with its own files and its own branch, but the same history as your original. It comes from Git, a proven tool. Your original is called main; the copy is the worktree.

One site, two changes at once

Make two copies and do one change in each — new colours in one, a new headline in the other. They never collide. Keep whichever wins and merge it back; throw the other away. Your live site was never touched. (It's the physical counterpart to an agent: an agent gets its own memory, a worktree gets its own copy of your files.)

You just ask Claude

You don't memorize commands. Claude Code has it built in: claude --worktree new-colours spins up an isolated copy on its own branch. Underneath it's plain git (add, list, remove) — you never type it. Just say: "Make a separate copy of this site so I can try [a change] without touching my live one."

Homework

Set up one worktree for a change you don't dare make on your real site — build it, then throw it away. Your live site stays exactly as you left it.

Want the whole lesson as a presentation? View the session slides →

Coming soon
07Business skill

Capture a recurring task from your business in a skill.

You do some tasks over and over: a price quote, a newsletter, a reply to a customer. A Skill lets you teach Claude Code that task once, so next time you just ask for it in one sentence.

Spot a task you repeat

Think about what you type or write again and again each week. Maybe it is a quote with your prices, a friendly reply to a booking, or your weekly newsletter. Pick one task that always follows the same shape. That is your first Skill.

Teach it once

Open Claude Code in your terminal and describe the task in plain words. Tell it what the result should look like, your tone, and any details that never change. Ask Claude Code to save this as a Skill. Now it remembers how you like it done.

Reuse it in one sentence

Next time you just say what you need, like "Make a quote for a one-day workshop for Lisa." Claude Code uses your Skill and fills in the rest. You check it, tweak a word if you want, and send. What took twenty minutes now takes one.

Homework

Pick one task you repeat every week and turn it into a Skill. Then test it by asking for it in a single sentence.

Coming soon
08Files & folders

Paths, files and folders — your computer as a workspace.

Your computer is a workspace, and right now it can feel like a black box. In this module you'll learn how files and folders are organised, so you can hand Claude exactly the right things to work with.

A path is just an address

Every file on your computer has an address, called a path. It tells your computer exactly where to find something, folder by folder. For example, /Users/you/Desktop/notes.txt means: the file notes.txt, on your Desktop, in your account. Once you see a path as an address, it stops looking like code.

Folders keep your work tidy

A folder is a box that holds files and other folders. Keeping related files together in one folder makes your work easier to find, for you and for Claude. Before you start a task, it helps to know which folder you're in. You can always ask Claude to show you the files in your current folder.

Give Claude the right files

Claude works best when you point it at the exact file or folder you mean. Instead of describing a document from memory, name it or share its path. The clearer you are about which files to use, the better the result. Think of it as handing someone the right paper instead of pointing vaguely at a pile.

Homework

Pick one folder on your computer that you use a lot. Ask Claude to list the files inside it, then ask it to tell you the full path of one file you care about. Notice how the path is just an address you can read.

open +
09Agents

A helper that knows who it is and works out the steps itself. From recipe to colleague.

Last week you built a skill: a recipe you call up. This week you build the worker that runs it on its own, an agent. You'll see what an agent really is, how it wakes up, and how to make one you can trust.

01The mental model
01

First, the mental model

Three words keep it clear. Every agent is a who, woken by a when, and made reliable by a how.

WHO
The agent

A helper with one role and its own workspace. Sits idle until called.

WHEN
The trigger

What wakes it: you, a schedule, or an event.

HOW
A skill

Your recipe, handed to the agent so the output is consistent.

Mental model

An agent is not a CLAUDE.md file (those are just your rules), and it never runs all day. It waits for its trigger, does its job in its own room, and goes back to sleep.

