Staring at a black terminal screen with a blinking cursor, wondering how this is supposed to make life easier? That feeling is universal. Terms like “context window,” “sub-agents,” and “MCP servers” sound like they were invented for someone who already knows everything. Here is the refreshing truth: they were not. This post walks through all 27 essential concepts of Claude Code, starting from the absolute basics and building to advanced workflows — in plain English, with the exact commands needed at each step.

Part 1: The Foundation

What Is Claude Code?

Regular chatbots — ChatGPT, Claude, Gemini — can only talk. They give advice. They cannot actually do anything on a computer. Claude Code is fundamentally different. It creates files, builds websites, sets up databases, installs packages, runs tests, and automates repetitive tasks. The user describes what they want in plain English and Claude Code takes action.

The Terminal — Do Not Be Intimidated

Claude Code does not run in a browser. It runs in the terminal — that black screen with white text that looks cryptic at first glance. Think of it simply as another way to talk to a computer. Instead of clicking buttons, commands are typed. Only a handful of commands are ever needed to get started.

That is genuinely the starter kit. No terminal expertise required.

Prompts — How to Talk to Claude

A prompt is simply what gets typed to tell Claude what to do — same as any AI chat window. No special syntax, no code, just plain English. The golden rule is to be specific. Vague prompts get vague results.

Better prompts produce better results. That is the whole game.

Permissions — Safety vs. Speed

Claude Code can make actual changes to a computer — create files, edit code, run commands, even delete things. By default it asks permission before doing anything significant. When starting out, keeping full permissions enabled is sensible so every proposed action is visible. As confidence grows, safe actions can be pre-approved so Claude works faster.

Tool Use — How Claude Takes Action

Regular Claude can only talk. Claude Code has three built-in tools that let it interact with a computer. There is no need to specify which tool to use — just describe the goal and Claude picks the right one automatically.

The Context Window — Claude’s Short-Term Memory

The context window is everything Claude can see and think about at any given moment — every message sent, every file it has read, every response it has given. It has a limit. When a conversation grows too long, Claude starts losing track of earlier information. This is called context rot and it produces the same kind of unreliable answers that AI users refer to as hallucinations. A green bar at the bottom of the terminal shows how full the context window is.

Conversation History — Never Lose Your Place

Claude Code automatically saves every session. Close the laptop, come back the next day, and pick up exactly where things were left off. The most-used command for returning to work: claude --resume. This drops back into the most recent conversation with all the same context intact. When that command is run, a list of all previous sessions appears — scroll through, see what each one covered, and choose.

Token Usage — How the Cost Is Calculated

Claude Code costs money. The cost is calculated in tokens, where roughly three-quarters of a word equals one token. Every prompt, every response, every file Claude reads — it all uses tokens. The model chosen determines the cost per token.

Monthly subscription (Pro or Max plan) beats pay-as-you-go for most users. Mid-build, worrying about token costs kills focus. A predictable fixed monthly cost removes that friction entirely. Use /cost or /stats to track usage in any session.

Claude Code runs in the terminal — a straightforward interface once the basics are understood.

Part 2: Making Claude Work Your Way

CLAUDE.md — The Personal Instruction Manual

Without guidance, Claude guesses. Guessing leads to results that need correcting. CLAUDE.md is a plain text file created in the project folder. It contains personal preferences, coding standards, tone requirements, and anything else that matters. Claude reads it before every single conversation in that project.

Write it once. Claude follows it forever. This is possibly the most important file in any project. Create one with /init and Claude will help set it up.

Memory — Claude Learns Over Time

CLAUDE.md is manual memory. Claude Code also has automatic memory — it stores preferences and facts across sessions without being asked. Preferred languages, coding styles, project conventions — Claude remembers these things without being told every time. Memory is managed conversationally.

Deny List — Files Claude Should Never Touch

Not every file should be accessible to Claude. Some are massive and waste context. Some contain sensitive data like API keys and passwords. Some are simply irrelevant. A deny list in .claude/settings.json puts specific files and folders completely off limits. Claude will not discover them, will not include them in search results, and cannot read them even if directly asked. To set one up, just ask: “Add a deny rule so you cannot read my .env file or anything in my secrets folder.”

Flags — Launch Settings for Each Session

Flags are options added when launching Claude Code. They customise behaviour for that specific session before it even starts. They are entirely optional but useful as shortcuts for common setups.

Example: claude --model opus --verbose

Compact Context — Managing Long Conversations

When conversations grow very long, the context window fills up and Claude starts forgetting earlier information. Claude handles this automatically at 85–95% capacity by summarising key information and clearing out the noise. When Claude seems confused mid-session, /compact can be triggered manually. Specific information to preserve can be named: /compact keep info about API calls, keep the database schema. Use compact proactively at natural break points in long sessions — it keeps Claude sharper throughout.

Models — Which AI Is Actually Running

When people say “Claude Code” they are talking about a family of AI models underneath. Each has different strengths, speeds, and costs. Use Sonnet for most things. Switch to Opus when the problem genuinely demands deep reasoning. Use Haiku for very simple tasks where speed and cost matter more than depth. Switch models mid-conversation with /model.

Part 3: Power Moves

Extended Thinking — Claude Reasons Before Acting

Some tasks are simple. Some genuinely require working through the problem before taking action. Extended thinking gives Claude a dedicated “thinking budget” in tokens to reason step by step before doing anything. The good news: there is nothing to configure. It is on by default with a maximum budget for thinking tokens. Just ask, and Claude thinks deeply before acting. Previously this required typing “think hard” or “ultra thinking” — now it is automatic.

