Claude & LLM Fundamentals Basics

9 examples to get you started with Claude & LLM Fundamentals - 6 basic and 3 intermediate.

Prerequisites

• No setup is required. Everything below happens inside a normal Claude.ai chat.
• It helps to know that a conversation is made of turns: what you type is a user message, what Claude sends back is an assistant message, and an optional system message (set by you or the app you're using) can quietly shape how Claude behaves for the whole conversation.
• Nothing here assumes a technical background.

Basic Examples

1. What Claude Is

Claude is a large language model (LLM): a neural network trained on huge amounts of text and code to predict likely next words.

• Claude is not a search engine and does not "look up" answers in an index.
• Claude is not a fixed script of pre-written replies either, the way an old-style rule-based chatbot is.
• Every response is generated fresh, based on patterns learned during training.
• Anthropic is the company that builds and trains Claude.

Related: How Large Language Models Actually Generate Text - the mechanics behind this

2. A First Message

"Explain what a token is, in two sentences."

• This is a user message, the most common thing you'll send.
• Claude reads your whole message before generating a reply.
• Short, specific requests like this one tend to get short, specific answers.
• The reply itself is built from an assistant message, generated one small piece at a time.

3. Tokens: The Units Claude Reads and Writes

Text is broken into small chunks called tokens before Claude processes it.

• A token can be a whole short word, part of a longer word, or a punctuation mark.
• Claude's pricing and limits are measured in tokens, not characters or words.
• A rough rule of thumb: a token is close to three-quarters of an English word on average.
• Longer or more technical text tends to use relatively more tokens.

4. The Context Window

"Given everything we've discussed in this conversation so far, summarize the three main points."

• The context window is the total amount of text (measured in tokens) Claude can consider at once.
• It includes your messages, Claude's earlier replies, and any documents or files attached to the conversation.
• Once a conversation grows past the context window, the oldest content can no longer be "seen" by Claude.
• Different Claude models offer different context window sizes, which matters for long documents or long chats.

5. System, User, and Assistant Roles

A single exchange is made of up to three message roles working together.

• The system message (if present) sets ground rules or a persona for the whole conversation, and is usually set by the app or account, not typed by you mid-chat.
• The user role is you: your questions, instructions, and any pasted content.
• The assistant role is Claude's generated reply.
• Understanding these roles helps explain why the same question can get a different answer in a different app that uses a different system message.

6. Knowledge Cutoff

"What happened in the news this week?"

• Claude's training data has a fixed end date, called the knowledge cutoff.
• Asking about events after that date, without an explicit browsing or research feature turned on, is likely to get an honest "I don't have that information" or an outdated answer.
• Claude does not automatically browse the live web in a plain chat.
• If current information matters, look for a dedicated research or browsing feature rather than assuming Claude already knows.

Related: Why Claude Sometimes Gets Things Wrong - what happens when Claude answers outside what it reliably knows

Intermediate Examples

7. Hallucination in Practice

"Who wrote the paper 'Quantum Synergy Theory of Marketing' and what year was it published?"

• If no such paper exists, Claude may still generate a fluent-sounding, plausible-looking answer, a known failure mode called hallucination.
• Hallucination happens because Claude is predicting likely-sounding text, not checking a fact database.
• It tends to show up most on obscure facts, exact citations, dates, and numbers.
• Asking Claude to say "I'm not sure" when it lacks confidence, or asking it to double-check itself, can reduce but not eliminate this risk.
• Treat any specific fact, citation, or number from Claude as something worth verifying independently.

8. Choosing Between Models

"I need a quick, cheap draft of ten product description variants." versus "Walk me through a rigorous, multi-step analysis of this legal contract's risk clauses."

• The first prompt is a good fit for a fast, low-cost model like Claude Haiku 4.5, since the task is high-volume and doesn't need deep reasoning.
• The second prompt benefits from a model with deeper, visible reasoning, like Claude Opus 4.8 or Claude Fable 5, which apply extended thinking before answering.
• Claude Sonnet 5 is the default model for most everyday and coding work, balancing quality, speed, and cost.
• Picking the right model is a trade-off between response quality, latency, and cost per task, not a single "best" choice.

9. Extended Thinking on a Hard Problem

"Before answering, reason step by step through the trade-offs, then give me your final recommendation."

• Some Claude models, including Claude Fable 5, apply extended thinking (also called adaptive reasoning) by default: a visible chain of reasoning steps generated before the final answer.
• This tends to improve accuracy on multi-step problems like planning, math, or weighing several options.
• The reasoning is still built from the same underlying next-token prediction mechanism, just applied more deliberately and at greater length.
• Extended thinking is not a guarantee of correctness, it reduces certain kinds of errors but does not eliminate hallucination.

Related: Inside the Transformer: How Neural Networks Power Claude - what's happening underneath a response like this

Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5. Model names, pricing, and product features move quickly - verify current specifics at platform.claude.com/docs before relying on them.