How to use Claude at work: Projects, artifacts, files, and privacy settings

Start here: the account and the three settings

Claude comes in a free tier, paid individual plans (Pro and Max), and Team and Enterprise plans for organizations. The plan decides two things that matter at work: how much you can use it, and what happens to your data.

On a personal plan (Free, Pro, or Max), open Settings, then Privacy, and look at the toggle that controls whether your chats can be used to improve future models. Anthropic made this an explicit choice in 2025, and the data retention period is longer if you opt in. Decide on purpose. Then follow the rule: never paste confidential company or client data into a consumer AI tool unless your organization has approved it.

On Team or Enterprise, conversations are not used for training by default, an admin controls membership, and Enterprise adds single sign-on, role-based permissions, and audit logs. If your company pays for it, use it for all work; it is the account your policy covers.

Two more settings worth knowing on day one:

• Incognito chat. Start one from the new-chat screen. The conversation is not saved to your history and does not feed memory. Use it for one-off sensitive questions and for experiments.
• Memory. On paid plans Claude can carry context across conversations, and on Team and Enterprise it keeps memory separate per Project, so one client's details do not leak into another client's work. Open the memory summary in settings, read it, and edit anything you would not want repeated.

If you work in healthcare, law, finance, HR, or insurance, assume PHI, privileged material, PII, and regulated financial data stay out of personal accounts until compliance signs off on a specific plan.

Model picker, thinking, and styles

The model picker offers a fast model for everyday work and a heavier one for hard problems. The names change with each release, so learn the rule instead: everyday drafting on the fast one, contract analysis and multi-step reasoning on the heavy one, and extended thinking switched on when you want Claude to work through a problem before answering.

Styles (Normal, Concise, Formal, or a custom style built from your own writing samples) set the default voice for every chat. Build one custom style from three emails you are proud of. It is the highest-return five minutes in this guide.

Projects and knowledge

A Project is a folder with three things: uploaded knowledge (documents, spreadsheets, transcripts), a set of standing instructions, and the chats that belong to that work. Every chat inside starts already knowing what is in the folder.

Good Projects at work look like this:

• Customer proposals: your last five winning proposals, your rate card, your style guide, and the instruction "match the structure of the attached proposals and never invent pricing."
• Board reporting: the last four board decks, the KPI definitions document, and "flag any number that contradicts a prior deck."
• Vendor contracts: the master services agreement, your negotiation playbook, and "always quote clause numbers."

Keep knowledge current. A Project with a stale rate card is worse than no Project, because Claude will use it with confidence. On Team and Enterprise you can share Projects, which is the simplest way to give a whole team the same reference set. Large knowledge bases are searched rather than read whole, so name your files clearly; "Q3 pricing v4 FINAL" is not a helpful name.

Artifacts: build small tools and documents

When you ask Claude for something substantial (a document, a chart, a web page, a small app), it appears in a panel beside the chat called an artifact. You ask for changes and watch the artifact update, then download it or share it with a link.

Three kinds of artifacts earn their keep at work:

Documents you will actually send. A one-page brief, a policy draft, a job description. Ask for the artifact, iterate in the panel ("shorten section two, make the tone warmer"), then export. On plans that include file creation, Claude can also produce Word, Excel, PowerPoint, and PDF files directly; check yours.

Small interactive tools. A calculator, a checklist, a form that turns inputs into a formatted output. No coding on your part. Describe the inputs, the rule, and the output, and Claude writes a working web page you can share with colleagues.

Charts and one-off dashboards. Upload a CSV and ask for an interactive chart. Useful for a meeting; not a replacement for your reporting tool.

Worked example: a quote checker the whole team can use

You run sales for a regional freight brokerage. Reps keep quoting below the floor margin. You want a tool that catches it before the quote goes out.

You are a web developer building a small internal tool for a freight brokerage. Build a single-page interactive artifact called Quote Checker.

Inputs: carrier cost, fuel surcharge percentage, accessorial charges, and the price quoted to the customer.
Rules: total cost = carrier cost, plus the fuel surcharge applied to the carrier cost, plus accessorials. Margin = (quoted price minus total cost) divided by quoted price. Our floor is [FLOOR MARGIN PERCENT]. Below the floor, show a red warning and the minimum price that meets the floor. Between the floor and [TARGET MARGIN PERCENT], show amber. Above target, green.
Output: total cost, margin as a percentage, and the status, updating live as I type.
Constraints: plain design, large numbers, works on a phone, no external libraries. If any rule is ambiguous, ask me one clarifying question before you build.