Slash Commands — Keyboard Shortcuts for Claude

Slash commands are shortcuts for repetitive actions. Type / followed by the command name. Custom slash commands can also be created by placing a file in the .claude/slash-commands folder — any daily task can become a one-word shortcut.

Skills — Expert Playbooks for Specific Tasks

Claude Code is a generalist — it can do many things adequately. Skills are pre-written instruction sets that teach it to do specific tasks exceptionally well: creating social media strategies, building presentations, writing polished copy, designing front-end interfaces. When a request matches a skill description, Claude reads that skill file and supporting files to understand best practices before building. Output quality improves dramatically. Skills use more context because of the files loaded, but the trade-off is almost always worth it. Skills can be found pre-built online or created from scratch based on personal expertise.

Hooks — Automated Background Rules

Sometimes specific rules should run automatically at certain moments without consuming AI tokens. Hooks are custom scripts that trigger at specific points in the workflow — completely in the background, completely free of token cost. Examples include automatically formatting every file Claude saves, or logging every command Claude runs. They are guardrails that work silently. To set one up, ask directly: “Set up a hook that auto-formats every file you save.”

MCP Servers — Connecting to External Tools

By default, Claude Code works with files on the local computer. But real work does not live only on a hard drive — it lives across Airtable, Notion, Asana, Slack, and dozens of other tools. MCP (Model Context Protocol) servers connect Claude Code to those external services. Once connected, Claude can pull data from those tools, push updates to them, and interact with an entire tech stack from a single conversation window. It takes actions not just on the desktop but inside the apps used every day.

MCP servers extend Claude Code’s reach beyond local files into every tool in the workflow.

Sub-Agents — Specialists with Their Own Context

Long sessions cause context pile-up. Even with /compact, a single sprawling conversation eventually becomes a burden. Sub-agents solve this by running in their own separate context window with their own system prompt and tool access, while still following the main CLAUDE.md file. The main Claude instance hands off a task, the sub-agent works independently, then reports back with the result.

Claude will automatically spin out sub-agents whenever the architecture makes sense. There is no need to instruct it to do so — it happens in the background.

Agent Teams — Sub-Agents That Talk to Each Other

Sub-agents have a limitation: they can only report back to the main agent. With five or more specialists, the main agent becomes a bottleneck — a researcher cannot share findings directly with a writer, for example. Agent teams (a newer feature) solve this. Instead of a hub-and-spoke model, teammates communicate directly with each other and share a task list. Building a SaaS application with an API developer, a front-end developer, and a tester all working in parallel — each aware of what the others are doing — becomes genuinely possible.

Multimodal Support — Screenshots and Images

Describing a visual bug in text takes far longer than simply showing it. Claude Code accepts screenshots pasted directly into the terminal — exactly like any other AI chat. Screenshot a bug and paste it in; Claude sees exactly what is wrong. The reverse works too: paste in a design reference and ask Claude to build to match it. The result will not be pixel-perfect, but it typically reaches 60–70% of the way there on the first attempt — far faster than describing the design in words. Images use context, but the time saved over written description is significant.

Checkpoints and Rewind — Undoing Mistakes

Claude does not get everything right. When working autonomously, it makes changes quickly and can occasionally head down the wrong path. Claude Code automatically creates session-level snapshots of the codebase before every file edit. No setup is required — it happens silently in the background. When something goes wrong, /rewind opens a list of every prompt from the session, showing which ones actually changed the codebase. Choose the moment to return to, and restore just the code, just the conversation, or both.

Git Integration — Tracking Changes Over Time

When Claude makes changes to a project, there needs to be a way to track what changed, when it changed, and why. That is Git — version control that records every single change over time. It powers the checkpoint system behind /rewind, enables reviewing Claude’s changes, rolling back specific edits, and working with a team without overwriting each other’s work. Claude runs Git in the background automatically. No Git expertise is required to benefit from it.

Headless / CLI Mode — Letting Claude Work Alone

Normally Claude Code is interactive: type, Claude responds, approve actions, go back and forth. Headless mode (officially called CLI mode) lets Claude work completely on its own with no human in the loop. The -p flag runs a full agentic loop independently — no approval prompts, no waiting for input. It processes the task and outputs the results automatically.

Cost Management — No Surprises

There are two ways to pay for Claude Code, and picking the wrong one costs more. Monthly subscription (Pro at roughly $20/month, Max at $100–$200/month) provides a fixed cost with usage limits and all models included. Pay-as-you-go via the API charges per token and only makes financial sense for very light usage. For most users the subscription wins: predictable cost, no mid-build anxiety, no stopping to check a spending dashboard. Just build.

Work Trees — Multiple Claudes Working at Once

Big projects often require working on multiple things simultaneously — a new feature, a bug fix, an experiment — without any of them interfering with each other. Work trees create separate working directories, each with their own files and branches, but sharing the same repository history. Three terminal panes, three Claudes, three completely isolated tasks running in parallel. When each finishes, that branch gets merged back into the main project. Sub-agents can also be given their own work trees, with automatic cleanup when each finishes.

The Only Commands Worth Memorising

All 27 concepts reduce to a surprisingly short list of commands for everyday use. Everything else is optional — learned when it is needed, not before.