02What wakes it — the trigger
02

What wakes an agent — the trigger

Three kinds of trigger wake an agent:

  • You ask (on-demand) — you type it: "plan my day." The agent runs there and then.
  • A schedule (the clock) — a fixed time: every day at 7:00, or weekly. A routine wakes it.
  • An event (something happens) — a new email, a WhatsApp "can we meet?", a form submission. The event is the trigger.
03Two ways to make one
03

Two ways to make one

  • Just ask Claude. "Make an agent that checks every page against my brand style." Claude writes the file for you — and picks up any change within seconds, no restart. Easiest start.
  • Write the file yourself. A markdown file in .claude/agents/ (this project) or ~/.claude/agents/ (everywhere) when you want full control. Run /agents to see the ones you already have.
04What's inside an agent?
04

What's inside an agent

Every agent — like the brand-checker.md below — is a short file with four parts:

  • name — what it's called.
  • description — the "use me when…" line. This is how it gets triggered automatically — so write it precisely.
  • tools — which tools it may use. Give it only what it needs.
  • instructions — its role and how it works: your standards, what "good" looks like, the output format. The heart of the agent.
~/.claude/agents/brand-checker.md
---
name: brand-checker
description: Use when a new page is done and needs a house-style check.
tools: Read, Grep          # read-only — it only reviews
model: sonnet
---

You are my brand guardian. Check the page against my house style:
  - colours: ink + electric blue only
  - font: Helvetica, no spaced-caps labels
  - tone: warm, plain, never salesy

Report only what's off, with the file and line. Suggest the fix.
Do not change files yourself.
05Where your agents live
05

Where your agents live

Two homes, depending on who should have them:

  • .claude/agents/ — this project only. Check it into version control so your team shares it.
  • ~/.claude/agents/ — all your projects: your personal helpers, everywhere.

Only name and description are required. Leave tools out and it inherits everything; add model: haiku to keep a simple agent cheap and fast.

You already have some, built in

Claude Code ships with subagents it uses on its own — Explore (searches your codebase in its own window), Plan (research during plan mode), and a general-purpose one. Your custom agents sit right alongside them.

06One agent, or a team?
06

One agent, or an agent with staff?

Same kind of worker either way — it runs the same loop of observe → think → act. What changes is the role. Think of a kitchen.

STRUCTURE A
The food truck · one agent
1 takes order→ 2 cooks→ 3 takes payment↻

One person does every step themselves, one after another. Simple and quick — but it's one pair of hands: tasks wait for each other.

STRUCTURE B
The professional kitchen · parallel
chef
startermaindessert all at once ⇉

A lead agent (chef) splits the work across subagents (staff) who cook in parallel — each focused, each handed the right recipe (skill). The chef combines the plates.

The one thing to remember

Same kind of worker — different role. Run alone at the top and it's the agent; hired by another to do one job and it's a subagent. Nothing else about it changes.

07Why a subagent?
07

Why reach for a subagent?

The real win is that own room. A subagent does the messy work in its own context window: reading dozens of files, digging through logs. It hands back only the summary, so your main conversation stays clean.

YOUR CONVERSATION
Stays clean and focused. You see the result — not the digging that got there.
delegates→
←returns only the summary
SUBAGENT · its own window
• read 40 files…
• grep the logs…
• trace the issue…
• weigh three fixes…
✓ found it

A tidy conversation is just the start. A subagent also lets you:

  • Preserve context — heavy exploration stays out of your main chat.
  • Enforce limits — restrict which tools it may touch, e.g. read-only.
  • Reuse it everywhere — a user-level agent works across all your projects.
  • Specialise — a focused system prompt for one domain.
  • Control cost — route simple work to a faster, cheaper model like Haiku.
  • Run on a trigger — you, a schedule, or an event.
08What makes it good?
08

What makes it a good agent

  • One clear role. One job, sharply scoped. Narrow beats broad.
  • A precise description. It's the trigger — write it like the moments you want it to fire. Add "use proactively" to nudge Claude to reach for it on its own.
  • Only the tools it needs. Review-only? Give it read access, nothing more.
  • Tell it how "good" looks. Your standard, the format, what to watch for. This is where a skill plugs in.
  • Keep its work self-contained. It should do its job without your whole conversation.
  • Iterate. Run it, read the output, refine. Never right the first time.
  • Choose the trigger on purpose. You · a schedule · an event.
The golden rule

Skill first, agent second. Build and test the recipe until you trust it. Then build the agent that runs it on its own — so the output is right, even when you're not watching.

09This week's assignment
09

This week's assignment — an agent that does the work

Last week's agent only looked. This week you build one that acts. Pick a task you do again and again, and let the agent run it from start to finish, using your skills as recipes along the way.

