Getting Started with Claude
This guide explains what Claude is, how it helps you do more of what you're good at, and how it can automate complex work.
Set up Claude CodeSee what's possible, then make it yours.
The video below showcases concepts we'll cover later (AI agents, different Claude models, context windows, and skills). What you see is a starting point. Never accept AI's first tryAI SlopContent made with artificial intelligence that is perceived as lacking in effort, quality, or meaning, and produced in high volume.Source: Wikipedia as the final version. Your job is to refine it, poke at it with prompts, and make something you'd put your name on.
"Claude" is both a model and a product.
You can think of modelsLarge Language Model (LLM)Large Language Models are AI systems trained on broad data to understand and generate human language. Claude's models (Opus, Sonnet, Haiku) differ in capability, speed, and cost.Source: Anthropic docs as the brains and productsProductA user-facing application built on top of one or more models. Each product (Claude Chat, Claude Code, Cowork) provides a different interface and toolset for interacting with the underlying AI.Source: Anthropic as the hands and feet. Models are the Artificial Intelligence itself, each with different tradeoffs between capability and speed. Products are the user interfaces that give those models different tools to work with.
The model sees things differently.
Large language models were trainedTrainingTraining happens in two stages. In pre-training, the model reads enormous amounts of text from the internet (books, articles, websites, code) and learns patterns in language, facts, and reasoning. Then, in post-training it learns to follow instructions, answer questions helpfully, and have conversations. on text from the internet, then taught to help with tasks. They each have a knowledge cutoff Knowledge cutoffThe date when the model's training data ends. Unless you tell it or it searches the web, the model has no awareness of events, discoveries, or changes that happened after this date. date, meaning they have no awareness of events after training data ends. They do not learn from your conversations, or improve the more you use them. But training also shapes how they "see" the world. The example here shows this in action: the model fails simple counting when left to its own devices, but succeeds when using code to find the answer. This comes down to how it sees your inputs.
A model does not process text, audio, video, and images like we do. It breaks all this data into
tokensTokenA small chunk of text, roughly one word or word-fragment. The model reads and generates text one token at a time. Everything (text, code, images) gets converted to tokens.
– small chunks, each corresponding to a unique ID in its own vocabulary. Your question, any
documents you attach, and every word the model writes back are all seen as tokens.
This is why some things that seem trivial to us, like counting dots in a message,
can trip it up – it can only see these tokens. Together these tokens make up the model's
context window.
(This visual uses a simplified tokenizer. To see how real
models tokenize text, try Tiktokenizer)
use code
Every conversation, the model starts with a fresh context window.
The model's context windowContext WindowA fixed-size container (measured in tokens) that holds everything the model can see in one conversation. When it fills up, older messages are summarized to make room (a process known as 'compaction'). holds tokens from your prompts, its own thinking, file uploads, and more. As the context window fills up, costs rise and performance degrades (sometimes called "context rot"). There are several ways to get information in, from simple copy-paste to connectorsMCP (Model Context Protocol)An open standard that lets Claude reach into external systems like SharePoint, Slack, Google Drive, Jira, and databases. Called "connectors" in the Claude Chat interface; in Claude Code, you configure them as MCP servers. that let the model interact with external apps and data sources. Each product will also preload a system promptSystem PromptHidden instructions sent to the model by the product or platform before your conversation begins. They shape Claude's tone, rules, and behavior without being visible to you. Every Claude product has one and they vary widely. and any instructions you've saved. Claude can also load skillsSkillA pre-written set of instructions and resources (e.g., templates) that load into the context window on demand, giving Claude expertise for a specific task (creating slides, working with PDFs, writing docs, etc.). – pre-packaged instructions with expertise for tasks like creating documents, building websites, and writing in a specific tone or style. Skills load only when needed, so Claude gets the right know-how without always taking up valuable context window space.
The tools below bring outside information into the context window
With the right building blocks, AI agents can automate work.
In the demo video, Opus reads templates and orchestrates the team, Haiku handles high-volume file scanning, and Sonnet summarizes only the relevant files. Each agentAI AgentA system that autonomously performs tasks by designing workflows with available tools. AI agents can encompass a wide range of functions beyond natural language processing including decision-making, problem-solving, interacting with external environments and performing actions.Source: IBM docs gets its own context window, so each one stays focused on its task without being distracted by what the others are doing. Skills play a role here too: when a task requires specialized knowledge (like how to orchestrate an AI team), each agent can load the relevant skill, giving it expert-level instructions without cluttering the other agents' windows. AI excels at using these building blocks. As a user, this means you need to build really good blocks – clear instructions, good templates, and well-organized files.
How Claude Code can orchestrate work
Each product gives you different capabilities.
Most of what you create – whether it's a presentation, a website, or a document – is code that gets written. When you learn to manipulate that, you can do anything. This guide starts with Claude Code because it gives you the most influence for bringing your ideas to reality. The other products are always available if you need them.
| Capability | Code |
Cowork |
Chat |
|---|---|---|---|
| Fully orchestrate a team of agents | ✓ | ✗ | ✗ |
| Manage context window usage | ✓ | ✗ | ✗ |
| Build, test, and deploy software end-to-end | ✓ | ✗ | ✗ |
| Read and write local files | ✓ | ✓ | ✗ |
| Work in the background | ✓ | ✓ | ✗ |
| Write and run code | ✓ | ✓ | ✓ |
| Connect to external tools | ✓ | ✓ | ✓ |
| Search the web | ✓ | ✓ | ✓ |
| Save instructions across conversations | ✓ | ✓ | ✓ |
Set up Claude Code and build something.
To set up Claude Code, copy the prompt below and paste in Claude.ai. The chat will guide you, one step at a time.