Test it with three real quotes where you know the answer. Then share the artifact link with the team. Total time: fifteen minutes, most of it testing.

File analysis

Claude reads PDFs, Word documents, spreadsheets, CSVs, images, and screenshots. Drag them into the chat or into a Project. Its strength is holding a long document in view at once: a hundred-page agreement, a year of meeting minutes, a full RFP.

For spreadsheets and numbers, Claude can run code in its analysis tool to compute, chart, and clean data rather than estimating in prose. Ask for it explicitly: "use code to calculate this." Then spot-check totals against the source file. Scanned PDFs, merged cells, and tables split across pages are where any assistant misreads.

Two prompts that work on almost any document:

• "Read this end to end. List every date, dollar amount, and named obligation in a table with the page it appears on."
• "Where does this document contradict itself or leave something undefined? Quote the passages."

Writing: where Claude is strongest

Claude produces first drafts that need less de-robotting than most. Three practices make the difference:

1. Give it samples. Paste three things you wrote and liked. "Match this voice" beats any list of adjectives.
2. Ask for one version. Three options is a way of not deciding. Ask for one, then ask for specific changes.
3. Ask it to argue back. "Tell me the weakest paragraph and why" gets an honest answer, and Claude will push back on a flawed premise when you invite it to.

It is also the tool most willing to say "I am not sure" or "this clause is ambiguous," which is exactly what you want in a second reader.

Connectors and where Claude reaches

On paid plans, Claude connects to Google Drive, Gmail, Google Calendar, Slack, and a growing directory of work tools (project trackers, design tools, CRMs), and on Team and Enterprise plans to Microsoft 365 sources such as SharePoint, OneDrive, Outlook, and Teams. These run on the Model Context Protocol, an open standard, which is why the list grows quickly. Availability depends on your plan and on what your admin has switched on; check the connectors directory in settings.

With a connector on, "summarize what happened in the Northwind channel this week and draft my reply to the open question" becomes a single prompt. Web search and the Research feature add cited work from the open web. Claude also runs as a desktop app, as a mobile app with voice, and inside Slack. Developers use Claude Code, a separate product, for software work.

Six workflows with prompts

• Email from bullets. "Write this as a reply to [PERSON]. Under 120 words, one ask, no exclamation marks. Match my style from the samples in this Project."
• Meeting notes to actions. Paste the transcript. "List decisions, actions with owners and dates, and unresolved questions. Quote the speaker for each decision."
• Redline a draft. Upload your document. "Act as a demanding editor. Give me a table of suggested changes: original sentence, revision, and a one-line reason."
• Executive summary from a long report. "One page for a reader with five minutes: what happened, what it means, what we should do, and what you are unsure about."
• Turn a process into a checklist. Describe how you do the task. "Convert this into a checklist artifact with tick boxes and the common mistake for each step."
• Learn a topic fast. "Explain [TOPIC] to a new [YOUR ROLE] in three levels: one paragraph, one page, and the five questions an expert would ask me."

Limits and what to verify

Claude does not generate images. Its knowledge stops at a training cutoff, so anything recent needs web search turned on. It can still invent a citation, misread a table, or agree too readily with a leading question; ask "what would change your mind?" and check every quote, number, and clause reference against the source. Usage limits on individual plans reset on a schedule, and heavy document work consumes them faster.

The bigger risk is the one it shares with every assistant: a confident, well-written wrong answer. Read How to fact-check AI output and decide in advance who signs off on anything with legal, financial, or medical consequences.

Next steps

• Compare the four major assistants in ChatGPT vs Claude vs Gemini vs Copilot and read the Claude tool page.
• Decide between Projects and their equivalents elsewhere with Custom GPTs vs Claude Projects vs Gemini Gems.
• Read AI privacy at work and the glossary entries for Model Context Protocol and context window.
• Use the ready-made prompts in writing and email and meetings and summaries.
• Practice with the prompt builder or take the course.