  • 1 · Give it the recipes — its skills. The skills it leans on: your brand-guide, your writing voice, your quote format. It reaches for one the moment the task needs it.
  • 2 · Let it act — the tools. This one isn't read-only. Give it the tools to do the work: write and edit files, run things, save the result. One job, sharply scoped.
  • 3 · Give it a goal and a trigger — the when. You ask, a schedule, or an event. It picks which skill to use, does the steps, and puts the result back in front of you to check.

Open Claude Code and paste this, with your own task dropped in:

Paste into Claude
Make an agent called [name] that does [recurring task] for me from start to finish. It uses my skills (brand-guide, [others]) as it works, and it may write, edit and run what it needs, not only review. One clear job. Trigger it when I ask. Show me the result to approve.
Because it acts now

An agent that writes and runs can do more, so it can also get more wrong. Keep it to a safe folder, use git so you can roll back, and read its work. Be stricter the moment real files, money or client data come into play.

Hand in — you're done when

The agent took a real task off your plate this week. It reached for at least one of your skills, did the work itself, and handed you something you only had to check.

Assignment

Build an agent this week that runs a recurring task from start to finish (panel 09): it reaches for your skills, it's allowed to act, and you check the result. Put it to work once and post what it did on the forum.

Coming soon
10Settings & permissions

Settings, a sound on permission prompts, and bypass permissions.

By now you have seen Claude Code ask for permission before it does things. This module shows you how to shape those moments: where the settings live, how to hear when Claude needs you, and when it is safe to let it run on its own.

Where your settings live

Claude Code keeps its settings in a small file you control. Type /config to open it and see your options. You do not need to memorise anything here. Just know this is the place you come back to when you want to change how Claude behaves.

A sound when Claude needs you

When you give Claude a big task, it works for a while and then pauses to ask permission. It is easy to miss that pause if you looked away. You can turn on a sound so your computer chimes the moment Claude needs your answer. That way you can let it work and trust it will call you back.

Bypass permissions, and the habit that keeps you safe

There is a mode that lets Claude act without asking each time, often called bypass permissions. It feels faster, but you give up the checkpoint where you can say no. Only use it for small, low-risk tasks where nothing important can break. Whatever mode you are in, keep one habit: read what Claude is about to do before you approve it, every time.

Homework

Open /config and find where permissions and notifications are set. Turn on the sound for permission prompts, then give Claude a small task and notice the chime when it pauses to ask you.

Coming soon
11Checking output

Spotting AI mistakes, checking output. A mid-point reflection.

Claude is fast and helpful, but it is not always right. Your job is to check the work before you trust it, especially when numbers, money, or customer details are involved.

Read it before you run it

When Claude writes code or makes a change, take a moment to look at what it did. You do not need to understand every line. Ask yourself: does this match what I asked for? If something feels off, say so and ask Claude to explain it in plain words.

Be extra careful with numbers, money and customer data

These are the things that hurt most when they go wrong. Double-check any total, price, invoice or calculation by hand. Never paste real customer details into a place you have not checked first. When in doubt, test with fake data before you use the real thing.

An honest check-in

You are halfway through. Be honest with yourself: are you really reading what Claude makes, or just clicking yes? There is no shame in slowing down. The people who get the most out of this tool are the ones who stay curious and keep checking.

Homework

Ask Claude to make a small change, then before you accept it, write one sentence explaining in your own words what it did. If you cannot, ask Claude to explain until you can.

Coming soon
12Choosing your project

One real task or workflow from your business. Setting the scope.

This is where Claude Code stops being an idea and starts being useful to you. Before you build anything, you pick one real piece of your own work to point it at.

Start with a task you already do

Look at your own week and find one task that keeps coming back. Maybe it's writing the same kind of email, sorting receipts, or turning notes into a tidy document. Pick something boring and repetitive, not your most exciting idea. The familiar tasks are the easiest place to see Claude Code actually help.

Scope it down until it feels small

Most first projects are too big. Take your task and shrink it to one clear step with a clear result. Instead of "organize my whole business," try "make one weekly summary from this one folder." If you can describe the finished result in a single sentence, your scope is good.

