What Claude Is and How Conversations Work

Claude is an AI assistant built by Anthropic that you interact with through natural conversation.

You type a message, Claude reads it along with everything else said earlier in that same conversation, and it responds in kind.

There is no menu of fixed commands to memorize and no rigid syntax to get right.

You describe what you want in plain language, and Claude does its best to help, ask a clarifying question, or push back if something is unclear.

This page builds the basic mental model for how a conversation with Claude actually works, so that the rest of the interface-specific guides in this section make sense.

Summary

Core Idea: Claude is a conversational AI assistant - you communicate with it the way you would with a knowledgeable colleague, through back-and-forth messages.
Why It Matters: Understanding how context accumulates in a conversation explains both why Claude gets better as you clarify a task and why old conversations can start to feel muddled.
Key Concepts: conversation, context, context window, turn, new conversation.
When to Use: Any time you need help thinking through a problem, drafting something, researching a topic, or getting a task done - whether that's a quick question or a long project.
Limitations / Trade-offs: Claude only knows what is in the current conversation plus its general training; it does not remember prior conversations unless you are using a feature that explicitly carries memory forward, and every conversation has a practical limit to how much it can hold before older parts start to matter less.
Related Topics: Claude's interfaces, choosing the right interface, Claude 101 basics.

Foundations

A conversation with Claude is a sequence of messages going back and forth between you and the assistant.

Each message you send is called a turn, and Claude's reply is also a turn.

Unlike a search engine, Claude does not just match keywords to a fixed answer.

It reads your message, considers the conversation so far, and generates a response written specifically for that exchange.

This means you can ask a follow-up question the way you would with a person.

If you ask Claude to summarize an article and then say "make it shorter," Claude understands "it" refers to the summary it just gave you, because that earlier turn is still part of the conversation.

The full set of everything said so far in a conversation - your messages, Claude's replies, and anything you have shared like a pasted document - is called the context.

Context is what lets a conversation feel coherent rather than like a series of disconnected questions.

Think of a conversation with Claude less like a Google search and more like a working session with a colleague who has good judgment, broad knowledge, and no memory of your past meetings unless you fill them in again.

Every reply Claude gives is shaped entirely by what is visible in that conversation, plus its general training.

Starting a fresh conversation is like walking into a room and meeting Claude for the first time - it has no idea what you discussed yesterday unless a memory feature is explicitly carrying that forward.

Mechanics & Interactions

Every model that powers Claude has a context window - a limit on how much text (your messages, Claude's replies, and any documents or files you have added) it can hold and actively reason over in a single conversation.

You do not need to calculate this limit yourself, but it is useful to know it exists, because two practical situations follow from it.

First, a very long, sprawling conversation that has drifted across many unrelated topics will eventually push earlier details toward the edge of what Claude can weigh, so instructions from far back in the thread may get less attention than something you just said.

Second, when a conversation has served its purpose, starting a new one is usually better than continuing to pile onto an old thread that has drifted off topic.

A new conversation is not a downgrade; it is a clean workspace, and it often produces a sharper answer than continuing a conversation that has accumulated a lot of unrelated back-and-forth.

The practical mental model is this: within one conversation, Claude is cumulative - it builds on everything said before.

Across conversations, Claude is stateless by default - it starts over unless a memory or project feature is explicitly configured to bring context forward.

This has a direct effect on how you should write your first message in a conversation.

Because Claude has no outside knowledge of your specific situation beyond what training gave it, the quality of its first response depends heavily on how much relevant context you provide up front.

A vague request like "help me write an email" gives Claude very little to work with, so it has to guess at tone, audience, and purpose.

A request that includes who the email is to, what it needs to accomplish, and any constraints gives Claude the material it needs to produce something close to what you actually want on the first try.

This is the same principle that applies to working with any capable collaborator: the more relevant background you share, the less back-and-forth it takes to land on a good result.

Advanced Considerations & Applications

The conversational model stays consistent across every way you can reach Claude - the web app at claude.ai, the desktop app, mobile apps, the Chrome extension, and the Slack integration all use the same underlying turn-by-turn, context-carrying conversation.

What changes between interfaces is not how conversation works, but what else is available around it - for example, the desktop app adds modes for autonomous task execution and coding, while Slack adds the ability to work inside an existing thread.

Conversations can also include more than plain text.

You can share documents, images, or other files as part of a message, and Claude treats that content as part of the conversation's context alongside what you have typed.

This is why Claude can discuss a pasted contract or an uploaded screenshot in the same conversational way it discusses a plain question - the file becomes part of what it is reasoning over.

Longer or more complex efforts benefit from treating the conversation itself as a tool to manage deliberately.

