Skip to content

Claude Code

Claude Code is Anthropic's command-line AI coding agent. Like Gemini CLI, it lives in your terminal — but it's designed for deeper, multi-step work. It can:

  • Read and understand your entire codebase
  • Write, edit, and refactor code across multiple files
  • Run commands and fix errors iteratively
  • Work with Git (commits, branches, pull requests)
  • Search the web and your local files

The key difference from Gemini CLI: Claude Code can chain multiple steps together autonomously. You give it a goal, and it figures out the sequence of actions needed to get there.

Account required

Claude Code requires a Claude Pro, Max, or API account — there is no free tier.

How it handles your files

Claude Code will ask for your approval before taking actions that modify files or run commands — unless you tell it to proceed automatically. You are always in control.

Prerequisites

  • Node.js 20+ — check with:

    node --version

    If you completed the Gemini CLI module, you already have this. If not, install Node 20+ from nodejs.org.

  • A Claude account — one of:

Cost

Claude Pro ($20/mo) or Max ($100/mo) subscriptions give you unlimited Claude Code usage. An API key is pay-as-you-go — typically a few cents per session for light use. For a masters project, a Pro subscription is the most practical option.

Installation

npm install -g @anthropic-ai/claude-code

Open Anaconda Prompt (not PowerShell or Command Prompt) and run:

npm install -g @anthropic-ai/claude-code

All subsequent claude commands should also be run from Anaconda Prompt.

Verify the installation

claude --version

You should see a version number printed. If you get command not found, see Troubleshooting below.

Authentication

Run Claude Code for the first time:

claude

You'll be prompted to choose an authentication method:

? How would you like to authenticate?
  ❯ Claude.ai account (Pro or Max subscription)
    Anthropic API key

Option 1 — Claude.ai account (recommended for most students): Select this and a browser window will open. Log into your Claude account and authorize the CLI. You'll be redirected back and the terminal will confirm you're authenticated.

Option 2 — API key: Select this and paste your key from console.anthropic.com when prompted. The key is stored locally and never sent anywhere other than Anthropic's servers.

Check your setup

After authenticating, run:

claude doctor

This checks your installation and reports any issues. A healthy output looks like:

✓ Node.js version: v20.x.x
✓ Claude Code version: x.x.x
✓ Authentication: OK
✓ Network connectivity: OK

How a Session Works

Navigate to any project directory and start Claude Code:

cd ~/your-project
claude

Claude Code reads the contents of your current directory automatically. You'll see:

✓ Loaded project context
>

From here you can type questions or tasks. When you ask it to modify something, it shows you what it plans to do and asks for approval before touching any files. You press y to approve, n to cancel, or type feedback to adjust.

You can also skip the interactive session entirely and give a one-shot command from the terminal:

claude "explain what this project does"
claude "add a --verbose flag to the CLI in main.py"

Exercise 1 — Explore Your Project

Try this now

  1. Navigate to any project you have locally (your capstone project, a class assignment, anything with code in it)
  2. Start Claude Code: claude
  3. Ask it: "Read this project and give me a summary of what it does, what the main files are, and what's still incomplete or unclear"
  4. Follow up with one or two questions based on its response

Goal: Get comfortable with the read-explain-ask loop before you ask it to change anything.

Start with questions, not edits

Before asking Claude Code to change anything, ask it to explain what's there. This builds your confidence in what it understands — and catches any misunderstandings early.

Exercise 2 — Make a Change

Try this now

  1. In the same project, ask Claude Code to make a small change:
    • Add a docstring to a function that doesn't have one
    • Rename a variable to be more descriptive
    • Add a comment explaining a confusing section of code
  2. Review the proposed edit before approving — notice it shows you the diff before touching your files
  3. After it applies the change, open the file and confirm it looks right

Goal: Experience the propose-review-approve workflow.

Slash Commands

Claude Code has built-in slash commands that control your session. Type / inside an interactive session to see the full list. Here are the most useful ones to know:

Command What it does
/help Show all available commands and keyboard shortcuts
/model Switch between models (Opus, Sonnet, Haiku) — trade capability for speed
/context See and manage what files Claude has loaded into its context
/usage Check your token usage and remaining quota
/compact Compress the conversation history to free up context window space
/clear Wipe the conversation and start fresh (does not undo file changes)
/commit Stage and commit your changes with an auto-generated commit message

A few keyboard shortcuts are also worth knowing:

Shortcut What it does
Esc Cancel a running action (useful when Claude is going down the wrong path)
Shift+Tab Enter multi-line input mode (for pasting code or writing longer prompts)

Try this now

  1. In your Claude Code session, run /model and note what models are available to you
  2. Run /usage to see your current token consumption
  3. Run /context to see what files Claude has loaded from your project

These are useful for managing long sessions. /compact is especially helpful when Claude starts losing track of earlier context, and /model lets you drop to a faster model for simple tasks.

Exercise 3 — Build Something with Your Data

This is where your capstone project enters the picture. You should have data, documents, or other materials from your project already on your machine. If you don't have capstone materials handy, any dataset will work — a CSV from a class assignment, a public dataset you've downloaded, anything.

