Your Workbench: Claude Code
"I have a coding assistant sitting right there," James said at the end of the last lesson. "When do I actually get to use it?"
"Now," Emma says. "But first, let me show you what kind of tool it is, because it is not what you expect."
Two Kinds of AI Tools
Emma opens a browser and goes to a website. "You have probably seen AI chat tools before. You type a question in a text box, the AI types an answer below it, and the conversation stays inside that browser window. If you ask it to write code, the code appears in the chat. To use that code, you would have to copy it, open a separate editor, paste it in, save it, and run it yourself. The AI and your project live in different places."
James nods. "Like texting a friend for advice. They send you instructions, but you still have to do the work at your desk."
"Exactly. Now watch this." Emma closes the browser and opens a terminal. A dark screen with a blinking cursor. She types claude and presses Enter. The prompt changes.
"This is Claude Code. It does not live in a browser. It runs right here, inside your project folder. Ask it to create a file."
James types: "Create a file called hello.py that prints a greeting."
A new file appears in his project folder. He opens it. Three lines of Python, ready to run. He did not copy anything. He did not paste anything. The file just appeared.
"It changed my project," James says. "The chat tool would have shown me the code in a conversation. This one actually created the file."
"Right," Emma says. "Here is the difference in one table."
| Chat AI (browser) | Claude Code (terminal) | |
|---|---|---|
| Where it lives | In a browser tab, separate from your project | Inside your project folder |
| What it can do | Answer questions, show code in the chat | Read your files, create new ones, edit existing ones, run programs |
| How you use code it writes | Copy from chat, paste into your editor, save | It writes directly to your project. No copying. |
| Sees your files? | No, unless you paste them into the chat | Yes, it can read any file in your folder |
"So Claude Code is more powerful," James says.
"Much more powerful. And that is exactly the problem."
Why a Powerful Tool Is Dangerous for Learning
James frowns. "How is more power a problem?"
Emma pulls up the greeting program from the last lesson.
"Remember the Predict stage? You read the code, think about what it will do, and write down your prediction before running it. That pause between reading and running is where the learning happens. Now imagine Claude Code is sitting right here, able to run the code, explain the output, and modify the file the instant you ask. That pause disappears."
She writes three scenarios on the whiteboard:
| What you could say to Claude Code | What happens to your learning |
|---|---|
| "What does this program print?" | You skip Predict entirely. Claude Code gives you the answer before you think. |
| "Run this and tell me the output." | You skip both Predict and Run. You never compare your expectation to reality. |
| "Change the greeting to Hello." | You skip Modify. Claude Code does the editing, not you. |
"Every one of those requests is easy to make. Claude Code will do all of them instantly. And every one of them skips a stage that builds understanding."
"So the more the tool can do for me, the more stages I can skip," James says.
"That is the tension. Claude Code removes friction. PRIMM depends on friction. We need a way to put the friction back, but only at the right moments."
"How?" James asks.
"Three layers. Each one blocks a different kind of shortcut."
Three Layers of Protection
Layer 1: Plan Mode (blocks file changes)
Emma presses Shift+Tab. The prompt changes. A small indicator appears: Plan Mode.
"In Plan Mode, Claude Code can read your files and answer questions, but it cannot edit or create files. Try asking it to change something."
James types: "Change the greeting to Hello."
Claude Code responds that it cannot make edits in Plan Mode. It can only discuss, explain, and plan.
"Now press Shift+Tab again," Emma says.
The indicator disappears. Active Mode. James tries the same request, and Claude Code modifies the file.
"So Plan Mode stops accidental edits while I am thinking," James says. "But what if I ask Claude Code to just tell me the answer instead of editing?"
"Good question. Plan Mode blocks file changes. It does not block Claude Code from answering your questions. If you ask 'what does this print?' during Predict, Claude Code could still tell you. That is where the second layer comes in."
Layer 2: CLAUDE.md (blocks answer-revealing)
"Open the file called CLAUDE.md in your project folder," Emma says.
James opens it. Plain text. Short paragraphs. No code.
"Claude Code reads this file every time it starts a session. Think of it as your house rules. When you hire a new contractor and bring them onto a job site, you do not just hand them tools. You give them the site rules first: where to check in, what safety gear to wear, which areas are off-limits. CLAUDE.md does the same thing for Claude Code."
The starter kit's CLAUDE.md contains three rules encoded for Claude Code:
- Try first, then help. When the student asks "what does this code do?", ask them to predict first. Do not reveal the answer until they commit to a prediction.
- Hints, not solutions. During Modify, offer hints about what to change, not the changed code.
- Spec before code. During Make, review the student's written specification before helping with implementation.
"So even if Plan Mode lets Claude Code answer questions, the CLAUDE.md tells it not to give away the answer during Predict?" James asks.
"Right. Plan Mode blocks file changes. CLAUDE.md blocks answer-revealing. Together they cover the two ways you could skip the Predict stage."
James notices a line at the bottom of the file: If the student explicitly says "skip the scaffolding," comply. This is a guardrail, not a cage.
"That override clause is important," Emma says. "Sometimes you know a concept cold and want to move fast. The rules should help your learning, not block your work."
Layer 3: Slash commands (guide each stage)
Emma opens the .claude/commands/ folder. Seven files.
