Ask a model about code it cannot see and it will still answer you, fluently and confidently, about the wrong file. Rephrasing the question does not fix that. Telling it what to look at does.

How this course works

You build Healthyfied Recipes. You paste in a recipe URL, the app fetches that page, and it matches the ingredients against a swap table you wrote by hand: butter becomes olive oil, white rice becomes cauliflower rice. It swaps the ones it knows, and the table is short on purpose. The only thing it ever fetches is the page you paste: no model call, no API key, no account to sign up for. There is no AI inside the finished app at all. It only shows up while you build. Before you start, confirm that your Anthropic plan includes Claude Code access.

The week is shaped concepts-first: Days 1 to 3 install the tools and explain how the machine works, then Days 4 to 7 build the app with it. I am stubborn about that order.

Every day runs the same five steps, and by Day 7 you should be running them without checking the list:

SPEC → CONTEXT → DELEGATE → VERIFY → CAPTURE

SpecWrite down what you want before you ask for it. Plain sentences, and say what “done” looks like.
ContextDecide what the model gets to see. It is the step people skip and the one this course is named after.
DelegateLet the agent do the typing. You decide what gets built and whether the result is right.
VerifyCheck something that can come back false. Does it build, and did you open the screen and look at it?
CaptureWrite down what you learned, so you are not re-explaining the same thing to it on Thursday.

Every day you do something with your hands, then read one or two official pages to see whether you did it right. Every day also ends with a ship check, which is just something concrete you can show another person. Running alongside all week: Anthropic Academy’s Claude Code 101, and the tutorial that ships inside the tool. Type /powerup in Claude Code and it walks you through 18 short interactive lessons.

One rule sits underneath all seven days: read the current documentation instead of trusting what you remember, and instead of trusting what the model remembers. Your memory and the model’s both stop somewhere, and neither of them tells you where. Every day here links the official page for whatever it teaches, and opening that link is part of the exercise rather than a footnote to it.

Week 1 stays on one agent and one project, running the same loop every day until it is dull. Week 2 stays on the same tool and goes deeper: subagents, hooks, MCP, automation.

The 7 days

Day 1: Install the toolsXcode, a GitHub account, and Claude Code running in your terminal, then a blank SwiftUI app booting in the iOS Simulator. You make no AI edits today. Ships: three working tools and a clean claude doctor. Start Day 1 →
Day 2: How models workWhy the output is probabilistic and not deterministic, and why tokens are the meter on every request. Your first conversation changes nothing: you ask Claude Code to explain your own project, and that is the whole exercise. Ships: a NOTES.md saying in your own words what a model does.
Day 3: Context, tools, agentsContext windows, why they get worse as they fill, and tool calling, which is the whole trick behind an agent: a model using tools in a loop. You poke around Pulse, a real open-source journaling app, and write your spec. Reading: Effective context engineering for AI agents and Building effective agents. Ships: PRODUCT.md for Healthyfied Recipes, plus a one-page tour of Pulse.
Day 4: First delegationPublish the repository to GitHub, learn @-file references, and spend one well-prepared prompt on the recipe-URL home screen. Then read the diff line by line before you accept it. Ships: a public repo and a screen Claude built and you reviewed.
Day 5: Persistent contextCLAUDE.md: generate a first draft with /init, then cut it by hand until every line earns its place. You also build the offline swap engine, the Swift lookup table behind the substitutions at the top of this page, and decide what it should say when an ingredient has no good swap. Ships: a CLAUDE.md that shortens every later prompt, and working swaps.
Day 6: Plan, build, verifyPlan mode (Shift+Tab): Claude researches and proposes, and touches no file until you approve. Then the full flow, plus Swift Testing over the engine. Ships: recipe in, healthier swaps out, with a green test suite behind it.
Day 7: The context stack and the storyYour first Agent Skill in .claude/skills, a LEARNINGS.md with real entries in it, a README with screenshots, and the 60-second version of the story for when someone asks in an interview. Ships: a repo you would happily show people.
Next up: Week 2, agents at scaleSubagents, hooks, MCP, and Claude Desktop’s iOS Simulator pane. Same loop, much more surface area, once one agent has stopped taking your full attention.

Day 1 is published in full below. Days 2 through 7 publish one a day for the rest of the week, so get Day 2 in your inbox the morning it lands.

Day 1: Install the tools

Budget 45 to 60 minutes of hands-on time, plus download time. Xcode is a big download, so start it before you read the rest of this page.

Today you make no AI edits and you publish nothing. Day 1 is install, verify, look around. The machine gets explained on Day 2 and you start delegating on Day 4. That order is deliberate. It keeps you from asking before you know what you are asking.

Before you start

You need a Mac that can run a current version of Xcode (Apple silicon recommended), about 30 GB of free disk space, and an email address.

Claude Code access is required. Anthropic’s setup documentation shows you how to confirm access and install the tool. Confirm access before you begin. The app you build never calls a model API.

Confirm that you can open Claude Code before you begin the later lessons. Day 1 does not use the model, but the rest of the week does.

Usage limits vary by plan and workload. Writing the spec first keeps your requests focused and helps you avoid sending large files without a clear reason.

Step 1: Install Xcode

Open the Mac App Store, search for Xcode, and press Get. It asks you to sign in with an Apple ID. Create one if you need it. The download is big, so start it now and keep reading.