Try this now

  1. Navigate to a directory containing data you're working with and start Claude Code
  2. Ask it to read your data and describe what it contains: 
    > Read the data files in this project and tell me what's there —
  columns, row counts, data types, and anything that stands out
  3. Then ask it to create a visualization: 
    > Using that data, create a Python script that generates a
  meaningful plot. Pick the most interesting relationship you
  see and save the figure as output.png
  4. Review the script it produces, approve it, and let it run
  5. If the plot isn't what you want, give feedback and let it iterate: 
    > Change this to a scatter plot instead, and increase the font
  size on the axis labels

Goal: Experience the full loop — data exploration, code generation, execution, and iteration — using your own project materials. This is the core workflow you'll use going forward.

Bring context, get better results

The more you tell Claude Code about what you're trying to do, the better the output. Instead of "make a plot," try "I'm investigating whether temperature affects yield in my experiment data — show me that relationship." Domain context makes a real difference.

A Workflow to Try: Structured Capstone Sessions

One way to get more out of Claude Code on a long-running project is to give it persistent context so you don't re-explain everything each session. This is optional — use what's helpful and skip what isn't.

CLAUDE.md — persistent project context

Claude Code automatically reads a file called CLAUDE.md at the root of your project at the start of every session. Think of it as onboarding instructions for a research assistant who's joining your project.

Create a CLAUDE.md in your project root with at minimum:

# CLAUDE.md

## Project
One sentence: what this project does and what the deliverable is.

## Setup
conda activate my-env  # or pip install -r requirements.txt

## Code Layout
src/          # core library code
scripts/      # standalone analysis and figure scripts
data/raw/     # raw input data (never modified)
data/processed/  # cleaned datasets

## Ask Before
- Deleting files
- Adding new dependencies
- Changing directory structure
- Modifying data files

The more specific and current this file is, the more useful Claude Code will be. Update it as your project evolves.

plan.md and todo.md as a session interface

Some students find it useful to maintain two lightweight documents alongside their code:

  1. docs/plan.md — a strategic overview: phases, milestones, approach, current status
  2. docs/todo.md — a checkbox task list seeded from the plan, one concrete action per line

Then at the start of each session:

> Read docs/todo.md and tell me what the next unchecked task is
> [Claude summarizes the next task]
> Go ahead and do that

This can save you from re-explaining your project each session — the agent picks up where you left off. You can also ask Claude Code to generate these files for you based on your existing project materials.

Example directory structure

Here's one way to organize a capstone project for this workflow. You don't need all of these — adopt the pieces that make sense for your project:

project/
  CLAUDE.md              # Agent instructions
  environment.yml        # Pinned dependencies

  docs/
    plan.md              # Milestones and strategy
    todo.md              # Living task checklist

  src/                   # Library code
  scripts/               # Standalone scripts (analysis, figures)
  data/
    raw/                 # Never modified — read only
    processed/           # Output of your pipeline
  figures/               # Generated outputs
  paper/                 # Your written deliverable

Example prompts for capstone work

> Read CLAUDE.md and docs/todo.md, then do the next unchecked task

> My data pipeline in scripts/preprocess.py is slow on large files —
  profile it and suggest optimizations

> Help me write the methods section for this analysis based on what
  you see in the code

> I just finished the feature engineering step — update docs/todo.md
  to check that off and tell me what's next

Let the agent update your docs

After completing a task, you can ask Claude Code to check off the item in todo.md and update any project docs that changed. This keeps your project state current with minimal bookkeeping.

Troubleshooting

claude: command not found after installation

The npm global bin directory may not be on your PATH. Add it:

echo 'export PATH="$(npm bin -g):$PATH"' >> ~/.bashrc
source ~/.bashrc

echo 'export PATH="$(npm bin -g):$PATH"' >> ~/.zshrc
source ~/.zshrc

Make sure you installed from Anaconda Prompt and are running claude from the same terminal. If it still fails, close and reopen Anaconda Prompt and try again.

Then verify: which claude (macOS/Linux) or where claude (Windows)

Authentication fails or browser doesn't open

Run claude doctor to diagnose. Common fixes:

  • Make sure you're logged into claude.ai in your browser before authenticating
  • Try the API key option instead of the browser OAuth flow
  • Check that your account type is Pro, Max, or has API credits
Claude Code modifies files without asking

By default, Claude Code asks before editing files. If it's not asking, check your session settings with /settings inside an interactive session. You can also add explicit instructions to your CLAUDE.md:

Always ask for approval before editing any file.
Can I use Claude Code for free?

There is no free tier. You need either a Claude Pro ($20/mo) or Max ($100/mo) subscription, or an Anthropic API key with credits. API usage for light sessions typically costs a few cents per session — check anthropic.com/pricing for current rates.

Claude Code is slow or says I've hit a usage limit

Pro plan subscribers have a daily usage cap. During heavy sessions — especially with the Opus model — you can hit this limit and get throttled or temporarily blocked. A few ways to manage this:

  • Use /usage regularly to check where you stand
  • Switch to a lighter model with /model (Sonnet or Haiku) for simple tasks like explaining code or writing comments
  • Use /compact to compress your conversation — shorter context means fewer tokens per request
  • If you hit the cap, take a break and come back later — it resets daily

Max plan subscribers have a significantly higher cap. If you're using an API key, you're billed per token with no hard cap, but costs can add up in long sessions.

It keeps misunderstanding my project

This usually means your CLAUDE.md needs more context. Add:

  • A clearer one-sentence project description
  • The entry points (which file to run, which function does X)
  • Any conventions or constraints the agent should follow

Think of CLAUDE.md as onboarding instructions for a new research assistant.

Back to: Home | Previous: Gemini CLI →