Check that you can tell when it's done

A good first project has an obvious finish line. You should be able to look at the result and say "yes, that's right" or "no, not yet." If you can't tell whether it worked, the task is still too fuzzy. Keep trimming until success is something you can see.

Homework

Write down three repetitive tasks from your own business. Circle the smallest one, then finish this sentence: "When this works, I will have one ___ that I can check by ___." Bring that sentence to the next module.

open +
13Building

Hands-on on your own project, with guidance where you get stuck.

This is where it all comes together. You are going to build your own project, one small step at a time, with Claude Code helping you along the way.

Start small and tell Claude what you want

Pick one clear thing you want to make. Describe it to Claude Code in plain words, just like you would explain it to a friend. You do not need the perfect prompt. Start with one feature, see what happens, and build from there.

Getting stuck is part of building

Something will break. That is normal, and it is not a sign you did anything wrong. Copy the error message and paste it straight into Claude Code, then ask what it means and how to fix it. Claude reads the error and walks you through the next step.

Iterate until it works

Building is rarely right the first time. Try something, look at the result, and tell Claude what to change. Keep these small loops going, and the project slowly becomes the thing you pictured.

Homework

Write down one small project you want to build in a single sentence. Open Claude Code, describe it, and ask for the very first step. Stop when you have something on screen, even if it is tiny.

Want the whole lesson, step by step? Open the full tutorial →

open +
14Connectors (MCP)

Give Claude access to your calendar, mail and files — with one click.

On its own, Claude only sees the files you open. A connector opens a door to a service outside — your calendar, Gmail, Drive. MCP is just the technical name for that plug.

Switch one on

On claude.ai, go to Settings → Connectors and click Connect on Google Calendar or Gmail. Sign in with Google and click Allow. No code to remember.

Check it with /mcp

Back in Claude Code, type /mcp to see your connectors and whether they're working. Green means connected — that screen is your dashboard.

Gmail or Apple Mail?

Gmail lives in the cloud, so it needs the connector. Apple Mail lives on your Mac itself — Claude can reach it directly, you just grant permission the first time.

Homework

Switch on one connector and ask Claude a real question that uses it, like "Am I free Thursday?".

Want the whole lesson, step by step? Open the full tutorial →

open +
15Automating

Letting your workflow run, with checks built in.

Until now, you said yes to each step. In this module you let Claude run a small workflow on its own, with checks built in so you stay in control. You decide what to hand over and what to keep in your own hands.

What you can let go

Repetitive, low-risk work is the easiest to hand over. Think of renaming files, formatting text, or running the same check every time. These are steps where a mistake is easy to spot and easy to undo. Start there, and let Claude do them without asking each time.

What you keep in hand

Keep a hand on anything that is hard to undo or that really matters. Sending an email, deleting something, or publishing a page are moments to pause and look first. A good workflow has these checkpoints built in, so it stops and waits for your yes. You are still the one who decides.

A first look at hooks

A hook is a small rule that runs automatically at a set moment. For example: every time Claude finishes a task, run a check. You set it once, and it works in the background without you thinking about it. You do not need to build one yet. For now, just know that hooks are how you make your checks happen on their own.

Homework

Pick one task you do over and over in Claude Code. Write down two things: which part you would happily let run on its own, and which part you always want to check yourself before it happens.

Want the whole lesson, step by step? Open the full tutorial →

open +
16Work without your laptop

Let it run in the cloud with your laptop shut — or keep your Mac awake for a long job.

You don't have to sit at your desk while Claude works. Either the cloud does the job with your laptop shut, or — for a long local job — you keep your Mac awake.

A routine in the cloud

At claude.ai/code/routines, set a task and a time. It runs on Anthropic's servers, even with your laptop closed. (At most once an hour, and it can't see your local files.)

Keep your Mac awake

For a long job on your own Mac, start Claude with a coffee in front of it — type caffeinate -is claude. Your Mac won't fall asleep while it works. Or click the pill in Amphetamine (free app). Keep it on the charger.

Which one?

Away, laptop shut, recurring? → a cloud routine. A long one-off that needs your own files? → run it locally and keep your Mac awake.

Homework

Set up one cloud routine that runs with your laptop shut.

Want the whole lesson, step by step? Open the full tutorial →

open +
17Showcase

Show what you built. What do you take with you, how do we keep sharing?

