Claude Code for Beginners: Every Concept Explained in Plain English (2026)

What is Claude Code, and how is it different from ChatGPT?

A chatbot talks; an agent acts. Ask ChatGPT for a website and you get a description of one. Ask Claude Code and it creates the files, writes the code, runs it, finds its own errors, and fixes them — on your computer, in a folder you can see. That single difference — taking actions instead of producing answers — is what every concept below exists to manage. See also our plain-language explainer on what an AI agent harness is.

Where it lives: in a terminal (the text window that scares everyone at first) or, friendlier, in the official VS Code extension — VS Code is free, and the extension gives Claude Code a proper visual interface. A project is just a folder: your files, its work, all in one place, like a Google Drive folder for one app or client.

Prompts are how you talk to it, and specificity is everything. "Build me a website" produces generic filler; "build a one-page site for a wedding-DJ business in Toronto, lead-generation focused, with a booking form" produces something usable. Give it a role ("you're a web designer"), name the outcome, and mention what you don't want.

How it thinks: context windows, tokens, and models

The context window is the agent's short-term memory: everything in the current session — your messages, its replies, the files it read — has to fit inside it. Long sessions fill it up; a nearly-full window means degraded answers and pricier replies, because the whole conversation is re-processed on every turn. (Deeper dive: our context-window glossary entry.)

Tokens are the units all of this is measured in. Anthropic's own rule of thumb: one token is roughly 4 characters, or about three-quarters of an English word (platform.claude.com docs, 2026-07-15). Your plan's limits are token budgets — which is why a two-hour conversation burns them faster than ten short ones: each new message carries the entire session behind it.

Claude Code ships several models you can switch between with /model: Opus 4.8 for the hardest work, Sonnet 5 as the balanced default tier, Haiku 4.5 when speed and thrift matter — plus Fable 5, which Anthropic calls the most capable model in Claude Code for long-running tasks; it's not the default, you select it deliberately. Extended thinking is a toggle that lets any of them reason longer before answering — slower, better on genuinely hard problems.

How you teach it: CLAUDE.md, memory, and voice files

CLAUDE.md is the employee handbook: a plain markdown file of standing instructions loaded at the start of every session. Project rules live in ./CLAUDE.md, your personal preferences in ~/.claude/CLAUDE.md. "Never touch the /seed folder", "always run the tests before saying you're done" — written once, obeyed every session. We keep a full guide with a real-world example file.

Memory is the other half: notes Claude writes for itself that persist across sessions — ask it to remember something and it files the note away, then recalls it next time (the docs call this auto memory). Without it, every conversation starts from zero; with it, the agent accumulates a working knowledge of you and your project. More in our guide to Claude Code memory.

Two smaller teaching tools: type @ to point at an exact file (@tone.md) instead of hoping it finds the right one, and voice files — paste samples of your own writing and have Claude extract your tone, vocabulary, and humor into reference files it consults whenever it writes for you. That's the difference between posts that sound like a model and posts that sound like you.

How you let it act: permissions, tools, hooks, and MCP

Permission modes decide how much the agent may do without asking. The ladder runs from Plan mode (think and propose, touch nothing) through Manual (the ask-first default) and Accept edits (edit files freely, ask for everything else) to Auto and finally Bypass permissions — fastest, and the one that can overwrite something you wanted. Cycle with Shift+Tab; start big tasks in Plan mode, then loosen (mode names per code.claude.com/docs, 2026-07-15).

Tools are the agent's hands — the built-in abilities it uses to act: read a file, search the project, edit, write, run a terminal command, fetch a web page. You don't configure these; you'll just see them named in the transcript as it works, and knowing the names helps you follow what it's doing.

Hooks are automations bolted to fixed moments in the agent's lifecycle: before a tool runs, after it stops, when a session starts. Play a sound when it finishes; block any command that would expose a secret. You describe the rule in plain English and Claude writes the hook for you.

MCP — Model Context Protocol — is the standard connector between the agent and your other apps: Gmail, Notion, Slack, databases (glossary). Browse ready-made connectors in the Anthropic directory at claude.ai/directory, or add servers with the claude mcp command. Where no MCP server exists, the fallback is the older way apps have always talked: an API with a key — which lives in a .env file, the one folder-file whose job is holding secrets that never get shared or committed.

How you scale it: skills, plugins, subagents, and checkpoints

Skills are your repeatable workflows saved as files: anything you do twice — a LinkedIn-post routine, a weekly report — gets written into a SKILL.md and invoked as a slash command (/linkedin). Plugins are skills, agents, and hooks other people already built, installed in one command via /plugin from marketplaces, including ones Anthropic runs itself. Check what exists before building your own; our tools roundup covers the repos worth installing.

Subagents are specialists the main agent can hire. The restaurant analogy holds up: a skill is a chef cooking at your table — you watch every step; a subagent is the kitchen — the order goes in, the dish comes out, and the mess stays out of your conversation's context. Agent teams take it further: a manager agent splits a five-page site among three workers building in parallel instead of one agent grinding sequentially.

Checkpoints are the undo button. Claude Code snapshots your files before each prompt; /rewind (or double-tap Escape) restores the code, the conversation, or both. One documented gap: changes made by shell commands aren't tracked — another reason git matters.

Which brings us to the last two: GitHub is offsite backup for code (and the road to deploying — hosts like Vercel publish straight from it), and scheduled tasks let the agent run without you — sort the inbox every 15 minutes, post the daily update at 9. That's the point where Claude Code stops being a tool you use and starts being staff you manage.