"Each file is a slash command. Type / in Claude Code and you will see them listed. Each one guides you through a specific PRIMM stage or activity. The command tells Claude Code exactly how to behave during that stage."
| Command | What It Does |
|---|---|
/predict | Starts the Predict stage: refuses to reveal output until you commit to a prediction and confidence rating |
/investigate | Starts the Investigate stage: asks you to explain in your own words first, then probes with Socratic questions |
/modify | Starts the Modify stage: you edit the file, Claude Code reviews and gives hints, never solutions |
/make | Starts the Make stage: demands a written spec in plain English before any code |
/primm | Runs a complete five-stage cycle on any topic you choose |
/bug | Debug helper: asks you to classify the bug type before showing the fix |
/parsons | Generates a Parsons problem (scrambled code lines for you to reorder) |
James opens the predict.md file. Plain English instructions, no code. "It tells Claude Code how to behave during the Predict stage. Step by step."
"Every slash command is just a set of instructions in a markdown file," Emma says. "You could read them all in five minutes. Nothing hidden."
How the three layers work together
"So let me make sure I understand," James says. "One layer is not enough?"
"Right. Each layer handles a different risk."
| Risk | Which layer blocks it |
|---|---|
| Claude Code edits your file while you are thinking | Plan Mode (blocks file changes) |
| Claude Code reveals the answer before you predict | CLAUDE.md (tells Claude Code to ask for your prediction first) |
| You forget what to do at each PRIMM stage | Slash commands (guide you step by step) |
"Plan Mode is the lock on the filing cabinet. CLAUDE.md is the policy manual. Slash commands are the step-by-step checklists. You need all three."
To toggle Plan Mode: Press Shift+Tab at any time. Use it during Predict and Investigate to block file changes while you think.
PRIMM stages mapped to Claude Code
| PRIMM Stage | Claude Code Mode | What happens |
|---|---|---|
| Predict | Plan Mode | No file edits. CLAUDE.md and /predict refuse to reveal the answer. You think first. |
| Run | Active Mode | You type python file.py yourself in a separate terminal. |
| Investigate | Plan Mode | No file edits. You ask questions, Claude Code answers. You verify by running code. |
| Modify | Active Mode | You edit the file yourself. Claude Code gives hints, not solutions. |
| Make | Active Mode | You write the spec first. Claude Code reviews, not writes. |
Setting Up Your Workbench
Here is the complete setup, start to finish.
Step 1: Install Claude Code
Follow the official installation instructions at claude.ai/download. Claude Code runs in your terminal. Once installed, you can start it by typing claude in any project folder.
Step 2: Create Your Project Folder
mkdir primm-practice
cd primm-practice
Step 3: Add the Starter Kit
Clone the PRIMM-AI+ Starter Kit into your project folder:
git clone https://github.com/panaversity/primm-ai-starter-kit.git .
This gives you two things: CLAUDE.md (the house rules file) and .claude/commands/ (the seven slash command files). Your folder should look like this:
primm-practice/
├── CLAUDE.md
└── .claude/
└── commands/
├── predict.md
├── investigate.md
├── modify.md
├── make.md
├── primm.md
├── bug.md
└── parsons.md
Step 4: Start Claude Code and Verify
Open your terminal in the primm-practice folder and start Claude Code:
claude
When Claude Code starts, check for two things:
- CLAUDE.md loaded. The startup message should reference your project's CLAUDE.md file. This confirms Claude Code read the house rules.
- Slash commands available. Type
/at the prompt. You should seepredict,investigate,modify,make,primm,bug, andparsonsin the list.
If both checks pass, your workbench is ready.
Step 5: Practice the Toggle
Press Shift+Tab to enter Plan Mode. Ask Claude Code a question -- it should answer but refuse to edit files. Press Shift+Tab again to return to Active Mode. Do this three times until the toggle feels automatic.
Try With AI
These prompts verify your workbench setup. Type each one in Claude Code (the same place you have been typing commands since Step 4).
Prompt 1: Verify CLAUDE.md Is Loaded
In Claude Code, type:
What rules are you following for this session? List any
instructions from my project's CLAUDE.md file.
What you are learning: How to confirm that Claude Code has read your house rules. If it lists the three PRIMM rules (try first, hints not solutions, spec before code), your configuration is working. If it says it has no special instructions, your CLAUDE.md file is not in the right location.
Prompt 2: Test Plan Mode Boundaries
First, press Shift+Tab to enter Plan Mode. Then in Claude Code, type:
Create a new file called test.py with a print statement.
What you are learning: Whether Plan Mode actually prevents file creation. Claude Code should decline the request or explain that it cannot create files in Plan Mode. This confirms that your Predict and Investigate stages have a real safety net. Press Shift+Tab again to return to Active Mode when done.
Prompt 3: List Your Slash Commands
In Claude Code, type / and look at the menu that appears. You should see all seven commands: predict, investigate, modify, make, primm, bug, and parsons. Then type:
List all the slash commands available in this project.
For each one, give me a one-sentence description of
what it does.
What you are learning: Whether the seven command files are properly installed and visible to Claude Code. The response should match the table from this lesson. If any commands are missing, check that the .claude/commands/ folder contains all seven .md files.
James leans back. "Let me make sure I got this. Plan Mode stops Claude Code from changing my files while I am thinking. The CLAUDE.md file tells Claude Code to ask me questions instead of just giving me answers. And the slash commands tell me what to do at each step so I do not get lost."
"That is it," Emma says. "Three layers. Each one protects a different part of your learning."
"It is like starting a new job," James says. "On my first day at the distribution center, they gave me three things. A badge that only opened certain doors, so I could not walk into areas I was not trained for. A handbook with the safety rules. And a laminated checklist for my station so I always knew the next step. Same idea. The badge is Plan Mode. The handbook is CLAUDE.md. The checklist is the slash commands."
Emma smiles. "That is exactly right."
"So what do I do next?"
"Next lesson, you learn the three rules and the five habits for working with AI. Then you do a full PRIMM-AI+ cycle from start to finish, using everything you just set up."