If you notice a conversation has grown long and started to lose the thread, it is often faster to open a new conversation, restate the essential context in a tight paragraph, and continue from there, rather than trying to salvage a conversation that has drifted.

Approach	Strength	Weakness	Best Fit
Continue an existing conversation	Full context is already present; no need to restate background	Gets harder to steer as it grows long or covers multiple topics	A single task you are actively working through
Start a new conversation	Clean, focused context; often a sharper first answer	You must re-supply any background Claude needs	A new task, or an old conversation that has drifted
Share a document or file in-message	Gives Claude precise source material instead of a paraphrase	Only as useful as the file's relevance to the actual question	Reviewing, summarizing, or discussing specific content

Common Misconceptions

"Claude remembers everything from all my past conversations." - By default, each conversation is self-contained; Claude only has what is in the current conversation unless a memory or project feature is explicitly enabled and configured to carry information forward.
"Claude is a search engine that looks things up live." - Claude generates responses from its training and from whatever you provide in the conversation; it is not browsing the live web unless the interface you are using explicitly gives it that capability for that request.
"Longer messages always confuse Claude." - The opposite is usually true: more relevant detail up front generally produces a better first answer, since Claude has more to work with. What hurts is irrelevant clutter, not length itself.
"A new conversation loses everything, so I should never start one." - A new conversation only loses what you don't restate. For a task that has drifted or grown unwieldy, restating the key facts in a fresh conversation is often faster than continuing an old, cluttered one.
"Claude answers the same way no matter what I ask." - Claude's replies are generated specifically for each conversation's context, so the same underlying question can get a different, more useful answer depending on the details and framing you provide.

FAQs

What exactly is a "conversation" with Claude?

A conversation is the full back-and-forth thread between you and Claude - every message you send and every reply Claude gives, in order.
It is the unit Claude uses to understand what you are asking for; everything inside it is visible to Claude, nothing outside it is.

Does Claude remember what I talked about yesterday?

Not by default. Each conversation starts fresh unless you are using a feature specifically designed to carry information across conversations.
If you need Claude to know something from a past conversation, the reliable approach is to restate it in the new one.

What is the "context" of a conversation?

Context is everything currently part of the conversation: your messages, Claude's replies, and any documents or files you have shared in it.
Claude's response to any given message is shaped by all of this context, not just the most recent message alone.

What is a "context window" and why should I care?

It is the practical limit on how much conversation content Claude can hold and actively reason over at once.
You rarely need to think about it directly, but it explains why very long, sprawling conversations can start to feel less sharp, and why starting fresh sometimes helps.

Why did Claude misunderstand my follow-up question?

Usually because the follow-up assumed context that was never actually stated in the conversation, or because the conversation had grown long and drifted across several topics.
Restating the specific detail you're referring to, or starting a new conversation with a tight summary, usually resolves it.

Should I keep using one long conversation, or start new ones often?

Use one conversation for a single task you are actively working through - it keeps all the relevant context in place.
Start a new conversation when you're switching to an unrelated task, or when an old conversation has grown long and lost focus.

Can I share a document or image in a conversation?

Yes. Files you share become part of the conversation's context, and Claude can discuss, summarize, or answer questions about them just as it would plain text you typed.

Is Claude searching the internet when it answers me?

Not automatically. Claude's responses come from its training and the conversation itself; live web access, when available, depends on the specific interface and request.

Why does giving Claude more detail up front help?

Claude has no outside knowledge of your specific situation beyond what you tell it in the conversation.
A request with relevant background - audience, goal, constraints - gives Claude the material to produce a useful answer on the first try, instead of guessing.

Does the conversational model change between claude.ai, desktop, mobile, and Slack?

No - the underlying turn-by-turn, context-carrying conversation works the same way everywhere.
What differs between interfaces is the surrounding functionality, such as autonomous task modes on desktop or thread-based interaction in Slack.

What happens if I ask about something from earlier in a very long conversation?

Claude will generally still have it in context, but as a conversation grows very long and covers many topics, earlier details can get less weight than more recent ones.
If an important instruction from early on isn't being followed, restating it is a reliable fix.

Is it a problem to abandon a conversation partway through?

No. Conversations don't need to be "finished" - you can pick a task back up in the same conversation later, or move it into a new one if it makes more sense to start clean.

Claude 101 Basics - a first walkthrough of starting a conversation in the claude.ai web app.
Claude's Interfaces: Web, Desktop, Mobile, and Slack - how this same conversational model shows up across every surface.
Choosing the Right Claude Interface for the Task - a decision guide once you understand how conversations work.

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.