Getting started with Claude.

The chat window is where you type prompts. The left menubar runs down the side and holds navigation, tools and your profile menu.

1. New chat, the plus icon near the top of the left menubar, starts a fresh conversation.
2. Chats opens your conversation history.
3. Your initial at the bottom-left opens the profile menu, including Settings (section 02).

To begin, click into the message box, type a request in plain English, and send. The conversation is saved in Chats.

To work from your own material, click the plus (+) button at the left of the message box and attach a file: a PDF, a document, a spreadsheet or an image. The file becomes part of the conversation's context, and Claude draws on it alongside your request when it answers (more on context in section 06).

One more thing before getting started: Claude's first answer may be as good as a first draft. Reply in the same chat to refine it: shorter, more formal, put it in a table, make it appropriate for the intended reader, etc. Each follow-up builds on what came before, so iterating in chat shapes the result.

Most controls sit around the message box.

Treat the chat box as a work brief, not just a search box. The more useful context you provide—a file, a source, the intended reader and a clear instruction—the better the work that comes back. The habit that matters most

Most account-level settings sit behind your profile menu.

1. Click the circle with your initial at the bottom-left.
2. Click Settings.
3. Settings opens as an overlay on the General tab. The section that matters most is Instructions for Claude.

Tell Claude how you work

The most useful setting in Claude. In Settings → General, find Instructions for Claude, the box where Claude asks what it should consider in responses. What you write here applies to every chat.

1. Open Settings.
2. On the General tab, below your profile details, find Instructions for Claude.
3. Click into the instructions box.
4. Write a few lines on tone, length and format.
5. Save your changes if prompted. Every new chat from here follows these.

A shortcut: let Claude interview you

Not sure what to write? Have Claude build it. Paste this into a chat, answer a few questions, then copy the block it produces into the instructions box above.

Review your privacy and memory settings

While you are in Settings, take two minutes before your first real conversation. Review two controls and learn one privacy habit.

1. In Settings, open Privacy. Find Help improve our AI models, which controls whether eligible chats and coding sessions can be used to improve Claude, and choose deliberately.
2. Open Capabilities, find Memory, then click View and manage memory. Claude curates your memory file from your chat history. Review it periodically and edit if necessary.
3. Incognito chat is not saved to your history or memory. In Claude Enterprise, it is still logged. To activate, click on the ghost icon at the top right corner of the Claude window.

Claude comes in several model families and each is suited to different kinds of work. Switch between them in the model menu next to the send button. Reasoning models such as Fable and Opus are very powerful but run slower and use more tokens, so match the model to the task. Which models appear depends on your plan, region, feature rollout and, on Claude Enterprise, administrator settings.

Model	Reach for it when	What it is good at
Fable 5	You are handing off a large, demanding assignment that may span several stages or workstreams.	Anthropic's most capable model for demanding tasks, reasoning and longer-horizon work. Use it when you want Claude to help architect and carry a whole workstream.
Opus 4.8	A bounded task genuinely needs deep thinking, judgment or reasoning.	Complicated analysis, deep reasoning, nuanced judgement and complex knowledge work. Uses more tokens, so save it for when it earns its keep.
Sonnet 5	You are not sure. Start here.	The daily driver: writing, analysis, documents, spreadsheets and most multi-step work. Strong balance of quality, speed and token use.
Haiku 4.5	You want a quick answer, instantly.	Summaries, quick lookups, pulling a figure from a document. Lightest on token use.

When in doubt, Sonnet. Reach for Opus when a task needs deeper thinking, judgment or reasoning, and Fable when you are handing off a massive assignment. The one rule of thumb to remember

Model labels and availability change as Anthropic rolls out new releases. The practical rule still holds: start with Sonnet for everyday work, use Haiku when speed matters, reach for Opus when you need deeper reasoning, and reserve Fable for the most demanding end-to-end assignments. In a chat you can switch models at any point, and the change applies from Claude's next reply.