You made it. This last module is about showing what you built and finding your people so the learning keeps going after the course ends.

Show what you built

Pick one thing you made and share it. It does not have to be finished or perfect. Tell people what it is, what problem it solves for you, and what surprised you while building it. A small real project says more than a long explanation.

What you take with you

You can now describe what you want, run it in the terminal, and fix things when they break. That skill stays with you, whatever you build next. Save your favourite prompts and notes in one place so your future self can reuse them. You are not starting from scratch anymore.

Keep learning together

Learning sticks better with other people around. Stay in the group, share your wins and your stuck moments, and ask questions out loud. When someone else gets stuck, help them. That back-and-forth is how a course turns into a community.

Homework

Share one thing you built with the group this week. Add two sentences: what it does, and one thing you learned making it.

Ready to share it? Put your site online and hand it over →

Reference

The tools you'll use

Tap a card to open it ↗
The AI
Claude Code

Your AI collaborator in the terminal. Tell it what you want in plain English and it writes the files, runs the commands, and looks things up online. The whole guide runs on it.

Open the docs ↗
Your account
Claude subscription

Where you sign up and manage your plan. The Pro plan (about €17/month) includes Claude Code — make your own account and log in with it. No API key needed.

Get Claude Pro ↗
Code hosting
GitHub

A free home for your code online. It keeps every version, so you can always go back to one that worked.

Open GitHub ↗
Hosting
Vercel

Turns your project into a real website with a public URL you can share. Free, and it connects straight to your GitHub.

Open Vercel ↗
Database
Supabase

A free database, for when your app needs to remember things like logins or comments. You can add it later.

Open Supabase ↗
Built-in
Terminal

Already on your computer, nothing to install. Mac: the Terminal app. Windows: PowerShell. This is where you run Claude.

Already on your Mac / PC
In plain words

What do these words mean?

No jargon — promise
Terminal
The text window where you type instructions to your computer instead of clicking. On Mac it's the "Terminal" app; on Windows it's "PowerShell." It's where you run Claude.
Command
A short line of text you give the computer to do something. Usually you copy it, paste it, and press enter.
Claude subscription
Your own paid Claude plan. The Pro plan (about €17/month) includes Claude Code — that's all you need to run it. You log in with your account; no extra code to paste.
Folder
A place on your computer that holds all the files of one project. Rule of thumb: one folder per project.
GitHub
A free online home for your code. It saves every version, so you can always go back to an earlier one.
Commit
Saving a snapshot of your work to that history, with a short note about what changed.
Deploy
Putting your project online so anyone with the link can open it in their browser.
Vercel
The free service that deploys your site and gives it a web address (a URL) you can share.
CLI
Short for "command-line tool": a program you control by typing instead of clicking. The GitHub and Vercel tools are both CLIs.
Database
A place your app stores things it needs to remember — sign-ups, saved data, comments. Supabase is one.
CLAUDE.md
A plain text file of your standing instructions that Claude reads at the start of every session.
Bypass permissions
Letting Claude make changes without checking with you each time. It keeps things moving while you learn.
Vibe coding
Building by saying what you want in plain language while the AI writes the code. You steer on the result, not on code rules. The term was coined by AI researcher Andrej Karpathy (2025).
Ask for help

Confused by a word? Ask here

Not your real terminal — a sandbox to ask questions and click the jargon
From the workshops

Questions from the group

Real participant questions — updated after each session
How is Claude Code different from ChatGPT?+

ChatGPT is mostly a chat window: you ask, it answers, and you copy the result somewhere yourself. Claude Code works inside your computer: it reads and creates files, runs commands, and actually puts a website live — no manual copy-pasting.

In short: ChatGPT talks, Claude Code does.

Read more: Zapier: Claude vs. ChatGPT ↗

What do you mean by "vibe coding"?+

Building by saying what you want in plain language while the AI writes and tests the code. You check the result and steer with short nudges — you "feel" your way there instead of planning everything up front.

The term comes from AI researcher Andrej Karpathy (2025) and was even the Collins Dictionary word of the year.

