Most people use Claude to answer questions. The real return comes from handing it the work.

When you ask for slides, a document or a spreadsheet, Claude creates a live artifact of that type. A deck opens in a slide editor where you can retitle a slide yourself; a document can be exported to Word or PDF; a deck to PowerPoint or PDF; a design to PDF or an image.

The usual habit is to ask for "some slides" and receive bullet points in the chat. Saying "make this a slide deck" produces the deck itself. You do not need to ask for a .pptx file unless you specifically want one: the export is available from the deck.

Multi-step work runs in Claude's own workspace and continues while you are away. A task list beside the conversation shows what is done and what remains, and you can pick the work up later from your phone.

If you tell Claude you will check back later, it will not stop to ask you questions: it takes the most reasonable reading, states at the top which reading it took, and carries on. For anything irreversible it prepares the work and waits for you.

Claude can set up a scheduled task that runs on a cadence you choose: every weekday morning, every Monday at 7, the first of each month. Each run is a fresh session, and you can be notified when it finishes.

Because each run starts with no memory of the conversation that created it, write the instruction so that it stands on its own: name the sources, the format and the recipient in full.

Claude builds a memory of your role, your projects and your preferences from your conversations, and it can also be told directly. "Remember that I write in British English" or "Forget what I said about the restructure" are acted on in that turn.

Standing preferences about format and tone apply to every response afterwards. You can ask "What do you know about me?" at any time. Memory is switched on and off with the "Generate memory from chats" setting, and an incognito chat runs without memory at all.

Claude can search and open your earlier chats. Refer to past work the way you would with a colleague, using its subject rather than its date, and Claude will find the conversation and continue from it.

The search matches the words that were actually used. "The parent survey analysis" will find it; "that thing we did" will not. This is controlled by the "Search and reference chats" setting.

A Project holds its own files, instructions and chats. Each Project also has a memory setting: connected, so that it shares context with your general memory, or kept separate in both directions.

For HR casework, a recruitment round or anything confidential, a separate Project keeps those details from surfacing in unrelated conversations. The setting is on the Project's own page, not in your general Settings.

Connectors link Claude to services such as Gmail, Google Drive, Slack and Canva. With a connector in place, Claude reads the source directly, and it can draft into the tool, for example a reply in Gmail.

If you ask about "my inbox" or "our shared drive" and no connector is set up, Claude will suggest the right one. Sending, posting or publishing on your behalf always requires your explicit confirmation.

A web page Claude publishes for you can do more than display. Depending on your account, it can keep a small shared database, so a sign-up sheet, an observation tracker or a booking list stores what people enter and everyone with the link sees the same records.

Pages start private to you and are one share away from colleagues. Claude can later add or change records in them without rebuilding the page.

Artifacts persist beyond the conversation that made them. In a new chat you can ask Claude to update one by name and it will edit the same artifact at the same address, so links you have already shared continue to work.

Comments left on a published artifact can also be sent to Claude, which makes a document something a team can review and Claude can act on.

A skill is a saved set of instructions that Claude loads automatically when a request matches it: your report template, your house style for letters, your checklist for reviewing a policy.

Whether a skill fires depends almost entirely on its description. Write it as a trigger ("Use when asked to draft a letter to parents") rather than a label ("Helps with writing"). The quickest route is to finish a task well once, then ask Claude to package what you did.

On high-stakes work, ask Claude to have a separate agent, one that has not seen the drafting, verify the result: every figure against its source, every citation against the paper, every date against the calendar.

A model checking its own draft tends to confirm it. A checker with no stake in the draft finds what the author missed. This costs a few extra minutes and is the closest thing to a second reader.

Claude will do what you ask. It is often more valuable to ask it first what is wrong with what you are asking for: the weak assumption in the plan, the objection a governor will raise, the section of someone else's document that is poorly designed.

Claude will disagree with you when asked plainly, and you can make this a standing preference so that it happens without asking.

For anything expensive to redo, the most efficient instruction is often to ask Claude what it needs to know. It will put a small number of multiple-choice questions to you, and the answers settle scope before any work begins.

Ten seconds answering questions avoids a twenty-minute deliverable aimed at the wrong reader.

Through Claude in Chrome, Claude can act on websites in your own browser with your existing sign-ins: navigate a portal, gather figures from several pages, fill in a form for you to review. With the desktop app it can also operate applications on your computer.

Use this for the steps no connector covers. Claude will not enter passwords or card numbers, and it asks before submitting a form, sending a message or making a purchase.