Every conversation you have with Claude runs on a small set of underlying mechanics: text gets broken into tokens, those tokens have to fit inside a context window, and the conversation itself is built from a sequence of labeled turns. You do not need to think about any of this to use Claude well, but understanding it explains almost every quirk you will run into - why a long chat starts to feel "forgetful," why pasting a huge document changes how Claude responds, and why starting a new chat gives you a clean slate. These examples walk through that machinery using an ordinary first conversation.
You open Claude.ai, type a question, and hit send.
"Can you help me write a short bio for my consulting website?"
- Your message becomes the first "user" turn in a new conversation.
- Behind the scenes, Claude does not receive your raw sentence - it receives the sentence broken into tokens, which are the actual units the model reads and generates.
- Claude replies with an "assistant" turn, and that reply is also generated one token at a time.
- Nothing before this message exists yet in this conversation - a brand-new chat always starts empty.
You type a sentence and wonder why "word count" and "token count" are not the same number.
"Tokenization isn't unbreakable - it's fascinating."
- A token is a chunk of text the model processes - often a whole short word, sometimes a word fragment, a piece of punctuation, or a space.
- "Tokenization" might split into a few token pieces rather than counting as one unit, while short common words like "is" or "the" are usually a single token each.
- As a rough rule of thumb, English prose runs about three to four tokens for every four words, so token counts are usually somewhat higher than word counts.
- This matters because usage limits, pricing, and the context window are all measured in tokens, not words or characters.
You paste a long article into the chat and ask Claude to summarize it.
"Summarize the article I just pasted in three bullet points."
- The context window is the total token budget for everything Claude can "see" in one exchange: your system prompt, every prior turn, and the reply it is about to generate.
- Depending on the model, that budget ranges from 200,000 tokens up to 1,000,000 tokens on Claude Fable 5.
- A long pasted article consumes a meaningful slice of that budget before Claude has written a single word back.
- If your conversation later grows close to the limit, the oldest parts of the chat become the first thing squeezed out or summarized, not the newest.
You scroll back through a conversation and notice the messages have a clear back-and-forth shape.
User: "What's a good icebreaker for a networking event?"
Assistant: "Try asking what project they're most excited about right now..."
- Every conversation is built from labeled turns: an optional system prompt that sets up behavior, your "user" turns, and Claude's "assistant" turns.
- On Claude.ai the system prompt is mostly invisible to you - it is set by the product, not typed by you - while user and assistant turns are the visible back-and-forth.
- Each new message you send is appended as another user turn, and Claude's reply is appended as another assistant turn.
- This turn structure is what lets Claude refer back to "the article I just pasted" or "the second option you gave me" - it is reading the whole labeled sequence, not just your latest line.
After Claude answers your first question, you ask a follow-up without repeating context.
"Make that bio shorter and more casual."
- Because this message lands in the same conversation, Claude can see your original request and its own prior reply as part of the context window.
- "That bio" resolves correctly because the earlier assistant turn is still present in the token budget Claude is reading from.
- Each new turn adds more tokens to the running total, so a long back-and-forth chat gradually uses up more of the context window.
- This is why follow-ups in one chat feel like a real conversation, while switching to a brand-new chat resets that shared understanding.
You open a new chat to ask about something unrelated to your last conversation.
"What's a reasonable structure for a two-week onboarding plan?"
- A new chat has an empty context window - none of the tokens from your previous conversation are included.
- Claude has no memory of the bio you wrote earlier unless you paste it in or reference it again.
- This is intentional: it keeps unrelated conversations from bleeding into each other and keeps each chat's context budget free for the topic at hand.
- Claude.ai Projects are the exception worth knowing about - they let you attach shared documents and instructions that persist across multiple chats inside that project, unlike an ordinary standalone chat.
You have a single long chat going for an hour, covering several topics and a few pasted documents.
"Going back to the first thing we discussed, can you also add a section on pricing?"
- Every turn so far - your messages, Claude's replies, and anything you pasted - is still sitting in the context window, adding to the token total.
- Claude can still "go back to the first thing" as long as that turn hasn't been pushed out of the window, which depends on the model's total token budget.
- Long chats with lots of pasted content are more likely to approach that budget than short, focused ones.
- If you notice Claude losing track of early details in a very long chat, that's usually a context-window symptom, not a sign Claude "isn't paying attention."
You need to have Claude read and reason over a very large set of documents in one sitting.
"I'm going to paste in five long reports and want you to cross-reference them."
- Context window size varies by model: Claude Haiku 4.5 offers 200K tokens, Claude Sonnet 5 and Claude Opus 4.8 offer larger windows up to 1M tokens depending on configuration, and Claude Fable 5 offers a full 1M-token window as standard.
- A task that involves pasting multiple long documents is exactly the kind of case where a larger context window matters most.
- Model choice also affects cost per token and speed, so the "biggest window" model isn't always the right default for short, simple questions.
- Picking a model with headroom above what you expect to paste in avoids running into the limit mid-task.
Claude gives you a fluent, specific-sounding answer that later turns out to be incorrect.
"According to a 2019 study, this approach improves results by roughly 40%..."
- Claude is a large language model: it generates each token by predicting what's statistically likely to come next, based on patterns learned during training - it is not looking anything up in a database.
- This is why a wrong answer can still read as fluent and confident - fluency and accuracy are produced by different parts of the process, and the model has no built-in "I don't actually know this" signal by default.
- A specific-sounding but unverifiable detail (a study, a statistic, a citation) is a common shape for this kind of error, sometimes called a hallucination.
- Claude's training data also has a fixed knowledge cutoff, so it can be confidently wrong about anything that happened after that cutoff unless it has search or another tool available to check.
Related: How a Claude Conversation Actually Works - traces one message through tokenization and reply generation | What Counts as a Token? Tokenization Explained - a closer look at tokenization | Context Windows: Why Claude Has a Memory Limit - why the token budget exists | How Claude Remembers (and Forgets) Within a Chat Session - session memory in depth
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.