Source: Wikipedia: Vibe coding ↗

What does Claude still refuse, even with bypass permissions on?+

Even with bypass on, Claude still asks before truly irreversible actions: permanently deleting files, spending money, publishing or sending something on your behalf, or changing sensitive settings.

That's a built-in safety net. It can't be switched off — and that's exactly the point.

Where do Vercel and GitHub come in? I don't seem to use them.+

GitHub keeps every version of your code — a memory you never have to look at once it's set up. Vercel puts your project live online. Claude Code usually handles both in the background; that's why you "use" them without noticing.

See also the word-by-word explanations in In plain words above.

Can you give us 10 must-know prompts?+
  1. "Explain this as if I have no technical background."
  2. "What exactly are you going to do before you start?" — makes Claude show its plan
  3. "Do this fully yourself, don't ask me in between, unless it's irreversible."
  4. "Save this as a skill so it's done my way every time."
  5. "What does this error mean and how do we fix it?" — paste the whole error
  6. "Put this online so I can show it."
  7. "Turn this into an invoice/quote in my own style."
  8. "Write this in my tone, not like an AI text."
  9. "What does this cost if I run it like this?" — when unsure about costs
  10. "Warn me before you delete or send anything."
How do I generate leads with Claude — and how far does that go?+

Give Claude context about your existing customers — who they are, what they buy — the way you'd brief a new intern. Then it can think along about who to approach, write outreach texts, and build a list.

For answering incoming enquiries there's a ready-made starter skill below: Lead catcher. Sending messages automatically is possible too, but needs a connection to that service.

Examples: Social Media Examiner: Claude Code for everyone ↗

Will agents really make invoices and post on socials?+

Yes — but not by itself. You ask "make an invoice for client X" or "schedule this post", and Claude executes it through the tools connected to it. Entrepreneurs already use this for weekly reports, content calendars, and repurposing long content into individual posts.

Automatically replying to leads via Instagram is possible too, but needs a connection to the Instagram/Meta API — a next step, not a first lesson.

Examples: MindStudio: repurposing content with skills ↗ · Skills for repetitive marketing tasks ↗

Can I pull the text and learnings out of an Instagram video?+

Not by just pasting the link — Instagram doesn't hand the video content to other apps. What does work: save your own reel to your camera roll (three dots → save), or use a transcript tool, and give that file to Claude.

From there Claude extracts the key points, quotes and post text. That's exactly what the new starter skill Insta learnings below is for.

Background: Transcribing reels to text (2026) ↗

Can Claude handle my Google Ads, instead of an agency?+

For the thinking work: yes. Claude writes ad copy, reads your numbers, and tells you in plain language where your budget is leaking. Export your figures as a CSV from Google Ads and use the Ads report starter skill below for a weekly overview.

If you want Claude to look into your Ads account directly and make changes, that works via a connection (MCP) — a next step for anyone serious about this.

Examples: Automate Google Ads reporting in 30 min ↗ · 10 marketing workflows for Google & Meta Ads ↗

Lesson 1

The one-page cheat sheet

Print it · keep it next to you
Claude Code cheat sheet (English)
English version

English

Everything from this lesson, on one page — in English.

Download (PDF)
Claude Code spiekbrief (Nederlands)
Dutch version

Dutch

Everything from this lesson, on one page — in Dutch.

Download (PDF)
Lesson 2

The skills cheat sheet

Print it · keep it next to you
Claude Code Skills cheat sheet (English)
English version

English

Everything about skills from module 2, on one page — in English.

Download (PDF)
Claude Code Skills spiekbrief (Nederlands)
Dutch version

Dutch

Everything about skills from module 2, on one page — in Dutch.

Download (PDF)
Starter skills

Skills you can download

Drop these into Claude to see what's possible
Brand deck

A short intake — your fonts, colours, an old deck, image inspiration — then a slide deck in your own style.

Download ↓
ZZP finance

Plain-language grip on your money: what to set aside for tax, and a simple monthly bookkeeping overview.

Download ↓
Handover

Wrap up a project cleanly: what's done, what's left, and where everything lives. Works in Dutch and English.

Download ↓
Insta learnings

Give Claude the transcript or text of a video (reel, talk, podcast) and get back the key points, quotes and a ready-to-use post text.

Download ↓
Ads rapport

Export your Google Ads figures as a CSV and get a weekly report in plain language: what works, what leaks, and what to change this week.