On models that support it, the same menu includes Effort. The model determines the capability tier; effort controls how much reasoning it applies. Higher effort can produce a more thorough answer, but it takes longer and uses more tokens. If Effort is missing for a model, use the default behavior or choose a model that exposes effort settings.

Level	When to use it
Low / Medium	Routine asks where speed matters. Uses fewer tokens.
High	The everyday balance of quality and speed for harder work.
Extra	Deeper reasoning for long-running, complex work, without Max's full cost.
Max	The hardest, get-it-right work. Deepest reasoning and slower.

The menu marks each model's recommended level with a Default badge. Trust the badge until you have a reason not to. Labels and availability can change; the options shown in your model menu are the source of truth for your account.

For everyday work the default is usually right. Raise it for complex tasks: document analysis, multi-step calculations, anything where being right beats being fast.

You get strong results when you define the task clearly. Give Claude what you would give a capable colleague picking up the work cold: the role to take, useful background and sources, the output you need, the reader, your preferences and what to avoid. It will not know your standards unless stated in chat or Instructions settings.

Think of six levers, not a checklist: persona, context, output, audience, preferences and guardrails. Most good prompts use two or three. Know all six so you can spot which missing one is hurting the output.

Cover	What it settles	Say something like
Persona	What role Claude should play.	"Act as an investment analyst with experience evaluating private-market transactions."
Context	What facts, source material or background Claude should use.	"Use only the attached CIM. Treat missing information as unknown."
Output	What should come back and in what form.	"Prepare a two-page first-pass investment memo by summarizing the attached CIM. Lead with the conclusion, then a metrics table, summary of the deal and key risks."
Audience	Who will read it and what they care about.	"Write for a CIO deciding whether to run a first pass."
Preferences	What tone, format and level of detail Claude should follow.	"Use direct, concise language. Use a table for metrics and short paragraphs for analysis."
Guardrails	What Claude should avoid, flag or ask about before drafting.	"Cite the page for every number. Flag unsupported claims and never invent a figure."

Context is everything Claude can see when it answers: your message, the files you attach or paste, the chat so far and your instructions from section 02. Files, email and other material count only after you explicitly attach, paste or connect them, so the practical skill is choosing what Claude sees.

The quality of Claude's answer is influenced by the quality of the context. When using Instructions (section 02) and projects (next section) context you rely on loads once instead of being retyped.

A normal chat does not automatically carry its files and task rules into the next one. A project gives every chat access to a pool of knowledge you load once and standing instructions for how Claude should behave inside it. Project Instructions supersede the global Instructions setting.

1. In the left menubar, click Projects, then New project. Give it a name and briefly describe what you are trying to achieve.
2. Add project knowledge: templates, past examples, a style note. Loaded once and available to every chat inside the project.
3. Set the project instructions: your preferences and instructions for this Project only.
4. Start chatting. Every chat can use the project knowledge and instructions, so you skip the repeated setup.

Your first complete run

1. Start a fresh chat and attach a non-sensitive document you understand well.
2. Use the prompt levers the task needs: persona, context, output, audience, preferences and guardrails.
3. Review the first answer and give one round of specific feedback.
4. Check every name, number, date and citation against the source.
5. If the work will recur, create a project and add the reusable files and instructions.

The concepts translate from ChatGPT. Everything on the right is covered in the section listed.

In ChatGPT	In Claude	Section
Custom instructions	Instructions for Claude, in Settings.	02
Memory	Memory, reviewed and edited in Settings.	02
Temporary chat	Incognito chat: not saved to history or memory, and still subject to data retention policies.	02
Projects	Projects, with knowledge files and standing instructions.	07
A custom GPT	The closest match is a project for persistent knowledge and instructions. A skill is a reusable set of instructions for a specific procedure that Claude can activate when needed.	07
Canvas (now writing and code blocks)	Artifacts, opening in a side panel you can edit in place.	01
Deep research	Research, in the plus menu by the message box.	01
The model picker	The model menu next to send: Haiku, Sonnet, Opus and Fable, sized by task.	03