On first launch, accept the license agreement and let Xcode install the iOS platform when it offers to. You do not need an Apple Developer membership for this course because everything this week runs in the Simulator.

Verify: Xcode opens to the Welcome window.

Step 2: Create your GitHub account

Go to github.com/signup and create an account. Pick a username you would be happy to put on a résumé rather than a joke handle, because this account is about to become your portfolio.

That matters more than it sounds like it does. From Day 4 onward, every day of this course ends with a push. By Day 7 the repository holds both your code and the evidence that you built it. The commit messages show the order and your reasoning. When someone asks what you can do, you send them the repo.

The account is all you need today. You publish this project on Day 4, once there is something worth pushing.

Step 3: Confirm access and install Claude Code

Sign in at claude.ai, then use Anthropic’s setup documentation to confirm that your plan includes Claude Code. Access is required for this step.

Then install the tool. If you have never opened a terminal in your life, Anthropic wrote a guide for exactly that person. It explains the window before it asks you to type anything into it.

Do not paste commands you do not understand. This command downloads Anthropic’s installer from claude.ai and runs it. It is the same command shown in the setup documentation linked above, which you can check before running it.

Open Terminal (press ⌘Space, type Terminal, press Return), paste this line, and press Return:

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

That installs a native binary. The installer needs no prior setup or Node.js, and it keeps itself up to date.

Verify, in the same window: run claude --version, which prints a version number, then claude doctor, which health-checks the installation and should come back clean. If something looks off, the full setup reference is where I would look first.

One aside. There is also a Claude Desktop app running the same engine with no terminal in sight. This course uses the terminal because that is where the complete feature set lives, and because the habits you pick up there transfer to every other agent you will ever touch.

Step 4: Create the project in Xcode

In Xcode, choose File → New → Project, pick the iOS tab, and select App.

Set Product Name to Healthyfied. Xcode product names cannot contain dashes; the repository gets the full name on Day 4. Set Interface to SwiftUI and Language to Swift. When you save, tick Create Git repository on my Mac and name the project folder healthyfied-recipes-ios.

Now press ⌘R. The iPhone Simulator boots and shows “Hello, world!”.

That is the whole app for now, and that is fine. What you have is a project that builds and launches. Getting there before you add anything is what keeps the next six days debuggable, because from here on, anything that breaks is something you just did.

Step 5: Start Claude Code once, then just read

Back in Terminal, move into your project folder: type cd followed by a space, drag the healthyfied-recipes-ios folder from Finder onto the Terminal window (that pastes the path in for you), and press Return. Then start the agent:

claude

The first run sends you to your browser to sign in. Do that, come back to the terminal, and then quit without asking it anything. Press Ctrl+C, or type /exit.

This leaves the project unchanged. Claude Code builds no index of your project. When you ask it something, it explores your code live with search tools, reading and grepping its way to an answer the way a new teammate would on their first day. Nothing has to be indexed first, so it is useful the moment it starts. It also means the quality of an answer depends mostly on what you point it at, which is what this course is about.

So today you point it at nothing. Open ContentView.swift in Xcode and read it line by line instead. Expand the project navigator and look at what Xcode generated for you. Change nothing.

You are about to spend a week directing an agent inside this project, and you cannot judge its work in files you have never read.

What today was about

1. What Claude Code is. An agent that lives in your terminal. It reads files, edits them, and runs commands. By default it asks before it edits a file or runs a command, and you decide how wide that permission gets. This week you use it read-only for three days before you let it write a single line. Claude Code quickstart →

2. Why the concepts come first. Claude Code is a probabilistic machine, not a compiler that happens to speak English. Understand it, then delegate to it, which is why nothing you installed today touches your code until Day 4. Alongside the week: Claude Code 101 from Anthropic Academy, plus the /powerup tutorial inside the tool itself, 18 short interactive lessons with no browser involved. Claude Code 101 →

Ship check: Day 1

The blank app runningXcode boots Healthyfied in the iOS Simulator. Screenshot it. The “before” picture is worth having on Day 7.
A GitHub accountCreated, under a name you would put on a résumé. Nothing pushed yet, that is Day 4.
Claude Code installedclaude --version prints a version, and claude doctor comes back clean.
Signed in, from the projectYou started Claude Code once from inside healthyfied-recipes-ios and finished the browser sign-in.

That is Day 1. Tomorrow, how the machine works.

Get Day 2 when it lands

The tools are running and you have not delegated anything yet. Day 2 explains the machine before you point it at your project. Leave your email and each day shows up the morning it publishes.

Six emails at most, one a day for the rest of the week. I don’t send anything else and I don’t pass your address on.

A note on sources

Every link here goes to a current official Anthropic property: code.claude.com/docs for Claude Code, platform.claude.com/docs for models and API concepts, anthropic.com/engineering for the research team’s essays, and anthropic.skilljar.com for Academy courses.

That narrowness is on purpose. Anthropic’s documentation moved domains recently and several features got renamed along the way, so a lot of the tutorials you will find elsewhere point at pages that no longer resolve and commands that no longer exist. Follow one of those and you go hunting for something that is not there, then decide you did it wrong. If anything here contradicts the official page, the official page is right and this page is stale.

One thing to know about and then set aside: Claude Desktop has an official iOS Simulator pane in public beta that streams your running app beside the conversation while Claude works. It shows up later in this course, not on Day 1.