Download ↓
Takeaways from previous workshops

Working with AI — field notes

Insights, not prompts
Play first, then purpose

The short version of what we've learned helping people use AI well. These are mindset tips, not prompts — read them once, then go build something.

A few truths

Where AI actually stands
InevitableAI is (probably) here to stay — it's being baked into Office, Google Docs, Canva, Miro. It'll be hard to escape.
Still acceleratingIt might "plateau" soon. I wouldn't bet on it.
Fear blocks skillMost people are afraid of it — and not very good at it. The fear prevents the play that builds the skill.
Raises the ceilingEveryone can build now. It's mostly an attitude.
New goalsAuthenticity, curiosity, intrinsic motivation. Experiment as much as you can.
How to work with it

The day-to-day habits
It's up to youAI augments you; it isn't autonomous. Bad results? Assume there's a better prompt.
Be specificVague questions get generic answers.
Play with itEffective use comes from intuition — and intuition comes from play.
Vibe, don't engineer"Prompt engineering" is too left-brained. Vibe with it.
Crap begets crapA lame reply poisons the ones after it. Edit your input until quality returns.
Quality breeds qualityPrime it with rich context before you ask the real question.
Start freshNew chat per task — quality degrades over a long thread. (This is what /clear is for.)
Brain onGive it your own ideas; it thinks better. And it can't finish the job — the last 10% is half the work.
Ask its planAsk what it's about to do before it does it, so you can steer.
In your work

Keeping your judgment in the loop
Think criticallyAssume anything it says could be false — it's a perspective, not the truth. Verify.
Be authenticTell it your real purpose. Use it to figure out what you actually want.
Have integrityKnow what you want to say. Don't cite what you haven't read.
Own the outcomeLike a camera — you didn't paint the pixels, but you framed it, and you own it.
Parts, not wholesAI makes parts, not finished things. Copy the good selections, not the whole.
Edit extensivelyLearn what "sounds like AI" means — and cut it.
Start earlyIt doesn't make you faster; it makes it easier to start.
Don't be lazySupplement your thinking, don't replace it. Raise your own standards.
Myths

Eight things people get wrong

Each one is worth questioning
The truth is usually more interesting than the fear

These are the common misconceptions. You don't have to agree — just don't take them on faith.

AI is not magic Using AI well needs a technical background First you must understand how it works You shouldn't treat AI like a person AI saves time but reduces quality AI is bad for the planet AI is bad for designers AI is bad for learning
Reality

What's annoying — and the 90/10 paradox
The 90/10 paradox

It can do 90% of the work — but the last 10% of a project is usually 90% of the work. Don't leave it all to the end.

Thinks it can't do something Talks too much Loses the plot Context rot Confident mistakes Sycophancy Slop Fluff Generic Makes you feel like you're getting work done
In their words

What people notice after trying it

"It was good to get started. You have to use your brain to finish it."

"It's great to put words to what you already thought."

"Use it as inspiration, not just an answer."

Where these come from

Field notes by J. Derek Lomas, from teaching people to work with AI. The full version, with prompts ↗  ·  dereklomas.me ↗

Learnings from earlier sessions

What teaching this keeps teaching us

Updated after each workshop

Every time we run this with a new group, a few things land harder than the rest. We keep the sharpest ones here — a short, growing record of what actually helps people get good with AI.

Milan · 2026

Polimi PSSD — AI for Social Transformation Design

A studio with Derek Lomas and master's design students, on using generative AI well.
  • Play first, then purpose. Intuition comes from play — let people mess around with it before handing them a goal.
  • Vibe, don't engineer. "Prompt engineering" is too left-brained. Steer it with short nudges and a feel for what's working.
  • The fear blocks the skill. Most people are afraid of AI and not very good at it — and the fear is exactly what prevents the play that builds the skill.
  • AI makes parts, not wholes. Copy the good selections, not the whole thing. You assemble; it supplies pieces.
  • Own the outcome. Like a camera — you didn't paint the pixels, but you framed it, and you own it.
  • The 90/10 paradox. It does 90% of the work, but the last 10% is usually 90% of the work. Don't leave it all to the end.
Read the full field notes from this session ↗